AI-Native Supply Chain Resilience: A Multimodal Architecture for Predictive Intelligence, Optimization, and Real-Time Decision-Making

 Corresponding author: Arunraju Chinnaraju Copyright © 2025 Author(s) retain the copyright of this article. This article is published under the terms of the Creative Commons Attribution License 4.0. AI-Native Supply Chain Resilience: A Multimodal Architecture for Predictive Intelligence, Optimization, and Real-Time Decision-Making Kannan Avalurpet Loganathan 1 and Arunraju Chinnaraju 2, * 1 Independent Researcher, California, USA. 2 Westcliff University, Seattle, USA. GSC Advanced Research and Reviews, 2025, 25(03), 078-126 Publication history: Received 27 October 2025; revised on 04 December 2025; accepted on 06 December 2025 Article DOI: https://doi.org/10.30574/gscarr.2025.25.3.0372 Abstract Supply Chain Global Systems are currently in constant states of uncertainty; Disruptions will occur and can rapidly move across all global connected Supply Chains due to their Interconnected Nature. Traditionally, Supply Chain Planning is based upon static forecasting with periodic optimization which does not allow for sufficient responsiveness and scalability. This paper proposes an integrated Architecture for Predictive and Adaptive Supply Chain Resiliency utilizing AI technology in conjunction with multilevel components which incorporate Large Scale Data Engineering, Hybrid Forecasting and Autonomous Optimization. The Framework utilizes Advanced Time Series Models including LSTMs, Temporal Convolutional Network, ARIMA-AI Hybrid Ensembles and Transformer Based Predictor Models for Disruption Sensitive Forecasting. The models utilized for Forecasting are inputted into Adaptive Decision Mechanisms using Reinforcement Learning (i.e. DQN, PPO), Multi-Objective Optimization Engines that balance Service Levels, Operational Costs and Carbon Efficiency. Additionally, Resilience Analytics including Agent-Based Simulation, Monte Carlo Stress Testing and Network Robustness Metrics measure how Disruptions Propagate and how Quickly Systems Can Recover. Finally, The Implementation Layer of the Framework utilizes Cloud Edge Orchestration, Micro-Services Architecture and Real-Time Decision Intelligence Dashboards to Support Human-AI Collaboration. Furthermore, this Paper discusses Governance, Model Transparency and Ethical Deployment Considerations to Ensure Responsible Use of Autonomous Decision Systems. Finally, the Paper discusses potential future Research Directions including Quantum Enhanced Optimization, Digital Twin Ecosystems, Generative Scenario Modeling and Federated Intelligence for MultiEnterprise Collaboration. Keywords: AI-Driven Resilience; Hybrid Forecasting Models; Reinforcement Learning; Deep Q-Networks; Resilience Analytics; Quantum Optimization 1. Introduction Supply Chain Management today is no longer based on traditional supply chain management (SCM) based on a stable and continuous flow of materials and information, but is increasingly dynamic and interconnected globally. Conventional SCM has traditionally been based on deterministic forecasting methods, periodic planning cycles and optimization procedures assuming constant and predictable market conditions. Today's supply chain disruptions occur rapidly and cascade quickly through supply networks; these types of disruptions are difficult for traditional SCM approaches to deal with (Christopher & Peck, 2004). The increased interconnectivity between different nodes and subnetworks in modern supply networks creates larger ripple effects from local disruptions creating large disruptions throughout the total system (Pettit et al., 2010). In addition to the above mentioned issues, political instability, environmental disasters and changes in consumer behavior can create a wide variety of complex fluctuations affecting the continuity of supply chains (Chan, 2011). The increasing volatility of today's logistics environment further reinforces GSC Advanced Research and Reviews, 2025, 25(03), 078-126 79 that resilience should be considered a fundamental design requirement rather than an additional design feature (Ivanov, 2020). Resilience has progressed from the basic concept of having redundant elements in a supply chain to being a broad capability that includes the concepts of anticipation, adaptation, and quick restoration of normal operating conditions after a disruption (Christopher & Peck, 2004). Today's supply chain resilience refers to the ability of a supply chain system to recognize early signs of potential disruption and to absorb the shock created by the disruption, thereby minimizing the negative impact of the disruption on the system's overall performance (Pettit et al., 2010). The ability to restore performance to normal operating levels in an acceptable timeframe is one of the key components of resilient design (Chan, 2011). Resilience is therefore an active and continually developing capability of an organization, rather than just a passive reliance on static safety buffers and/or overstocked inventories (Ivanov, 2020). Real-time Optimization using Artificial Intelligence (AI) has become a fundamental component of this transformation. Machine Learning Models and Predictive Analytics provide the capability to identify risks and anomalies much earlier in complex operational environments than previously possible (Zamani et al., 2023). Dynamic Optimization enables Supply Chains to run simulation scenarios for various responses to disruptions and to develop the optimal strategy for responding to the disruption (Ivanov, 2020). Unlike traditional models that rely primarily on past data, current AI-based models incorporate high frequency multi-modal input streams including demand signals, transportation conditions, supplier actions, and external risk indicators (Zamani et al., 2023). These capabilities support automatic adjustments to sourcing strategies, transportation routes, inventory placement and production schedules, thereby enhancing the adaptive response to disruptions of the supply chain (Ivanov, 2020). The primary objective of this study is to design an intelligent architecture for continuous forecasting, adaptation, and optimization of supply chain operations across all supply chain functions. The proposed architecture meets the requirements of the Viable System Model (VSM), where structural coordination and feedback loops are required for long-term survival of the system (Chan, 2011). It also represents modern ideas regarding resilient network design by combining predictive intelligence with response and recovery capabilities (Pettit et al., 2010). The integration of predictive models, multilayer optimization and real-time decision interfaces reflects emerging frameworks for AIenabled supply chain transformations (Zamani et al., 2023). Treating resilience as a computationally solvable problem allows the study to position the supply chain as a self-regulating and self-correcting autonomous system able to maintain its performance in highly unstable and uncertain global environments (Ivanov, 2020). 2. Theoretical Foundations and Conceptual Framework The theoretical underpinning for AI driven supply chain resilience lies at the intersection of predictive analytics, optimization theory, and decision intelligence — three areas that collectively provide the architectural components necessary to predict, adapt, and learn within changing operational environments (Baryannis et al., 2019). Predictive analytics represents the systems' interpretative function (Makridakis et al., 2018), while optimization theory represents the systems' ability to act strategically under constraints (Tang, 2006). The combination of predictive analytics and optimization theory forms the base for the decision intelligence that connects the data, models, and organizational processes within an organization and creates decision pathways at the enterprise level (Tang & Musa, 2011). Together, these three elements form the conceptual basis for an intelligent supply chain that can anticipate disruption and execute response activities with precision and adaptability (Sunmola & Baryannis, 2024). From an analytical perspective, predictive analytics relies on the statistical and computational procedures used to enable systems to recognize patterns, produce forecasts, and recognize emerging anomalies (Makridakis et al., 2018). A wide range of analytical procedures including those based on regression analysis, temporal models using deep learning algorithms, and other analytical procedures provide the means for the system to interpret demand signals, logistics patterns, and risk indicators (Baryannis et al., 2019). Pedagogically, this constructivist perspective is consistent with the notion that intelligence is constructed through ongoing interactions with data streams (Tran et al., 2025). Additionally, predictive modeling serves as a dynamic memory system that captures operational experiences and converts them into actionable foresight (Sunmola & Baryannis, 2024). Time series deep learning models that process the continuous flow of data generated by modern supply chains represent the foundational principles for predictive analytics (Makridakis et al., 2018). The primary reason for this is because the Long Short-Term Memory (LSTM) networks have the ability to capture long-range temporal dependencies in nonlinear sequences (Hochreiter & Schmidhuber, 1997). These characteristics make LSTMs particularly well-suited for modeling the variability of demand, the instability of transportation systems, and the uncertainties related to the environment (Baryannis et al., 2019). Additionally, transformer-based architectures extend temporal reasoning even GSC Advanced Research and Reviews, 2025, 25(03), 078-126 80 further by utilizing attention mechanisms that allow the system to recognize long distance relationships across multiple variables simultaneously (Vaswani et al., 2017). These deep learning systems utilize back propagation as their primary learning principle that enables iterative error correction and model refinement over time (Rumelhart et al., 1986). Collectively, these architectures enable the forecasting framework to represent the multivariate and nonlinear nature of supply chain dynamics (Tran et al., 2025). Hybrid model architectures that combine traditional statistical methodologies with deep neural networks improve predictive performance (Makridakis et al., 2018). Traditional time series models have limitations in terms of representing nonlinear disruptions and structural breaks (Tang & Musa, 2011). Deep neural networks overcome these limitations by representing nonlinear residual behavior that traditional models cannot represent (Hochreiter & Schmidhuber, 1997). Hybrid models increase robustness, stability, and adaptability in volatile or unpredictable operational contexts (Baryannis et al., 2019). By integrating the interpretability of statistical models with the representational capabilities of deep learning, hybrid models enable more resilient and generalizable forecasting under uncertain supply conditions (Tran et al., 2025). Causal and structural modeling techniques represent a fundamental aspect of predictive analytics by providing assurance that the predictions made by a system reflect the underlying generative mechanisms and not just superficial correlations (Tang, 2006). These types of models assist organizations in identifying if changes in supplier lead times, transportation conditions, or inventory levels are structural causes of downstream disruptions (Tang & Musa, 2011). By mapping causal pathways between critical variables, the system becomes more transparent in terms of decision making (Baryannis et al., 2019). This capability is essential to developing decision architectures that respond to the actual dynamics of a system rather than reacting to temporary data artifacts (Sunmola & Baryannis, 2024). To maintain predictive accuracy over time, sophisticated mechanisms must be implemented to detect evolving patterns and changes in the underlying distribution of the data (Makridakis et al., 2018). As customer behavior, supplier reliability, and transportation networks evolve, models must adapt to avoid declining predictive performance (Baryannis et al., 2019). Continuous monitoring and calibration of models will assure that the analytical outputs remain relevant to the operational realities of the organization (Tran et al., 2025). These adaptive capabilities support resilience by recognizing emerging vulnerabilities prior to the point where they become widespread systemic disruptions (Sunmola & Baryannis, 2024). Additionally, predictive analytics must incorporate external pipeline signals that provide organizations with knowledge of macroeconomic, geopolitical, and environmental factors that may affect the stability of upstream and downstream supply chain operations (Tang, 2006). AI-based processing of these external signals will improve organizations’ situational awareness and operational readiness (Baryannis et al., 2019). By incorporating both structured and unstructured data into the predictive layer of the system, the system will develop a more complete understanding of risk, uncertainty, and opportunity (Tran et al., 2025). Finally, optimization theory provides the structural logic to determine how supply chains evaluate alternatives, allocate resources, and behave under constraints (Tang, 2006). Optimization theory defines feasible solution sets and prioritizes action in situations where uncertainty exists (Tang & Musa, 2011). AI increases this capability by providing systems the ability to pursue adaptive strategy development through feedback-based learning processes (Sunmola & Baryannis, 2024). Strategically, this position optimization as a key resilience mechanism that transforms intelligence into competitive advantage (Baryannis et al., 2019). AI paradigmatically form the operational foundation of the proposed architecture. Systems using learning paradigms use back propagation to continuously refine the internal representations of the system over time (Rumelhart et al., 1986). Architectures designed around sequence-based processing allow systems to understand temporal dependencies in logistics and demand patterns (Hochreiter & Schmidhuber, 1997). Mechanisms using attention allow the system to dynamically weight the importance of different temporal inputs (Vaswani et al., 2017). Collectively, these mechanisms allow for continuous improvement, contextual reasoning, and adaptive decision making in highly complex environments (Tran et al., 2025). Systems thinking and dynamic network theory provide additional context to view the supply chain as an interconnected system rather than a sequential process (Tang, 2006). Disruptions propagate through networks using relational dependencies and feedback loops (Tang & Musa, 2011). Modeling relational dependencies will provide a more accurate representation of cascading failure effects and resilience thresholds (Baryannis et al., 2019). From this perspective, resilience is defined by how effectively a system processes information, reorganizes structure, and remains coherent under stress (Sunmola & Baryannis, 2024). GSC Advanced Research and Reviews, 2025, 25(03), 078-126 81 In bringing all of these disciplines together, the conceptual framework focuses on the continuous interaction between prediction, adaptation, and resilience (Tran et al., 2025). Prediction provides visibility into emerging risks and trajectories (Makridakis et al., 2018). Adaptation allows data-driven responses through the optimal reconfiguration of the system (Baryannis et al., 2019). Resilience arises as the cumulative result of when predictive intelligence and adaptive control are operated in synchronization (Sunmola & Baryannis, 2024). This architecture views the supply chain as a self-regulating, analytically empowered system that has the potential to sustain performance under uncertainty (Tang, 2006). 3. Architectural Design of Intelligent Predictive Optimization Systems An architecture for an AI-enabled resilient supply chain has to work like a multi-layered computational eco-system where data, models, optimization algorithms and decision-making interface are continuously interacting with one another. The modularity of this architecture is deliberate; since each layer of the architecture will perform different cognitive functions while being dependent on other layers via feedback loops. The architecture is designed hierarchically starting from raw data to be captured, then progressing to predictive knowledge, followed by planning for optimization to decision-making in real time allowing the system to learn, adapt and reconfigure itself as operational conditions change. Thus, the multi-layered architecture is a technical and conceptual representation of Systems Theory (ST), where Resiliency does not arise from the independent performance of individual parts, but arises from the coordinated behavior of the total architecture as shown in Fig. 1. Figure 1 AI-native supply chain decision architecture Data Layers — It has a large-scale distributed nature, and a highly intensive data environment that is used to continue to ingest and transform heterogeneous types of data streams (Gunasekaran et al., 2017). The integration of Enterprise Resource Planning (ERP) and Supply Chain Management (SCM) allows enterprise systems to provide access to structured data for analytical processes using standardized and interoperable formats (Teixeira et al., 2025). IoT generated inputs provide real-time operational visibility to asset, logistics infrastructure, and production environments; thereby increasing situational awareness (Gunasekaran et al., 2017). Access to external data sources such as weather indicators, transportation indicators, and macro-level market indicators provide contextualized intelligence to the system (Zamani et al., 2023). The Feature Store is a structured and versioned data repository that keeps model training and model inference consistent in an AI-driven environment (Teixeira et al., 2025). Therefore, this tier acts as the system's sensory foundation and provides high-resolution signals to be processed in the subsequent analytical layers (Gunasekaran et al., 2017). GSC Advanced Research and Reviews, 2025, 25(03), 078-126 82 Decision Intelligence Layer — Provides operational backbone through scalable orchestration of machine learning services and analytical workflows (Teixeira et al., 2025). In modern AI infrastructures, structured governance mechanisms are required for model lifecycle management, auditability, and version control (Zamani et al., 2023). Continuous monitoring will identify any prediction bias, concept drift, or performance degradation issues; and can then address them by recalibrating or retraining the models (Gunasekaran et al., 2017). Scalable environments enable the system to adaptively allocate resources to maintain low-latency responsiveness to changing operational demands (Teixeira et al., 2025). Through the use of these mechanisms, the system maintains reliability and operational trustworthiness. Predictive Intelligence Layer — Functions as the analytical core of the architecture and incorporates forecasting and anomaly detection into specialized model ecosystems (Teixeira et al., 2025). Convolutional and Recurrent Sequence Models enable the extraction of long-range patterns and non-linear temporal structures from complex datasets (Bai et al., 2018). These architectures are well-suited to learn from irregular disruptions and fluctuating behavioral cycles present within supply chain data (Bai et al., 2018). Gradient-Boosted Decision Trees enhance structured feature learning by providing scalable and high-performance pattern recognition capabilities (Chen & Guestrin, 2016). Overall, these modeling techniques translate multidimensional data streams into reliable forward-looking intelligence (Gunasekaran et al., 2017). Hybrid Modeling — Further increases predictive precision by integrating advanced deep learning architectures with statistical reasoning (Teixeira et al., 2025). Decomposition-based approaches improve stability in environments that have identifiable trends and seasonality (Gunasekaran et al., 2017). Deep sequence learners compensate for nonlinearity and sudden shifts by capturing hidden temporal dependencies (Bai et al., 2018). Ensemble-based frameworks, specifically those that utilize gradient-boosting, provide additional robustness by minimizing over-fitting while maintaining sensitivity to subtle patterns (Chen & Guestrin, 2016). This integration improves resilience to volatility in data and structural uncertainty (Zamani et al., 2023). Causal Modeling — Increases the interpretive ability of predictive analytics by identifying meaningful relationships between variables versus surface-level correlations (Teixeira et al., 2025). Causal modeling helps the system distinguish between true drivers of disruption and coincidental variance (Zamani et al., 2023). By establishing causation, the system also gains transparency, interpretability, and a stronger foundation for strategic decision-making (Gunasekaran et al., 2017). Optimization Layer — Translates analytical insights into actionable strategies for routing, sourcing, scheduling, and resource balancing (Teixeira et al., 2025). Mathematical optimization provides decision-making support under complex constraint scenarios including cost, speed, availability, and risk (Gunasekaran et al., 2017). Machine learning extends mathematical optimization by providing adaptability and the ability to learn from environmental feedback (Zamani et al., 2023). Therefore, the system can respond strategically to both structured plans and unforeseen situations (Teixeira et al., 2025). Decision Interface Layer — Enables interaction between computational intelligence and human cognition through transparent visualization, simulation, and action mechanisms (Gunasekaran et al., 2017). Dashboards convert complex outputs into understandable scenarios that support planning and evaluation (Teixeira et al., 2025). Autonomous agents extend this layer by taking action based on pre-defined conditions being met, thereby reducing response time and improving consistency (Zamani et al., 2023). Event Bus — Establishes a real-time communication fabric that connects all architectural layers, enabling asynchronous exchange of predictions, decisions, feedback, and system states (Gunasekaran et al., 2017). This event-driven approach enables components to react dynamically while coordinating activity within a unified system (Teixeira et al., 2025). As a result of the continuous flow of information, the architecture continually adapts and self-corrects without centralized micromanagement (Zamani et al., 2023). This architectural framework operates as a closed-loop learning organism in which each operation influences data, predictions, decisions, and outcomes in a circular fashion (Gunasekaran et al., 2017). With each operational cycle, the architecture improves its predictive accuracy, optimizes its decision-making efficiency, and develops greater adaptability (Teixeira et al., 2025). Through recursive learning and intelligent reconfiguration, the system does not develop a static form of resilience but instead a dynamic, continually evolving form of capability (Zamani et al., 2023). GSC Advanced Research and Reviews, 2025, 25(03), 078-126 83 4. Predictive Modeling and Forecasting Algorithms Predictive modeling (the use of AI) is the brain or core of the proposed AI-driven Supply Chain Resiliency Architecture. Predictive modeling is the primary mechanism for the system to go from reactive decision-making to anticipatory intelligence. This section will lay out the analytical basis for predicting what the system will look like in the future, identifying disruption patterns prior to materialization, and developing adaptive strategies in real-time. The research positions time-series forecasting as part of a layered intelligent process where hybrid AI models such as LSTM, Prophet, ARIMA-hybrid models and advanced deep architectures work together to take in external signal fusion, causal reasoning and probabilistic inference. In particular, these models are designed to capture both short term and long term structural shifts in supply networks; and recognize non-linear relationships among fluctuations in demand, variability in lead times, capacity constraints, and global environmental conditions. From a technical standpoint, the predictive modeling component of this framework combines temporal deep learning, statistical modeling, causal inference, ensemble methods, and exogenous data streams into a single, integrated forecasting engine. The macroeconomic signals, geopolitical instability, climate anomalies, supplier reliability indicators, and transportation constraints are represented as both structured and unstructured features that add to the predictive context. Thus, the traditional univariate forecasting model is transformed into a multivariate, multi-scale, and multi-source analysis process. The architecture assumes that real world supply networks will exhibit nonstationarity, concept drift, regime shifts and complex interdependencies; therefore, it requires models that continually learn, dynamically update, and quantify uncertainty in each projection. Therefore, the predictive layer is not restricted to generating numerical forecasts; rather, it generates a range of probabilistic outcomes that support optimization engines, reinforcement learning agents, and enterprise decision interfaces. Figure 2 Long Short-Term Memory (LSTM) based supply chain forecasting architecture LSTM based Supply Chain Forecasting Architecture acts as advanced Decision-Intelligence Engine for converting Dispersed Operational Data into Forward-Looking Actionable Insights (Wang & Yeh, 2014). The System Aggregates Multivariate Input Signals in Time Dependent Sequence Xt = [x₁, x₂, …, xₙ] (Xun & Li, 2024) That Includes Critical Sources of Information Such as Inventory Stock Levels, Reorder Thresholds, Historical Sales Patterns, Shipment Records, Supplier Lead Times, Real-Time Customer Order Streams (Xun & Li, 2024). Prior To Entering Model, These Inputs Pass Through Preprocessing Stage Where Data Is Standardized, Synchronized Across Timestamps and Structured Into Consistent Temporal Sequences (Nguyen et al., 2021). The Preprocessing Stage Is Important Because Real Supply Chain Data Are Typically Noisy, Asynchronous And Heterogeneous (Nguyen et al., 2021). Therefore, By Aligning These Diverse Data Flows Into Unified Time-Series Structure, The Model Receives Coherent and Context-Rich View of How Demand, Inventory and Supply Conditions Evolve Over Time (Wang & Yeh, 2014). GSC Advanced Research and Reviews, 2025, 25(03), 078-126 84 In The Heart of This Architecture Lies The Long Short-Term Memory Network, Which Is Specifically Designed For Capturing Both Short-Term Fluctuations and Long Range Dependencies In Sequential Data (Krawczyk & Wójcik, 2021). In Contrast to Traditional Forecasting Models That Rely Only On Recent Trends or Fixed Historical Windows, LSTMs Can Preserve Information Across Extended Time Horizons, Making Them Suitable for Modeling Long-Term Patterns in Sales and Demand Behavior (Xun & Li, 2024). Thus, They Can Learn Complex Patterns Such as Seasonal Demand Cycles, Long Term Behavioral Shifts, Regional Purchasing Trends and Delayed Responses to Operational Changes (Nguyen et al., 2021). More Advanced Hybrid Architectures Combine LSTM with CNN and GRU Layers to Improve Interpretability and Enhance Extraction of Local and Global Patterns in Supply Chain Data (Li et al., 2025). Output Of The Model Predictive Outcome Is Represented As ŷᵗ, Which Represents System's Forecast of Future Demand Conditions, Inventory Risk Levels and Potential Stock-Out or Over-Stock Situations for Specific SKUs, Locations, or Distribution Points (Krawczyk & Wójcik, 2021). Instead of Generating Only Single Point Estimate, Output Reflects Broader Assessment Including Expected Demand Intensity, Volatility Patterns and Risk Classifications Across the Operational Network (Li et al., 2025). Predictive Outputs Are Immediately Transferred to Optimization and Decision-Support Engine. At This Stage, System Converts Forecasts into Concrete Operational Actions Such as Recommended Reorder Quantities, Adjusted Replenishment Cycles, Dynamic Safety Stock Adjustments and Risk-Based Prioritization of Products and Locations (Wang & Yeh, 2014). Event-Sensitive Data Integration Further Improves This Process by Incorporating Contextual Information and Multi-Dimensional Relationships that Improve Quality and Robustness of Decision-Making Under Uncertainty (Du et al., 2025). When Implemented in Real Supply Chain Operations, These Recommendations Are Continuously Monitored for Outcomes Such as Actual Sales Performance, Inventory Turnover and Fulfillment Success Rates (Nguyen et al., 2021). This Feedback Is Then Reintegrated Into Learning Loop, Allowing LSTM Driven System to Adjust Internal Parameters, Recalibrate Weights and Improve Predictive Accuracy Over Time Through Iterative Learning (Krawczyk & Wójcik, 2021). Closed Feedback Mechanism Enables Continuous Model Adaptation in Response to Structural Changes and Evolving Operational Dynamics (Li et al., 2025). From Organizational and Business Perspective, This Architecture Reduces Excess Inventory Accumulation While Simultaneously Minimizing Revenue Loss Caused by Stockouts (Xun & Li, 2024). It Also Improves Warehouse Efficiency by Streamlining Inbound and Outbound Flows and Strengthens Service Reliability Across Fluctuating Market Conditions (Wang & Yeh, 2014). Improved Forecast Accuracy and Anomaly Detection Also Contribute to Better Risk Visibility and Faster Response During Disruptive Events (Nguyen et al., 2021). Ultimately, LSTM Driven Forecasting Architecture Transforms Supply Chain Management From Reactive and Retrospective Function into Predictive, Anticipatory and Strategically Optimized System Capable of Sustaining Performance, Resilience and Long Term Profitability in Highly Dynamic Markets (Du et al., 2025). 5. GRU Low-Latency Edge Inference Pipeline The Diagram 3 illustrates a High Speed Predictive Architecture for Ultra-Fast Inference, Minimized Transmission Delay and Localized Decision-Making in Environments where these are Critical to Success. Real-Time Edge Based Predictive Architectures are becoming Increasingly Common in Distributed and Edge-Based Supply Chain Environments for Supporting Immediate Context Aware Intelligence (Adewumi & Oyekunle, 2025). Operational Environments Where Rapid Micro-Level Decisions Have Significant Impact on System Efficiency and Resilience include Warehouses, Logistics Terminals and Transportation Hubs (Gour & Yadav, 2025). Input to the system consists of Continuously Streaming Signals such as Barcode Scans, Arrival Timestamps, Dock Sensor Readings etc. Each of these Time Ordered Signals Represents a Micro-Level Event and Collectively Form a Structured Temporal Sequence Defined As: Xₜ = [xₜ₋₁ , xₜ , xₜ₊₁] Event-Based Learning Architectures are Particularly Effective in Transforming Sequential Data from Input Streams into Actionable Intelligence Due to Their Ability to Preserve Temporal Dependencies (Cho et al., 2014). The Input Data is Filtered, Synchronized and Structured in Machine Interpretable Form Prior to Entry into the Learning Core to Maintain Signal Quality and Reduce Noise in Edge-Based Learning Systems (Li et al., 2025). Unlike Centralized Batch-Oriented Forecasting Systems, the Architecture Uses Event-Driven Processing Whereby Each New Occurrence Instantaneously Influences the Predictive State (Gour & Yadav, 2025). GSC Advanced Research and Reviews, 2025, 25(03), 078-126 85 Domain-Specific Feature Engineering Transforms the Refined Input Data and Applies Rolling Averages, Change Rates, Congestion Ratios, and Inter-Event Time Gaps. These Transformed Features Enhance the Semantic Clarity of Raw Input and Allow the System to Identify Complex Behavioral Patterns Such as Bottlenecks, Burst Arrivals, or Irregular Cargo Flows (Ravichandran, 2025). The Engineered Features Then Act as Primary Inputs to the Gated Recurrent Unit Model That Forms the Computational Core of the Pipeline (Li et al., 2025). Figure 3 Gated Recurrent Unit (GRU) based forecasting pipeline GRUs are Specifically Designed for High-Speed Resource-Efficient Inference in Constrained Environments. GRUs Use a Lightweight Internal Architecture That Allows for Faster Execution While Preserving Temporal Memory (Cho et al., 2014). Therefore GRUs are Suitable for Edge Deployment Where Computational Resources, Power and Latency Are Constrained (Adewumi & Oyekunle, 2025). The GRU Dynamically Balances Newly Arriving Data With Relevant Historical Patterns Allowing for Immediate Adaptation to Disruptions and Fluctuations in the Local Supply Chain Environment (Li et al., 2025). The Output of the GRU Is Expressed as a Prediction State: hₜ = f(Xₜ , hₜ₋₁) This Output Reflects Near-Term Operational Conditions Including Queues Sizes, Expected Delays, Congestion Risk, Throughput Estimations, etc. (Ravichandran, 2025). Rather than Relate on Distant Cloud Infrastructure, the Output of the GRU is Briefly Held in a Local I/O Buffer and Published to an Event Bus Enabling Connected Systems to React Immediately via Updated Schedules, Modified Routing, Automated Alerts, etc. (Thompson & Hall, 2025). The Architecture Also Includes a Residual Learning and Update Mechanism That Continuously Compares Actual Outcomes to Predicted Outputs. The Difference Between Observed and Predicted Performance is Reintegrated Into the Learning Loop Allowing the System to Self-Adjust and Refine Its Internal Parameters in Real-Time (Du et al., 2025). Adaptive Feedback Mechanisms Such as Those Used in This Architecture Are Essential to Maintaining Forecasting Accuracy Under Changing Operational Conditions Such as Demand Spikes, Labor Constraints or Environmental Disruptions (Gao et al., 2025). Organizational Benefits of the Low-Latency GRU-Pipeline Include Improved Operational Response, Reduced Congestion at Key Nodes and More Efficient Utilization of Physical and Logistical Assets (Thompson & Hall, 2025). Organizations Achieve Localized Autonomy Without Reliance on Centralized Decision Systems by Embedding Intelligence Directly At the Point of Action (Gour & Yadav, 2025). In Uncertain and Volatile Environments, This Type of Real-Time Edge Intelligence Provides a Measurable Advantage in Improving Responsiveness, Throughput and Ensuring Resilient and Adaptive Supply Chain Performance (Gao et al., 2025). GSC Advanced Research and Reviews, 2025, 25(03), 078-126 86 6. TCN Multi-Sequence Disruption Detection Model Figure 4 Temporal Convolutional Network-based forecasting model In contrast to other models based on recurrent neural networks (RNNs) for time series modeling that have shown to capture trends but lack robustness and adaptability to extreme outliers, the study proposed TCN-based outlier-resistant supply chain forecasting model provides a robust temporal intelligence framework to represent and predict the real world irregularities, volatility, and disruptions in supply networks. Unlike RNNs, the model uses a convolutional neural network (CNN), particularly dilated CNNs that have proven to be highly stable and effective in sequential learning tasks that include large-scale dependencies and irregularities (Bai et al., 2018). To preserve the temporal relationship among variables and enhance the model's ability to be resistant to extreme anomalies or abrupt structural changes in the time series data, the model use causal padding and dilated convolutions (Lai et al., 2018). The model inputs multiple information streams including: • Historical Demand (Dt) • Current Inventory State (It) • A variety of External Factors (Et) that impact Operational Stability These signals can be represented as: Xt= [Dt, It, Et] Historical demand (Dt) represents past demand for products; Current inventory state (It) indicates the current inventory condition; Exogenous Variables (Et) include a variety of external influences on operational stability such as weather disturbances, macroeconomic indicators, supplier instability, and geopolitical constraints. By incorporating these exogenous and disruption related variables into the model, the study enable the model to go beyond simply extrapolating trends and reflect the true operational complexities of real-world supply chains (Jha & GSC Advanced Research and Reviews, 2025, 25(03), 078-126 93 9. Bayesian Network causal forecasting architecture Figure 8 Bayesian Network causal forecasting architecture Instead of simply forecasting future events based on past trends, the architectural design shown in the diagram 8 creates a sophisticated method to forecast and understand the behavior of supply chains through cause-and-effect modeling (Heckerman, 1995). The architecture consists of two primary components; a Structural Causal Model and a Bayesian Network, which work together to allow the system to determine why disruptions occurred, how those disruptions propagated through various parts of the interconnected systems, and ultimately, how different types of interventions could have altered the eventual outcomes (Brintrup et al., 2025). The Bayesian Network, positioned at the top of the architecture, portrays important supply chain variables (lead time, port congestion, etc.) as interconnected nodes with directed paths of influence. As opposed to portraying each variable independently, the Bayesian Network illustrates how variables affect each other (for example, how extreme weather affects port congestion and port congestion subsequently causes instability in lead times) (Tsang, 2025). Each of the causal pathways is documented through conditional probability tables, therefore enabling the system to provide possible outcomes in terms of probabilities rather than deterministically, which is crucial when dealing with unpredictable and disruptive environments (Heckerman, 1995). At a higher-level, the joint probability distribution of the variables can be represented as follows: P(X₁, X₂, X₃) = P(X₁ | U) · P(X₂ | X₁, V) · P(X₃ | X₂) GSC Advanced Research and Reviews, 2025, 25(03), 078-126 94 Where, X₁ may represent lead time, X₂ may represent port congestion, X₃ may represent environmental disruption, and U and V represent additional unseen contextual variables (labor strikes, geopolitical instability, infrastructure failures, etc.) (Ojha & Ghadge, 2018). This type of representation allows the system to show how changes to the upstream variables result in downstream outcomes in a clear and mathematically proven manner (Mohammadi & Esmaili, 2025). Below the Bayesian Network is the Structural Causal Model, which defines causal relationships through mathematical functions rather than through correlation alone (Brintrup et al., 2025). Within this layer, variables are defined as having explicit causal relationships, or as having an explicit causal mechanism defining how one factor leads to change in another. For instance, the causal relationship between lead time and outside interference can be defined as follows: Y = g(X₁, X₃, V) Where, Y is a disruption-related outcome (delivery failure probability), X₁ is lead time, X₃ is some form of port or weather interference, and V is context-specific, unmeasured noise (Heckerman, 1995). This function definition allows the system to reason about interventions in addition to predictive associations, which is a hallmark of causal modeling approaches (Mohammadi & Esmaili, 2025). Within the inference layer, the do-operator is used to invoke counterfactual reasoning as follows: P(Y | do(X₁ = x')) This allows the architecture to explore the results of hypothetical interventions, such as decreasing the level of congestion or changing transportation modes (Brintrup et al., 2025). Through this functionality, supply chain decision makers will be able to examine the potential outcomes of proposed courses of action prior to their implementation, thereby converting predictive analytics into a testing ground for decision-making (Tsang, 2025). Ultimately, the causal output produced by the model is combined in the forecasting layer to create actionable information such as causal risk scores, disruption likelihoods, and scenario-conditional performance predictions (Ojha & Ghadge, 2018). These outputs are then utilized directly to inform decisions related to inventory planning, route selection, supplier diversity, and resilience investments (Mohammadi & Esmaili, 2025). From a strategic standpoint, this architecture shifts companies from being reactive in terms of analyzing data to being proactive and intelligence-driven (Brintrup et al., 2025). Not only does it indicate what may happen within a highly complex supply network, it also indicates why something may happen and what specific actions may be taken to alter the eventual outcome (Tsang, 2025). This degree of transparency and ability to intervene is necessary to manage systems that include significant interdependence and ongoing uncertainty. Through the integration of probabilistic reasoning, structural modeling, and counterfactual simulation, this causal forecasting framework establishes a flexible, adaptive and decision-ready intelligence layer that enhances both the resilience of operations and the effectiveness of strategic planning in today's modern supply chain environments (Heckerman, 1995). 10. Structural Equation Impact Path Model The Structural Equation Model (SEM) in Figure 9 depicts a complete model for examining how internal and external elements affect the total strength and responsiveness of a supply network (Bollen, 1989). This is particularly well-suited to the subject matter since resilience is the product of many observable and non-observable variables that have complicated interactions with each other and with each other's interactions (Ivanov et al., 2025). The model divides factors affecting the system into exogenous and endogenous elements, and introduces latent constructs that mediate system behavior (Yoon & Kim, 2023). Exogenous variables are located on the left-hand side of the figure and represent factors beyond the immediate operational control of the organization, including market volatility, logistics flexibility, and supplier collaboration (Shahzad et al., 2025). Market volatility is the reflection of variations due to demand variability, price volatility, geopolitical issues, and macro-economic uncertainties (Chen et al., 2025). Logistics flexibility refers to the organization's ability to make dynamic route choices, change transportation mode options, and allocate capacity based upon the changing environment (Ariadi et al., 2025). Supplier collaboration refers to the extent of trust, information exchange, and coordination between the organization and its supply partners. All of these will help determine the extent to which risk information and mitigation strategies can be communicated throughout the supply network (Xiao & Li, 2025). As GSC Advanced Research and Reviews, 2025, 25(03), 078-126 95 all of these factors are not created by the internal system itself, they are represented as exogenous inputs that will create directionally causal effects on the internal construct. Figure 9 Structural Equation Model – Supply Chain Path Modeling architecture There are also three key latent constructs at the center of the model; system agility, shock adaptability, and operations robustness (Ivanov et al., 2025). Latent constructs represent internal organizational abilities that describe how well a supply chain responds to disruption (Bollen, 1989). System agility describes how quickly and responsive a decisionmaking process is in response to sudden changes (Shahzad et al., 2025). Shock adaptability defines the ability of the system to modify its structure, sourcing patterns, and operational logic in response to extreme or unforeseen events (Yoon & Kim, 2023). Operations robustness describes the reliability and continuity of internal processes as they experience external stress (Chen et al., 2025). Since the three constructs are mutually reinforcing, the bidirectional arrows between them represent the fact that improving one capability enhances the effectiveness of the other two (Ivanov et al., 2025). The endogenous variables are found on the right-hand side of the model and describe measurable outcomes of the internal resilience mechanisms (Xiao & Li, 2025). The endogenous variables include lead-time recovery and network redundancy. Lead-time recovery describes the rate at which the system returns to its normal state of operation after experiencing a disruption (Shahzad et al., 2025). Network redundancy describes the existence of backup suppliers, alternative logistics pathways, and spare capacity that would prevent the loss of the entire system if the primary pathway were to fail (Ariadi et al., 2025). Endogenous variables are defined as being inside the system because they are created by the system itself, as opposed to being created independently by external causes (Bollen, 1989). Simplified representations of the structural equations relating the variables can be written as: GSC Advanced Research and Reviews, 2025, 25(03), 078-126 96 η = Bη + Γx + ζ In the above equation, η represents the latent constructs, such as system agility, shock adaptability, and operations robustness, x represents the exogenous variables, such as market volatility, logistics flexibility, and supplier collaboration, B represents the structural relationships between the latent constructs, Γ represents the influence of the exogenous variables on the latent constructs, and ζ represents the internal noise within the system (Bollen, 1989). Using this type of mathematical representation allows us to quantitatively evaluate how much of a difference various forces make in creating internal system capabilities (Ivanov et al., 2025). The relationship between the latent constructs and the observable variables can be conceptualized as: y = Λyη + ε In the above equation, y represents the observable variables, such as lead-time recovery and network redundancy, Λy represents the factor loadings connecting the latent constructs to the observable variables, and ε represents the measurement errors (Bollen, 1989). Through this connection the model can map abstract system capabilities to tangible performance metrics that can be measured, compared against one another, and continually improved (Yoon & Kim, 2023). Business-wise, the model gives companies both diagnostic and predictive tools. Companies can use the model to see which external stresses most greatly influence their internal resilience and which internal capabilities provide the greatest increases in performance when they are enhanced (Chen et al., 2025). Instead of simply reacting to disruptions, company leaders can proactively target their investments in flexibility, collaboration, and adaptive infrastructure (Xiao & Li, 2025). In terms of strategy, the SEM framework transforms resilience from an abstract concept into a tangible and controllable asset that offers science-based justification for companies to improve their supply chain design, increase their risk readiness, and establish sustainable competitive advantages in unstable global markets (Shahzad et al., 2025). 11. Ensemble Stacking Pipeline (LSTM + TCN + ARIMA) Figure 10 Ensemble Stacking Architecture for Supply Chain Forecasting GSC Advanced Research and Reviews, 2025, 25(03), 078-126 97 The Ensemble Stacking Architecture for Supply Chain Forecasting Illustrated Above Formalizes How Multiple Heterogeneous Models Can Be Combined Into One Single Risk Aware Forecasting Engine That Is More Accurate Than Any Individual Model. The Raw Input Data Streams Of Order Histories, Shipment Logs, Lead-Time Records, Inventory Positions, And Exogenous Signals Such As Prices, Disruptions Or Promotions Are Standardized By A Dedicated Preprocessing Layer To Align Timestamps, Handle Missing Values, Encode Categorical Attributes, Transform Raw Input Into Clean Feature Matrix Suitable For Machine Learning, And Construct Lagged Features, Rolling Statistics, And calendar variables so temporal structure is explicitly encoded in model inputs. (fu et al., 2025) Cleaned data is then partitioned using cross validation splits and model folds. The architecture adopts a k-fold stacking protocol where each base learner is trained on k−1 folds and predicts on the held-out fold, producing out-of-fold predictions for every observation (giannopoulos et al., 2025). These out-of-fold predictions are used to train the metalearner at the stacking layer only on predictions generated from data not seen during corresponding base model training (bandara et al., 2020). Let xₜ denote the feature vector at time t and fᵢ(·) the i-th base model. The base-level prediction matrix, often called the stacking feature matrix, is constructed from these individual model outputs across time (fu et al., 2025). The individual model layer combines models with different inductive biases and strengths. Recurrent and decomposition based methods capture temporal and seasonal structure associated with sequential demand, while convolutional and decomposition based deep learning approaches emphasize pattern extraction and resilience to noise (heng et al., 2024). Other machine learning models specialize in capturing nonlinear interactions among explanatory variables such as promotions, holidays, and regional dynamics complementing the temporal focus of recurrent and convolutional networks (giannopoulos et al., 2025). It is precisely the diversity in error characteristics that enables the stacking process to outperform individual models under varying conditions (bandara et al., 2020). The stacking layer operates as a meta-learner that takes each row of the stacking matrix and learns how to blend the base predictions into a single, more reliable forecast. In its simplest form, the stacked prediction for time t can be expressed as a weighted combination of base forecasts, where the stacking weights are learned from data to minimize forecasting error (fu et al., 2025). In more advanced settings, the stacking function may be implemented as a regularized regression, a gradient boosted model, or a shallow neural network that can capture nonlinear interactions between the base predictions and additional covariates such as forecast horizon, product class, or volatility regime (giannopoulos et al., 2025). Above this, an ensemble optimization module selects stacking parameters by minimizing a task-specific loss function, such as weighted absolute error or quantile loss, which can be tailored to reflect the business impact of over and underforecasting (heng et al., 2024). Regularization constraints are applied to reduce overfitting and improve generalization, while asymmetric penalty functions can be used to prioritize service level protection for high value or high-risk products (fu et al., 2025). This stage transforms the ensemble from a purely statistical construct into a financially aligned and risk-conscious decision support tool (giannopoulos et al., 2025). The enterprise computing layer operationalizes the ensemble within a scalable, microservices-based architecture. Each individual model can be deployed as an independent service with its own monitoring and update cadence, while the stacking service orchestrates the generation of the final forecast in real-time (bandara et al., 2020). Continuous evaluation mechanisms allow new models or updated versions to be tested and validated online before replacing existing components, enabling safe model evolution without disrupting operations (heng et al., 2024). Risk-adjusted forecasts generated by the ensemble are ultimately consumed by planning and optimization systems to inform reorder quantities, safety stock levels, production schedules, and transportation capacity decisions (giannopoulos et al., 2025). Because the stacking architecture explicitly integrates diverse modeling paradigms and learns dynamic weighting schemes, it is more resilient to regime shifts, sudden demand surges, and localized disruptions than any single model (fu et al., 2025). Strategically, the ensemble framework allows organizations to treat forecasting as a modular and continuously improvable capability, where new modeling innovations can be incorporated over time while aligning predictive performance with real world operational and financial objectives (bandara et al., 2020). GSC Advanced Research and Reviews, 2025, 25(03), 078-126 98 12. External Signal Fusion Architecture Figure 11 External Signal Fusion Architecture The External Signal Fusion Architecture formalizes the transformation of heterogeneous macro level signals into a unified representation that conditions supply chain forecasts on geopolitical, economic and climate risks and recognizes that external environments have strong effects on supply chain stability and visibility (Brintrup et al., 2024). Unlike many models that treat demand as a singular time series, the architecture provides explicit evidence that purchasing behavior, transportation efficiency, and supplier reliability are continually affected by dynamic external variables (Pounder et al., 2013). Therefore, the architecture has a dedicated feature pipeline for processing external data, and after feature extraction, the external and internal feature streams are merged into a composite risk context for forecasting and scenario analysis (Lamola, 2023). The left-hand side of the architecture extracts three primary categories of external data; geopolitical risk indicators, economic indicators, and climate indicators. Examples of geopolitical indicators could be: conflict intensity measurements, regional stability indices, and country-specific risk scores, all of which directly impact the feasibility of logistics and the reliability of suppliers (Pounder et al., 2013). The economic indicators would reflect: exchange rate fluctuations, consumer confidence, unemployment rates, industrial output, and commodity price volatility, both influencing demand behavior and production viability (Rashid & Rasheed, 2025). Climate indicators would include: short-term weather anomalies, and long-term environmental trends including temperature anomalies, precipitation anomalies, and frequency of extreme climatic events, and these indicators would create instability in production cycles and transportation networks (Supply Chain Model on Uncertainty Demand, 2015). All of these external data streams pass through their own respective feature extraction modules to assure that high-frequency noise is removed and the macro-patterning is preserved (Lamola, 2023). Each external feature extraction module produces an external feature vector x_geo, x_eco, and x_clt representing the current and future state of the macro environment (Brintrup et al., 2024). The external feature vectors are then passed into a fusion network that generates a single composite representation of the external risk. In its simplest form, this fusion is performed via a weighted linear combination where: x = α_geo * x_geo + α_eco * x_eco + α_clt * x_clt, GSC Advanced Research and Reviews, 2025, 25(03), 078-126 99 the fused external feature vector is denoted as x and the weights α_geo, α_eco, and α_clt represent the modeled sensitivities of demand and logistics behaviors to each class of external signal (Rashid & Rasheed, 2025). More complex versions of the fusion layer are accomplished using a neural network with attention mechanisms providing the ability to account for both linear and non-linear interactions between the geopolitical, economic, and climatic signals (Brintrup et al., 2024). The External Influences Attribution Module uses an attention based neural layer to further refine the composite representation of external risk by determining which factors most significantly contribute to the observed and predicted changes in demand. During training, the model will learn to focus on the relevant features and time frames associated with the greatest disruption or shift, such as currency volatility causing shifts in demand for import intensive products or extreme weather creating difficulties in regional distribution (Supply Chain Model on Uncertainty Demand, 2015). The resulting composite vector x^R is a risk enriched external context aligned with the internal operational data streams (Lamola, 2023). The external risk context is then incorporated into the core supply chain forecasting engine. As previously stated, the internal demand and inventory histories are used to form the standard time series input stream X_int, while the composite external feature vector x^R influences the forecast through either feature concatenation, feature gating, or feature modulation. The forecasting function can be represented as follows: ŷ_t = f(X_int(1:t), x^R_t), where ŷ_t is the risk adjusted demand prediction at time t and f(·) represents the deep learning forecasting function (Brintrup et al., 2024). Therefore, the model is conditioned upon the actual macro-risk environment at each point in time and does not rely solely upon past internal behavior (Rashid & Rasheed, 2025). The bottom portion of the architecture provides support for performing scenario-based risk simulations. In this phase, the external vectors are systemically perturbed to simulate hypothetical futures, such as increased geopolitical tensions, economic recessions, and sequences of extreme weather events. The modified vectors then undergo the same fusion and forecasting processes as those in the previous phase, yielding a distribution of potential outcomes instead of a single deterministic forecast (Supply Chain Model on Uncertainty Demand, 2015). The resultant scenario-based forecasts provide a means to establish a set of risk envelopes that quantitatively characterize the uncertainty and reveal vulnerabilities across sourcing, inventory, and capacity planning decisions (Pounder et al., 2013). From a theoretical standpoint, the architecture operationalizes multi-modal representation learning and late-stage data fusion in a supply chain context, connecting macro-environmental modeling to micro-level operational forecasting (Lamola, 2023). Through the separation of extraction, fusion, attribution, and forecasting into distinct yet interconnected layers, the architecture facilitates model interpretability while maintaining the predictive capability of deep learning (Brintrup et al., 2024). From a strategic standpoint, the architecture allows organizations to transition from reactive and disruptive-driven decision making to proactive and anticipatory decision making based upon a unified and mathematically consistent framework for understanding the influence of external factors on demand and supply dynamics (Rashid & Rasheed, 2025). 13. Uncertainty Quantification & Probabilistic Forecast Architecture (MC Dropout, Bayesian Inference, Quantile Regression) The Architecture shown in Figure 12 illustrates a probabilistic and predictive framework for uncertainty assessment and forecasting; in other words, a probabilistic forecasting system for the prediction of future supply chain performance (Abdar et al., 2021). Traditional forecasting methods predict a single expected value of demand or lead times, while actual supply chains face multiple sources of uncertainty due to variations in demand, instability among suppliers, geopolitical events, and disruption in logistics and transportation (Kendall & Gal, 2017). The system depicted here addresses the need for a probabilistic understanding of uncertainty by combining Monte Carlo dropout, Bayesian inference, and quantile regression to provide a predictive distribution (instead of a single deterministic outcome) for every forecast (Gal & Ghahramani, 2016). The first stage of the process is the supply chain input vector x, which includes all possible relevant inputs including past demand, current inventory, supplier reliability metrics, and external indicators of variability such as weather and economic changes (Abdar et al., 2021). The inference model uses Monte Carlo dropout (a process in which dropout GSC Advanced Research and Reviews, 2025, 25(03), 078-126 100 layers are kept active during prediction) and processes the input vector to produce a predictive distribution of demand or lead times instead of a single deterministic output (Gal & Ghahramani, 2016). Figure 12 Uncertainty Quantification & Probabilistic Forecast Architecture Each output is generated by a forward pass of the model, where the model's parameters have been randomly perturbed by a dropout mask (Gal & Ghahramani, 2016). The process can be represented as follows: ŷ⁽ᵗ⁾ = f(x, θ⁽ᵗ⁾) Where t is the index of the specific forward pass and θ⁽ᵗ⁾ represents parameters modified through the application of a random dropout mask (Gal & Ghahramani, 2016). Through multiple forward passes of the model, the model produces a distribution of outputs, from which the mean and variance of the predictive distribution can be determined: μ = (1/T) Σᵀₜ₌₁ ŷ⁽ᵗ⁾ σ² = (1/T) Σᵀₜ₌₁ (ŷ⁽ᵗ⁾ − μ)² These measures capture two types of uncertainty in the model's predictive outputs: uncertainty inherent in the data used to train the model, and uncertainty related to the model itself (Kendall & Gal, 2017). In addition to Monte Carlo simulation, the architecture also employs Bayesian inference, which treats model weights as probability distributions as opposed to fixed points (Blundell et al., 2015). The Bayes theorem describes the relationship between these concepts as: P(θ | D) ∝ P(D | θ) P(θ) Here, P(θ | D) is the posterior probability of model parameter weights given training data D; P(D | θ) is the likelihood of observing the training data; and P(θ) is the prior belief regarding the model parameters (Blundell et al., 2015). As new data arrive, the model updates its internal beliefs regarding the model parameters and its confidence in those beliefs (Abdar et al., 2021). In particular, Bayesian inference is useful when the environment evolves quickly (i.e., when there GSC Advanced Research and Reviews, 2025, 25(03), 078-126 101 are frequent disruptions or volatility in supply chains), and the model must adapt to changing environmental conditions (Abdar et al., 2021). Finally, the architecture uses quantile-based regression techniques to create confidence intervals for forecasts as opposed to simply reporting a "best guess" for demand or lead times (Lakshminarayanan et al., 2017). While traditional models focus on estimating the mean of a probability distribution, the model described above estimates multiple quantiles of the predictive distribution (e.g., the 10th, 50th, and 90th percentiles) (Lakshminarayanan et al., 2017). The model estimates these quantiles using the pinball loss function: ρᵩ(r) = { φr, if r ≥ 0 ; (φ − 1)r, if r < 0 } Where φ is the selected quantile and r is the residual between the actual and predicted values (Lakshminarayanan et al., 2017). Using these estimated quantiles, the model creates an interval-based forecast (also referred to as a prediction interval): P(Yₐ ≤ ŷ ≤ Yᵦ) = 1 − α Using this type of forecast, supply chain managers are able to provide explicit uncertainty bounds for their demand or lead time predictions (Lakshminarayanan et al., 2017). As a result of processing the input vector and applying the inference model, the system ultimately provides a probabilistic forecast distribution instead of a single deterministic number. The critical risk metrics of Value-at-Risk, Conditional Value-at-Risk, and Interquartile Range can then be derived from this distribution to support decisionmaking at the operational level (Abdar et al., 2021). These measures allow planners to manage their inventory levels (including safety stocks) based on the organization's risk tolerance and desired service levels (Blundell et al., 2015). Strategically, this uncertainty-considered architecture enhances an organization's resilience (Abdar et al., 2021). It transforms forecasting from being solely about achieving high accuracy, to being about providing decision-makers with a risk-informed decision-support system that enables them to plan for potential adverse scenarios and to avoid overstocking and under-stocking of critical items (Kendall & Gal, 2017). By modeling and incorporating uncertainty into decision-making processes, organizations can develop more robust contingency plans, utilize resources more efficiently, and deliver better service reliability during uncertain and unstable periods (Gal & Ghahramani, 2016). Overall, this architecture does not aim to remove uncertainty from decision-making, but instead to measure, explain, and incorporate uncertainty into decision-making processes (Abdar et al., 2021). Thus, the transition from certaintybased prediction to probabilistic intelligence represents a fundamental advance toward developing truly resilient, risk aware, and adaptable supply chain management systems (Abdar et al., 2021). 14. Unified Predictive Intelligence Layer Diagram The Unified Predictive Intelligence Layer shown in Figure 13 represents a fully integrated engine for predicting demand and sensing risks, providing a comprehensive anticipatory view of future demand and disruption (Ivanov & Dolgui, 2020). Conceptually, the Unified Predictive Intelligence Layer functions as the cognitive core of the AI-driven supply chain, transforming disparate and noisy data into structured intelligence that enables proactive and risk aware decisionmaking (Aljohani et al., 2023). In contrast to separate analytic components, the Unified Predictive Intelligence Layer was designed as an integrated ecosystem to support real-time forecasting, adaptive learning and resilience planning (Hosseini Shekarabi et al., 2025). There are two major input channels at the beginning of the architecture. The first input channel contains all internal operational supply chain data, e.g., inventory levels, sales orders, supplier lead times and replenishment history, etc. These signals define the internal state of the network and provide the foundation for predicting demand and optimizing inventory (Aldahmani et al., 2024). The second input channel contains all exogenous data, i.e., geopolitical risk indices, climate patterns, macroeconomic indicators and market volatility etc. These signals define significant uncertainties and are generally underrepresented in traditional planning systems (Ivanov & Dolgui, 2020). All input channels have a common preprocessing stage where the raw time series data are cleansed, normalized, aligned temporally and transformed into high dimensional feature spaces (Tao et al., 2019) GSC Advanced Research and Reviews, 2025, 25(03), 078-126 102 Figure 13 Unified Predictive Intelligence Layer architecture Feature processing includes lagged observations, rolling statistics, seasonal encodings, volatility measures and regime indicators to make the temporal relationships between the data sets more apparent (Aljohani et al., 2023). The internal and external streams are then combined to form a single aggregate feature vector defined as: z ₜ=ϕ(x¹ⁿᵗ_(1 : t),xᵉˣᵗ_(1 : t)) Where x¹ⁿᵗ_(1 : t) and xᵉˣᵗ_(1 : t) represent the histories of the internal and external signal sources up to time t; and ϕ(·) is a nonlinear mapping of the internal and external features to produce a combined feature space (Aldahmani et al., 2024). Multiple hybrid models operate in parallel within the prediction component to generate candidate predictions based on the aggregated feature vector. Hybrid ARIMA-AI models identify the linear trend and seasonality from the data whereas the deep neural models, such as LSTMs and GRUs, identify long range nonlinear temporal dependencies in the data (Hammam et al., 2025). Temporal convolutional models identify local outliers resistant to changes in the pattern and the transformer models use temporal self attention to determine the relative importance of past events and exogenous shock to the current state (Tao et al., 2019). Additionally, there are separate models focused on capturing the interactions between the external conditions, such as weather anomaly and economic fluctuations, and the base line demand dynamics (Ivanov & Dolgui, 2020). Each model provides a different candidate prediction for the same feature representation. The candidate predictions are processed using a fusion framework that serves as a meta-learning layer. This layer uses residual analysis to detect persistent forecasting errors, performs drift detection when the data distributions change, and applies ensemble weighting to combine the candidates into a single, stabilized prediction (Hosseini Shekarabi et al., 2025). The weighted combination of the predictions is expressed as: ŷ ₜ = Σₖ₌₁ᴷ wₖ fₖ(z ₜ) Where fₖ(z ₜ) is the prediction generated by the k-th model and wₖ is the learned weight for that model, determined historically by its performance and currently by its relevance to the current regime (Hammam et al., 2025). This GSC Advanced Research and Reviews, 2025, 25(03), 078-126 109 speed and safety of autonomous decisions (negri et al., 2017). As these simulations become more synchronized with real operations, the control system gains greater foresight, further reducing operational uncertainty and reaction time. From a resilience perspective, this layer directly supports the three core attributes of adaptive capacity, redundancy and recovery effectiveness. Rather than responding only after disruption has occurred, the system is able to sense degradation trends early and take preventative action, strengthening the supply chain's ability to absorb shocks and recover faster (ponomarov & holcomb, 2009). Standardized resilience assessment approaches emphasize that proactive adjustment and response speed are fundamental dimensions of resilience effectiveness (pettit et al., 2013). By embedding these principles directly into an autonomous execution layer, resilience is transformed from a reactive concept into an operationalized, real-time function. Probabilistic feedback mechanisms further enhance this control capability. Demand and lead-time uncertainty are incorporated through autoregressive forecasting structures that continually update belief states as new data arrives. This allows the control layer to switch strategies automatically when volatility increases or when distributional shifts are detected (salinas et al., 2020). Instead of assuming static environments, the system continually re-evaluates the future landscape, adapting actions to align with changing probability distributions rather than outdated point estimates. In practical business terms, this autonomous and adaptive layer significantly reduces the time between disruption detection and corrective action. It improves asset utilization, minimizes idle time, improves on-time delivery and decreases overall system fragility. For organizations operating in highly volatile global conditions, this capability creates a critical competitive advantage. The supply chain is no longer constrained by slow reaction cycles or fragmented decision hierarchies. Instead, it evolves into a self-learning, self-correcting network that can anticipate, absorb and adapt to change continuously, supporting both operational strength and long-term resilience. 20. Integration with Enterprise and Cloud Ecosystems The point at which theoretical intelligence is realized in practical form is the integration of an AI-driven supply chain resilience framework into both enterprise and cloud systems. In essence, this layer takes the high-level capabilities of advanced predictive analysis, optimization and control and converts them into scalable, secure, and interoperable systems that may be deployed across the wide variety of digital environments present in today's organizations. Rather than being a separate analytical tool or engine that sits outside of the organization's IT environment, this framework is embedded directly into the organization's information environment. Therefore, the framework connects multiple sensors, operational systems, and digital planning systems together as part of a single adaptable system (Ben-Daya et al., 2019). As such, this integration reflects the underlying principles of the digital supply chain twin, where physical activities are constantly replicated and improved upon through a digital mirror image to improve responsiveness and resilience to disturbances (Ivanov & Dolgui, 2021). As noted above, cloud-edge architectures serve as the primary technology architecture for implementing distributed AI processing across the supply chain. Cloud architectures offer the scalable computing resources required to support the training of large, complex machine learning models, the flow of data across geographically dispersed locations, and the simulation of large numbers of variables and scenarios. Edge architectures, located near production and distribution locations, provide real-time decision-making capabilities at those points. This architecture provides the capability to make immediate decisions such as reroute shipments, reprioritize tasks, etc. at the point of operation; however, longer term strategic forecasting and optimization will continue to reside in the cloud (Wang et al., 2022). Collectively, the use of a hybrid cloud-edge architecture provides the organization with a fast, flexible, and contextually aware intelligence platform, which accurately reflects the distributed and highly interdependent nature of today's cyber-physical supply chains (Ivanov & Dolgui, 2021). A microservices-based architecture is used to create the ability to scale and adapt each component of the framework independently and reliably. Unlike traditional monolithic applications that integrate all aspects of the framework into a single, inflexible unit, the microservices architecture breaks the framework down into smaller, independent units called microservices, such as demand forecasting, disruption detection, route optimization, supplier risk assessment, and autonomous control. Each microservice has the ability to be scaled up or down independently, providing the organization with greater flexibility to respond to changes in demand and other environmental factors (Wang et al., 2022). Additionally, the independence of each microservice facilitates the ability to dynamically reconfigure the overall system, a critical requirement for systems that need to operate in uncertain and rapidly changing environments (Sarkis et al., 2011). GSC Advanced Research and Reviews, 2025, 25(03), 078-126 110 Each microservice is integrated with existing enterprise systems using Application Programming Interfaces (APIs) that provide seamless communication between AI capabilities and existing systems including Enterprise Resource Planning (ERP), Warehouse Management Systems (WMS), Transportation Management Systems (TMS), and Procurement Platforms. This allows AI-enhanced intelligent functions to be easily integrated into operational workflows that exist in the current systems, rather than requiring new systems to be implemented. For example, an enterprise purchasing platform can receive real-time ordering recommendations provided by the AI engine, and a TMS can obtain adaptive routing guidance provided by the AI engine based on the current status of the network (Ben-Daya et al., 2019). The integration of AI with existing systems also promotes the adoption of digital twins in enterprises. Digital twins represent the practice of embedding intelligence into existing operational systems to augment their performance and value (Wang et al., 2022). On a behavioral basis, the integration of AI with existing systems reduces the perception of technological disruption associated with adopting AI and increases the perceived utility and relevance of the AI. The architectural vision of enterprise-integrated supply chains must draw upon lessons learned from operationalizing AI in latency-sensitive and compute-constrained domains like gaming. Recent advancements in edge–cloud optimization, originally implemented for scalable game infrastructure, have become directly transferable to AI-native supply chains, especially those requiring real-time responsiveness and federated intelligence. By leveraging architectural patterns such as observability frameworks, dynamic edge scheduling, and latency–intelligence trade-off models, the framework enhances response accuracy across distributed environments (Chinnaraju, 2024a). These latency-aware edge-cloud mechanisms are particularly vital when synchronizing predictive AI pipelines with physical infrastructure constraints in logistics and manufacturing, ensuring both computational efficiency and resilience (Chinnaraju, 2024b). Furthermore, supply chain resilience increasingly depends on deploying AI models across heterogeneous platforms ranging from IoT sensors to web and mobile systems requiring rigorous benchmarking and operationalization standards. Integration of MLOps blueprints developed for high-volume game applications can accelerate the deployment of AI agents in enterprise environments. Key enablers include modular pipelines for real-time inference (Chinnaraju, 2024b), adaptive control loops triggered by infrastructure feedback (Chinnaraju, 2024c), and crossplatform execution leveraging ONNX Runtime, TVM, and WebAssembly for consistent performance (Chinnaraju, 2025a; Chinnaraju, 2025d). By incorporating these proven techniques, supply chain AI deployments become more robust, context-aware, and capable of real-time adaptation in volatile operating conditions, aligning with the core objectives of predictive intelligence and multimodal decision optimization. To protect against unauthorized access or manipulation of data flowing between edge devices, cloud services, and enterprise platforms, security and governance mechanisms are integrated throughout the interconnected environment. Encryption protocols, centralized identity management, and continuous anomaly detection are examples of the mechanisms that enforce the protection of data integrity, confidentiality, and operational reliability (Ben-Daya et al., 2019). Data ownership, usage rights, and compliance boundaries across different jurisdictions are defined by governance frameworks to ensure that digital transformations meet the expectations of ethics and regulations (Sarkis et al., 2011). Trust is a prerequisite for organizations to delegate real-time operational authority to AI systems. At the execution level, adaptive control mechanisms embedded within the ecosystem allow for automatic updates to strategic and operational policies. The adaptive control mechanism uses reinforcement learning principles to optimize actions such as rerouting, reprioritizing, or shifting capacity based on trial, feedback, and continuous improvement (Schulman et al., 2017). Over time, the organization's policies are transformed into more robust policies that reflect long-term business objectives such as cost efficiency, resilience, and sustainability. This dynamic learning loop transforms the organization's supply chain from a reactive process into a self-optimizing and anticipatory process. Strategically, the integration of an AI-driven supply chain resilience architecture into enterprise and cloud ecosystems allows an organization to be viewed as an intelligent and adaptive network. The organization is able to coordinate its operational processes more effectively, make more timely and effective decisions in response to events, and embed ongoing learning into the organization's culture and practices (Ivanov & Dolgui, 2021). Perhaps most importantly, the organization is able to transform its supply chain into a proactive and forward-looking system that is better positioned to anticipate disruptions, adapt quickly in response to disruptions, and sustain its competitive position in highly volatile and complex global markets. GSC Advanced Research and Reviews, 2025, 25(03), 078-126 111 21. Evaluation Metrics and Benchmarking Frameworks Evaluating and Benchmarking is the scientific discipline that validates whether the AI-driven Architecture for supply chain resilience, in terms of Technical Excellence, delivers meaningful Business Impact; and, without Systematic Measurement, the most Complex Systems for Forecasting and Optimizing remain Conceptual rather than Operational. This Study identifies Evaluation as both a Quantitative Process and Strategic Process intended to measure Improvements in Adaptability, Predictive Reliability, Operational Efficiency and Disruption Responsiveness within Highly Dynamic Supply Environments (Makridakis et al., 2018). At the Computational Level, primary emphasis is placed upon Predictive Performance; this includes Analyzing how Accurately the System Forecasts Demand for SKUs at the Item Level, Lead-Time Variation, Inventory Behavior, and Probability of Disruptions across Nodes in the Network. Notably, Accuracy is not viewed as a Single Static Outcome but is Evaluated Dynamically across Fluctuating Demand Patterns and Unstable Conditions to Capture Consistency During both Normal and High-Volatility Periods (Gunasekaran et al., 2017) and to Identify Tendencies toward Systematic Overestimation or Underestimation (Bias), which could result in Chronic Overstocking or Service Shortfalls if Corrected (Makridakis et al., 2018). While Traditional Prediction Metrics will be used to Evaluate Predictive Performance, this Research will introduce Disruption-Centric Indicators that Reflect Anticipatory Intelligence. The Disruption Prediction Reliability Score Measures how Early and Precise Potential Risks such as Labor Shortages, Supplier Failures, and Transportation Congestion are Identified; this is Critical since Contemporary Supply Chain Risk Landscapes are Complex and Multifaceted (Ghadge et al., 2012). Also, the Anomaly Detection Precision Rate Evaluates the Balance Between Sensitivity and False Alarm Reduction, a Critical Factor in Ensuring that Decision-Makers are not Overwhelmed by Unnecessary or Misleading Alerts (Heckmann et al., 2015). Optimization Performance is Another Major Dimension of Evaluation. This Study Measures how Effectively AI-Driven Decisions Improve Outcomes Compared to Conventional Approaches. Key Outcomes include Reductions in Excess Inventory, Improvement in Asset Utilization, Stabilization of Production Schedules, and Minimization of Waste. The Inventory Optimization Efficiency Ratio is particularly Important as it Represents the System's Ability to Balance Cost Reduction with Service-Level Reliability (Ho et al., 2015). Supporting this is the Dynamic Reorder Accuracy Metric, which Assesses the Alignment of AI Generated Replenishment Quantities with Real World Consumption Patterns Under Fluctuating Market Conditions (Gunasekaran et al., 2017). Resilience is Represented through Temporal and Structural Measures that Quantify System Recovery Capability. The Time-to-Recovery Index Calculates the Duration Required for a Disrupted Node or Process to Return to Equilibrium. A Shorter Recovery Time Directly Corresponds to Higher Resilience Under Stress Conditions (Ghadge et al., 2012). Similarly, the Supply Network Redundancy Coefficient Evaluates how Effectively Alternative Pathways and Backup Capacities Can be Activated, Representing Adaptive Strength in Complex Interconnected Systems (Heckmann et al., 2015). The Speed and Quality of Intelligence are Measured through the Adaptive Response Velocity Metric, which Captures How Quickly the System Sens Environmental Changes and Initiates Corrective Action. This is Particularly Critical in Real-Time Digital Environments Where Even Minor Delays Can Amplify System-Wide Consequences (Chae, 2015). Closely Related to this is the Autonomous Decision Effectiveness Score, which Measures the Quality and Stability of Decisions Generated Without Human Intervention. These Metrics Validate Whether Autonomous Mechanisms Can Be Trusted During Time-Sensitive Disruptions (Ho et al., 2015). Sustainability and Efficiency-Oriented Metrics Further Extend the Evaluation Scope. The AI-Driven Carbon Reduction Index Assesses how Optimized Routing and Demand Consolidation Reduce Unnecessary Transport Activities, Leading to Lower Emissions and Improved Environmental Performance (Sarkis et al., 2011). The Resource Utilization Efficiency Metric Examines how Intelligently the System Allocates Available Capacity Including Warehouse Space, Transportation Assets, and Production Facilities. These Measures Connect Technological Advancement with Long-Term Sustainability Responsibility (Gunasekaran et al., 2017). At the Organizational and Strategic Level, Composite Indicators Capture Holistic Transformation. The Resilience Maturity Score Integrates Operational, Structural and Predictive Metrics to Evaluate how Deeply Resilience is Embedded into Decision Making. This Reflects a Progression from Reactive Management Toward an Adaptive Data Driven Culture (Ghadge et al., 2012). The Decision Confidence Index is Introduced to Assess Managerial Trust in AI GSC Advanced Research and Reviews, 2025, 25(03), 078-126 112 Recommendations Based on Clarity, Consistency, and Transparency. This Aligns with the Growing Recognition that Psychological Trust and Perceived Reliability are Essential for Organizational Adoption (Chae, 2015). Benchmarking is Conducted Using Both Real-World and Synthetic Data Environments. Large Scale Demand Datasets Used in Global Forecasting Challenges Provide External Baselines for Performance Comparison (Makridakis et al., 2018). Social Media and Other Digital Data Signals Provide Contextual Stress Testing for Disruption Modeling (Chae, 2015). Also, Simulated Global Network Data Enables the Framework to be Evaluated Under Extreme Conditions Such as Geopolitical Instability, Climate Events, and Economic Turbulence, Supporting a Broader Generalization of Findings (Heckmann et al., 2015). Lastly, Performance is Compared Against Traditional Statistical Forecasting and Rule-Based Heuristic Systems. Long Established Models Such as Regression Based Approaches and Historical Smoothing Methods Serve as Baselines for Determining Relative Improvement. The Integration of Deep Learning Capabilities, Specifically Sequential Memory Based Architectures, Demonstrates Clear Advancement Beyond Traditional Temporal Dependency Limitations (Hochreiter & Schmidhuber, 1997). This Comparative Process Makes it Possible to Isolate the Tangible Performance Gains Resulting Directly From AI Integration. 22. Business Applications and Strategic Impact Business uses of the architecture of a predictive and adaptive supply chain driven by artificial intelligence will go well beyond incremental efficiency gains. Strategically, the integration of intelligent forecasting, autonomous optimization and real-time decision intelligence will create new ways to design, manage, and protect supply networks. Supply networks will be considered as strategic capabilities for companies to increase their competitiveness, responsiveness, sustainability and resilience in unstable markets (Tukamuhabwa et al., 2015) in place of being seen as simply reactive operational functions. This transformation signifies a move from linear operational thinking to systems thinking and that flexibility and learning are now key components of performance (Chowdhury & Quaddus, 2017). An example of one of the main applications of the architecture is to analyze and predict risks of suppliers through the means of a dynamic assessment of the risk associated with suppliers (risk analytics). Traditionally, the analysis of the risks related to suppliers depends mainly on historical data and periodic evaluations, which are unable to identify quickly changing realities. On the contrary, a system of artificial intelligence continues to evaluate and assess the behavior of suppliers in real time, by way of a number of indicators including variations in lead times, behavioral patterns of reliability, changes in geopolitical conditions, regulatory changes, environmental threats, etc. These inputs generate a dynamic and adaptable profile of the risk associated with each supplier instead of static categorizations (Fan & Stevenson, 2018). Through predictive modeling, the system does not only identify current vulnerabilities, but also predicts potential instabilities due to political tensions, climate disruptions, or economic pressures, thus increasing proactively the ability to govern the risk of the company (Dubey et al., 2022). The predictive nature of this system has the effect of transforming the procurement strategies of companies. Companies can no longer react to the failures of suppliers; they have the opportunity to diversify their sources of supply, to restructure their portfolio of suppliers, and to implement strategic buffers before the disruptions occur. This transformation leads to a slow evolution of the dependence of companies on single suppliers, towards a dependence on suppliers in a network, more robust and diversified (Brandon-Jones et al., 2014). From a theoretical point of view, this corresponds to the resource-based view, since the critical elements of supply are not only acquired, but are protected as essential resources of the organization that contribute to its resilience and to the generation of added value (Chowdhury & Quaddus, 2017). Another important area of application of the architecture concerns the balance between the demands of consumers and the supplies available to meet those demands, which is one of the most enduring structural problems in the management of supply chains. In the traditional systems, the static cycles of planning create frequently misalignments between the needs of the consumer and the availability of production, purchase and delivery. The AI-driven framework provides for continuous sensing of the demands of the consumer and allows the companies to dynamically adjust the production, purchase and delivery to the real time of the market. This approach directly facilitates the development of the capacity of adaptation and responsiveness, two of the most important aspects of the resilience of the supply chain (Tukamuhabwa et al., 2015). Instead of considering the demand forecasting as a periodic activity, the system transforms it into a permanent learning activity. Increases in demand can be identified early enough to allow the producers and suppliers to respond to them in advance. Decreases in demand can be identified in sufficient time to avoid excesses of stocks. This results in increased GSC Advanced Research and Reviews, 2025, 25(03), 078-126 113 rotation of inventory, less waste, and a better quality of services, which reinforce simultaneously the sustainability of the finances and the operations of the company (Dubey et al., 2022). The routing of deliveries and the automation of the inventory represent additional examples of the application of predictive intelligence to operational advantages. Decisions relating to routes were traditionally made in accordance with static rules and are continuously revised using in real time the information related to the traffic, the climatic conditions, the state of the infrastructures, the criteria of the fuel consumption, etc. These inputs enable the system to give priority to the transportations of high value-added and to redirect the flows of deliveries in case of disruption of the network (Fan & Stevenson, 2018). The inventory policies are also adapted in real time, and the quantities of the security stock are adjusted according to the volatility of the demand and the degree of vulnerability of the products, in order to define differentiated strategies for the products sensitive to the risks and for those with low sensitivity (Brandon-Jones et al., 2014). When observed through the lens of sustainability, the architecture is even more transformative. Companies are increasingly pressured to reduce their greenhouse gas emissions and to respect environmental regulations. By integrating the criteria of carbon footprint and environmental indicators into the optimization algorithms, the system ensures that the decisions taken into account not only the costs and the deadlines, but also the ecological impacts. This facilitates the adoption of practices of green purchasing, reductions in the emissions generated by transports, and the establishment of structures of supply chain more sustainable (Tukamuhabwa et al., 2015). These operational efficiencies produce a strategic advantage. Companies that deploy AI-driven, resilient supply chains do not only react more quickly than their competitors. They have the possibility to anticipate the disruptions and to maintain the continuity of the operations of the companies, while continuing to serve their clients, even during the periods of general crisis. This resilience creates an operational barrier that is difficult for the competitors to reproduce, particularly if the technologies used are supported by proprietary assets of data and systems of learning continuously evolving (Brandon-Jones et al., 2014). Finally, ethical and governance considerations further strengthen the importance of the business of the model. The use of AI in the processes of decision making generates issues relative to the accountability, the transparency and the correct utilization of the data. By incorporating the principles of ethics and the procedures of governance into the designs of the systems, the companies are able to combine the powerfulness of the technology with the sense of responsibility and the compliance to the regulations (Floridi & Taddeo, 2016). 23. Ethical and Governance Considerations in AI-Driven Supply Chains Artificial Intelligence (AI) has already transformed the world's supply chains and will continue to have an increasing impact moving forward. The incorporation of AI in Supply Chain Systems provides companies with new levels of capability, but also brings with it many complex ethical, organizational, and governance issues that need to be addressed as well as other secondary issues. As AI-driven architectures increasingly determine where goods are sourced, routed, inventoried and how suppliers are evaluated, they will not only shape the operation of the company, but also the lives of people, their environmental practices, regulatory compliance and the level of equity in the marketplace. Thus, the development of AI in supply chains needs to be built on a solid ethical base and include a defined governance structure that addresses the social, legal and environmental responsibilities of the technology advancements (Floridi & Cowls, 2019). In this regard, ethics and governance should be seen as the building blocks for developing a sustainable, transparent and trustworthy use of AI across all types of complex global networks (Jobin et al., 2019). One major ethical issue in automated decision-making systems is the lack of transparency and explainability in decisionmaking processes. Many traditional AI models are based on opaque "black box" algorithms which produce results without providing explanations for those results. In the context of supply chains, the lack of transparency in decisionmaking processes can result in serious consequences. For example, if an AI system de-prioritizes a supplier, changes shipping routes, or reduces production in a certain region, the affected entities may incur economic losses without knowing the reason behind the loss. This creates distrust among the stakeholders and raises questions about fairness and responsibility (Floridi & Cowls, 2019). To mitigate this risk, designers must incorporate transparency and explainability into their designs. Each important decision that is made by an AI system must provide a humaninterpretable explanation of the decision and the variables that led to the decision, along with real-time data that triggered the decision (Jobin et al., 2019). Therefore, Explainable AI is a bridge between complex computation and human understanding. It allows managers, partners and regulators to assess not only what decisions were made, but also why decisions were made. Explainable GSC Advanced Research and Reviews, 2025, 25(03), 078-126 114 AI therefore enhances the ability of stakeholders to oversee and monitor automated decision-making systems; diminishes the reliance on unbridled confidence in automation; and enables stakeholders to intervene ethically when automated decisions contradict the values of the organization or society. In the context of supply chains, explainable AI facilitates engagement and informed participation in AI-supported decision-making processes; and prevents stakeholders from completely relinquishing cognitive control to automation (Floridi & Cowls, 2019). Transparency, however, is only part of the equation; beyond explainability, governance frameworks establish the institutional framework that regulates how data and intelligence are created, deployed, monitored and corrected. Governance frameworks that are effective establish the owner of data; clarify the responsibility for the data; and define limits of the autonomous behavior of the system. At the organizational level, governance operates in multiple dimensions (technical, legal and human) to guarantee that models are sufficiently validated; that data are used legally; and that ultimate decision authority resides with accountable individuals rather than automated systems (Jobin et al., 2019). Sustainable sourcing represents another fundamental dimension of the ethical implications of AI-driven supply chain decision-making. The use of automated systems for assessing suppliers influences volumes of contracts awarded to suppliers; priority of suppliers; and long-term viability of suppliers. If these systems are trained using historical data that is biased, they may reproduce existing inequalities by giving preference to large, established suppliers and exclude small or socially-responsible suppliers. In order to avoid reproducing existing inequalities, sustainability, working conditions and environmental responsibility need to be incorporated into the logic of sourcing and not simply treated as secondary objectives to cost efficiency (Seuring & Müller, 2008). In this manner, the evaluation of suppliers evolves from a strictly economic function to a function that includes broader social and environmental accountability. Similarly, dual-sourcing and diversified procurement strategies acquire both ethical and resilience value in AI-based systems. Although dependence on a single source of supply may appear to be more efficient, it also exposes companies to greater risks of crises and ethical dilemmas related to the supply chain location. The integration of intelligent sourcing diversification allows companies to minimize their dependence on fragile systems while promoting stability throughout a larger network of suppliers and local communities (Sun & Van Mieghem, 2019). Thus, resilience and ethics are no longer mutually exclusive objectives, but are becoming complementary. Finally, the responsible application of AI to supply chains must be guided by a human-centric perspective. The introduction of automation leads to changes in the dynamics of workforces and may lead to the displacement of some workers if not introduced responsibly. Responsible implementation necessitates a broad range of training programs for re-skilling workers; open communication regarding the introduction of automation; and the phased introduction of automation in a manner that allows workers to adapt and evolve with intelligent systems rather than being displaced by them. This human-centered approach ensures that the knowledge and expertise of experienced professionals are retained and expanded upon rather than eliminated through automation (Jobin et al., 2019). Compliance with regulations is also a key component of implementing AI in supply chains. Because companies operating globally rely on supply chains that traverse a multitude of countries with different laws and regulations related to data protection, labor standards, environmental impacts, and trade practices, the development of AI systems must be done in a way that incorporates safeguards that protect privacy, prevent discriminatory outcomes and facilitate the lawful exchange of data. Failure to implement such safeguards may result in significant reputational, legal and social costs (Floridi & Cowls, 2019). Environmental governance is another factor that reinforces the ethical aspects of AI-driven supply chains. While decision systems may optimize for only speed and cost, they may inadvertently increase emissions and accelerate resource depletion. In order to counteract this tendency, environmental metrics such as carbon footprint, waste generation and energy consumption must be incorporated into the optimization criteria of decision systems rather than remaining as optional elements (Seuring & Müller, 2008). Thus, the use of technology to achieve efficiency must also be aligned with environmental sustainability. Furthermore, incorporating human values into the representation of human values in computational systems poses a much deeper philosophical question: How to represent human values in terms of computing? Human values are not universal and vary across cultures, industries and social settings. Therefore, ethical AI governance must be flexible; inclusive; and receptive to input from various stakeholders. Models of participative governance involving communities, regulators, suppliers and workers increase the legitimacy and social acceptability of AI and ensure that the values represented by AI reflect human values and not override them (Jobin et al., 2019). GSC Advanced Research and Reviews, 2025, 25(03), 078-126 115 In summary, this study demonstrates that the consideration of ethical and governance factors are essential components to the use of AI in supply chains. The inclusion of transparency, explainability, accountability, fairness, sustainability and compliance with law must be reflected at both technical and organizational levels. Ethical governance, when applied effectively, does not impede innovation; it increases trust; accelerates the adoption of AI; and ensures that AI contributes to both the effectiveness of supply chain operations and to the overall wellbeing of society (Floridi & Cowls, 2019). 23.1. Future Research Directions • A New Frontier in Supply Chain Resilience Research Opens as Artificial Intelligence, Advanced Analytics and Emerging Computational Paradigms Converge Rapidly The current supply chain resilience research architecture establishes a strong base using predictive modeling, adaptive optimization and autonomous decision systems. However, future research must go beyond incremental improvements to develop transformative technologies that establish new ways of thinking about conceptualizing, simulating, governing and evolving global supply networks. Therefore, future research will reimagine the supply chain as an intelligent, self-learning and multi-entity ecosystem that anticipates uncertainty, creates resilience across organizational boundaries and evolves continuously through advanced computational intelligence. Four promising and impactful research areas have emerged as the most promising for research and application: Quantum Enhanced Optimization, Federated AI for Collaborative Intelligence, Digital Twin Driven Predictive Simulation and Generative AI for Autonomous Scenario Construction and Resilience Forecasting. • Quantum Enhanced Optimization Presents Revolutionary Potential for Complex Supply Chain Problems Optimization techniques for complex supply chain systems, even those that employ heuristics or meta-heuristics, struggle to efficiently solve large-scale, multi-dimensional supply chain problems characterized by thousands of variables, constraints, interdependencies and uncertainty conditions. Examples of such problems include global route optimization across multiple transportation modes, multi-tier supplier selection with risk diversification, real-time demand-capacity balancing under uncertainty, and integrated cost-emissions minimization across international networks. Such problems typically fall within the realm of Combinatorial Optimization, in which the number of potential solutions increases exponentially as system size increases. Therefore, classical computers-even with advanced parallelization-are fundamentally unable to search large solution spaces effectively. • Quantum Computing Offers Radically Different Computational Approach for Evaluating Multiple Solution States Simultaneously A future supply chain research context will exploit quantum enhanced methods to enable real-time resolution of routing problems across millions of possible paths, optimal supplier network configurations balancing cost and risk, and dynamic inventory redistribution strategies currently beyond the capability of classical computers. Future research will focus on hybrid quantum-classical architectures, in which classical AI systems generate candidate solutions, and quantum processors refine them towards near-optimal states, providing unprecedented levels of efficiency, resilience, and environmental sustainability. • Development of Federated AI for Multi-Enterprise Collaborative Forecasting Will Represent Another Critical Direction Modern supply chains are decentralized networks of independent organizations including suppliers, manufacturers, distributors, logistics providers, and retailers, each possessing sensitive and proprietary data. One of the fundamental limitations of current predictive frameworks is the inability to safely and securely share data across these organizations. The lack of trust between organizations limits visibility and reduces the accuracy of global planning decisions. • Federated AI Provides Transformative Solution to Secure Sharing of Data Across Organizations Federated AI allows different participants to collaborate to train shared forecasting models without revealing their local data. Each organization trains a local version of the model on their own data set, and only encrypted parameter updates are exchanged globally to aggregate the local models, thus preserving data privacy while increasing collective predictive intelligence (Wang et al., 2022). Through this method, a more complete and accurate demand forecasting model can be developed without violating confidentiality or regulatory requirements. Future research will investigate how federated architectures can be applied to heterogenous data sets, inconsistent reporting standards, variable data quality, and GSC Advanced Research and Reviews, 2025, 25(03), 078-126 116 variable levels of computational preparedness among supply chain partners (Wang et al., 2022). Additionally, researchers will investigate the organizational and behavioral aspects of developing trust, establishing incentive mechanisms, and developing governance structures that support long-term interorganizational collaboration. • The Development of Digital Twin Technology Represents Another Powerful Frontier for Developing Resilient Systems and Practices A digital twin is a continuously updated virtual representation of the physical supply chain, which reflects its real-time state, structural dynamics, performance metrics, and evolving risk factors. Digital twins allow researchers and practitioners to analyze and experiment on the supply chain in a virtual environment without disrupting real-world operations (Wang et al., 2022). Future studies will focus on developing AI-embedded digital twins that reflect not only the current state of the system but also predict future states by combining forecasting models, reinforcement learning agents, and predictive analytics engines. This will allow researchers to simulate future disruptions such as climate-related events, geopolitical instability, demand surges, supplier failures and infrastructure failures in a safe, but highly realistic way (Wang et al., 2022). Through iterative scenario testing, organizations will be able to identify weaknesses in their systems, test mitigations for identified weaknesses and improve their systems' defenses before disruptions occur. From a theoretical standpoint, digital twins represent the intersection of Cyber Physical Systems, Systems Theory and Predictive Intelligence, creating a powerful living lab for conducting resilience engineering experiments. Generative AI for Autonomous Scenario Creation and Strategic Network Redesign Will Represent the Fourth Transformational Research Area. Traditional approaches to scenario planning rely on humans to think about possible future states, which limits the scope of the problem space because humans are influenced by cognitive biases and lack of knowledge. Generative AI can create hundreds of thousands of alternative future states by identifying patterns in historical data, emergent trends and probabilistic variability. The scenarios created by Generative AI can include different paths for various types of geopolitical risks, consumer behavior, climate volatility, energy scarcity, technological adoption and changes to regulation. AI-generated scenarios can be used to test the effectiveness of various strategies under the conditions represented in the scenarios, and can be passed through predictive and optimization models to determine how different strategies would perform under varying conditions. Thus, Generative AI will convert traditional static managerially focused scenario planning into a dynamic computationally-driven process that continually searches through all possible worlds. Furthermore, Generative AI can suggest entirely new network design options by systematically exploring combinations of supplier locations, transportation routes, warehouse placement and inventory strategies that maximize both resilience and efficiency. This parallels evolutionary system concepts in which adaptation, selection, and transformation lead to continued intelligent structural improvement of systems over time. Taken together, these research areas describe a future in which supply chains evolve to become self-designing, collaborative and cognitively-adaptive ecosystems. Quantum-enhanced optimization will unlock previously unachievable levels of computational efficiency. Federated AI will allow for large-scale, secure intelligence sharing across supply chain partners. Digital Twins will provide living simulation environments for continued learning and improvement. Generative AI will enable the creative, data-driven reconfiguration of network structures. When combined with the research architecture outlined in this paper, these advancements will not only enhance the resilience of global supply chains but also transform the nature of supply chains into intelligent, self-learning, and ethically-aligned global systems. From a broader theoretical perspective, these advancements represent a progression of supply chains from mechanistic infrastructures to complex adaptive organisms. Similar evolutions in the field of neuroscience, ecology and organizational learning theory show that intelligence arises through interaction, adaptation, feedback, and continued learning. This supports the premise of this paper that future resilience will arise not from the imposition of rigid control but from intelligent adaptability, distributed cognition and ethically-guided technological development. As research in these areas continues to advance, the global supply chain will evolve into a system that is not only more efficient, but more resilient, more equitable, more sustainable and better aligned with the long-term interests of business and society. 24. Conclusion This research has developed a complete, integrated, and forward-looking framework for developing the AI-driven supply chain resilience, and thus, redefined how today's supply chains are designed, controlled, and optimized in an era GSC Advanced Research and Reviews, 2025, 25(03), 078-126 117 of continuous uncertainty. Instead of seeing disruptions as external threats against which the supply chain must react, the study has re-conceptualized the ability to respond to disruption as an internal, computationally soluble and continuously improvable system characteristic that results from the intentional combination of predictive analytical, adaptive optimization, decisional intelligence and governance structures. Through the combined use of these structures, the supply chain can be converted from a passive, operational response mechanism to a pro-active, intelligent and strategic supply chain capability embedded in the heart of the organization. At its theoretical base, the research demonstrated a strong conceptual linkage between systems theory and dynamic network theory (e.g., Ijiri), and advanced AI paradigms (e.g., decision intelligence). The predictive intelligence capabilities of the system were based upon deep learning architectures including, but not limited to, LSTM, GRU, TCN, and Transformer-based models as well as hybrid ARIMA-Neural Systems, all of which enable the system to predict disruption patterns and changes in demand before those disruptions or demand changes result in operational failure. Unlike many studies that treat individual predictive models as separate technical tools, each of these models was considered part of a larger, unified intelligence fabric operating synergistically. To elevate predictive intelligence from surface-level pattern recognition to structural and probabilistic analysis, the study incorporated techniques such as causal modeling, anomaly detection, ensemble learning, and uncertainty quantification. These methods collectively increased the likelihood that the system could distinguish between random fluctuations in demand and meaningful structural changes in the demand environment. From an architectural perspective, the study demonstrated the potential of a multi-layered, closed loop system design that integrates data engineering pipeline functions, predictive intelligence functions, optimization engine functions, real-time decision interface functions, and autonomous control functions to create a continuous flow of adaptation where data is not passively consumed but instead actively transformed into decisions, actions, learning, and refinement. In addition, the study demonstrated the potential for this type of architecture to leverage scalable cloud-edge computing architectures and microservice orchestration to achieve both global scale and local responsiveness in support of distributed, heterogeneous, and multi-organization supply chain systems. Using real-time logistics, sourcing, scheduling, and inventory decisions as examples, the study demonstrated how reinforcement learning, multi-objective optimization, and adaptive policy iteration can be used to optimize and control real-time supply chain decision-making processes. The mechanisms presented in this study demonstrate that supply chain decision making does not have to occur using static rules or historical averages; rather, it can evolve as a selfcorrecting, experience-based system that learns from past successes and failures to continually improve its performance. A second major contribution of this study is the development of a specialized evaluation framework for assessing the effectiveness of AI-driven supply chain resilience. As described previously, traditional supply chain metrics were extended to include predictive accuracy, disruption prediction reliability, adaptive response velocity, supply network redundancy, resilience maturity scores, and AI-driven carbon reduction. These metrics provide a bridge between the performance of algorithms and business outcomes and provide a structure for organizations to assess their supply chain performance and compare that performance to other organizations' performance. Although this study demonstrates the potential for significant technical innovations in the area of AI-driven supply chain resilience, the study also illustrates the direct impact that this technology can have on the business practice of companies. Specifically, this technology has the potential to transform the practice of resilient sourcing, real-time demand-supply matching, intelligent logistics routing, dynamic inventory stratification, and ESG-aligned decisionmaking. Furthermore, by incorporating sustainability and social responsibility into the optimization process, the study demonstrates the ability of this technology to move beyond the traditional focus on profit maximization and promote a more responsible approach to value creation. This focus on sustainability and social responsibility is consistent with emerging global trends toward environmentally and socially responsible corporate practices. In addition to demonstrating the technical feasibility of AI-driven supply chain resilience and illustrating its direct impact on the business practice of companies, this study emphasizes the importance of ethical integrity and governance in the use of AI technology in supply chain management. Given the influence of AI in supply chain management on communities, environments, and economies, the study incorporates explainability, transparency, accountability frameworks, and social responsibility standards into the design of the AI technology. Therefore, the study establishes a requirement that human oversight, participatory governance, and algorithmic responsibility must be implemented as structural requirements for the sustainable deployment of AI technology in supply chains. Thus, this study addresses a significant gap in current AI-centric supply chain models. GSC Advanced Research and Reviews, 2025, 25(03), 078-126 118 Finally, the study concludes with a discussion of future research opportunities related to the proposed architecture. For example, recent advances in quantum-enhanced optimization, federated AI, digital twins, and generative scenario modeling suggest that supply chains may eventually develop into fully cognitive, self-designing networks. Although these future directions do not contradict the study's proposed architecture, they represent a natural extension of the study's proposed architecture and expand the scope of its applicability beyond the current state-of-the-art in AI and machine learning technologies. In summary, this study represents a new paradigm for supply chain systems. The study advocates that the supply chain should no longer be viewed as simply a passive connector of products but as a living, intelligent being that can learn, adapt, and evolve in real-time. By providing a rich framework for the study of AI-driven supply chain resilience, the study provides scholars with a framework for future research and provides practitioners with a blueprint for creating competitive, resilient, and responsible global supply chains. Ultimately, the study affirms that supply chain resilience is no longer dependent upon chance or historical strength. Instead, it is an engineered outcome of intelligent design, adaptive learning, and strategic foresight. As the world continues to evolve in an uncertain environment, organizations that can develop and implement intelligent supply chain resilience through AI-driven architectures will not only survive disruption but will lead the way in the next generation of global innovation and sustainable enterprise. Compliance with ethical standards Disclosure of conflict of interest The authors declare that they have no conflicts of interest related to this study.” References [1] Abdar, M., et al. (2021). A review of uncertainty quantification in deep learning: Techniques, applications and challenges. Information Fusion, 76, 243–297. https://doi.org/10.1016/j.inffus.2021.05.008 [2] Abudu, R., Thompson, E. A., Dzawu, F. S., & Roa-Henriquez, A. (2025). A meta-analysis assessment of adaptive and transformative approaches to supply chain resilience. Supply Chain Analytics, 10, 100124. https://doi.org/10.1016/j.sca.2025.100124 [3] Adewumi, I. O., & Oyekunle, V. B. (2025). Lightweight deep learning for agricultural loss forecasting: GRU with transfer learning and edge deployment. Biosys Journal of Biotechnology & Bioengineering, 1(2), 1–10. https://doi.org/10.21203/rs.3.rs-7370153/v1 [4] Ahumada, O., & Villalobos, J. R. (2009). Application of planning models in the agri-food supply chain: A review. European Journal of Operational Research, 196(1), 1–20. https://doi.org/10.1016/j.ejor.2008.02.014 [5] Akbas, T., Keles, M., & Ertugrul, A. (2022). A hybrid forecasting model using LSTM and Prophet for energy consumption with decomposition of time series data. PeerJ Computer Science, 8, e1001. https://doi.org/10.7717/peerj-cs.1001 [6] Aldahmani, E., Alzubi, A., & Iyiola, K. (2024). Demand forecasting in supply chain using Uni-Regression Deep Approximate Forecasting Model. Applied Sciences, 14(18), 8110. https://doi.org/10.3390/app14188110 [7] Aljohani, A., Qureshi, M. R. N., Aljahdali, H., & Dabaghzadeh, F. (2023). Predictive analytics and machine learning for real-time supply chain risk management. Sustainability, 15(20), 15088. https://doi.org/10.3390/su152015088 [8] Aravind Chinnaraju. (2024). Optimizing Latency and Intelligence Trade-Offs in AI-Driven Games: Edge-Cloud Architectures, Scheduling Policies, and Observability Frameworks Aravind Chinnaraju. International Journal of Advanced Research in Science Communication and Technology, 647. https://doi.org/10.48175/ijarsct-17097 [9] Aravind Chinnaraju. (2025). Operationalizing AI in game development: MLOps infrastructure patterns and frontline insights. International Journal of Science and Research Archive, 15(2), 081–101. https://doi.org/10.30574/ijsra.2025.15.2.1288 [10] Aravind Chinnaraju. (2024). Real Time Adaptive AI pipelines for edge cloud systems: Dynamic optimization based on infrastructure feedback. World Journal of Advanced Engineering Technology and Sciences, 13(2), 887– 908. https://doi.org/10.30574/wjaets.2024.13.2.0636 GSC Advanced Research and Reviews, 2025, 25(03), 078-126 125 [127] Vaswani, A., Shazeer, N., Parmar, N., et al. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30. https://doi.org/10.48550/arXiv.1706.03762 [128] Vaswani, A., Shazeer, N., Parmer, N., Uszkoreit, J., Jones, L., Gomez, A. N., … Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30, 5998–6008. https://doi.org/10.48550/arXiv.1706.03762 [129] Wamba, S. F., Dubey, R., Gunasekaran, A., & Akter, S. (2020). The performance effects of big data analytics and supply chain ambidexterity: The moderating effect of environmental dynamism. International Journal of Production Economics, 222, 107498. https://doi.org/10.1016/j.ijpe.2019.09.019 [130] Wan, R., Mei, S., Wang, J., Liu, M., & Yang, F. (2019). Multivariate temporal convolutional network: A deep neural networks approach for multivariate time series forecasting. Electronics, 8(8), 876. https://doi.org/10.3390/electronics8080876 [131] Wang, H., Xie, F., Duan, Q., & Li, J. (2022). Federated learning for supply chain demand forecasting. Mathematical Problems in Engineering, 2022, 4109070. https://doi.org/10.1155/2022/4109070 [132] Wang, J., & Patel, P. (2023). Employing long short-term memory and Facebook Prophet model in environmental forecasting. Communications in Statistics - Simulation and Computation, 52(9), 4574-4592. https://doi.org/10.1080/03610918.2020.1854302 [133] Wang, L., Deng, T., Shen, Z.-J. M., Hu, H., & Qi, Y. (2022). Digital twin-driven smart supply chain: New research opportunities. Frontiers of Engineering Management, 9(1), 56–70. https://doi.org/10.1007/s42524-021-01869 [134] Wang, L., Deng, T., Shen, Z.-J. M., Hu, H., & Qi, Y. (2022). Digital twin-driven smart supply chain: New research opportunities. Frontiers of Engineering Management, 9(1), 56–70. https://doi.org/10.1007/s42524-021-01869 [135] Wang, Q. (2025). A Hybrid Transformer-ARIMA Model for Forecasting Global Supply Chain Disruptions Using Multimodal Data. International Journal of Advanced Computer Science and Applications, 16(1). https://doi.org/10.14569/ijacsa.2025.0160153 [136] Wang, T.-Y., & Yeh, D.-H. (2014). An Integrated Forecasting DSS Architecture in Supply Chain Management. Operations and Supply Chain Management: An International Journal, 24–41. https://doi.org/10.31387/oscm030120 [137] Wen, R., et al. (2017). A multi-horizon quantile recurrent forecaster. arXiv preprint arXiv:1711.11053. https://doi.org/10.48550/arXiv.1711.11053 [138] Wu, H., Xu, J., Wang, J., & Long, M. (2021). Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. NeurIPS 2021. https://doi.org/10.48550/arXiv.2106.13008 [139] Xiao Yang and Lan Li. 2025. Quantifying the Impact of Digital Transformation on Supply Chain Resilience: A Machine Learning-Based SEM Study in China's High-End Manufacturing Sector. In Proceedings of the 2025 International Conference on Big Data, Artificial Intelligence and Digital Economy (BDAIE '25). Association for Computing Machinery, New York, NY, USA, 86–91. https://doi.org/10.1145/3767052.3767066 [140] Xun, H., & Li, W. (2024). The supply chain demand forecasting model based on LSTM and multiple clustering techniques. Highlights in Science, Engineering and Technology, 101, 390–394. https://doi.org/10.54097/y5c6pb47 [141] Yoon, J., & Kim, S. (2023). Triple bottom line aspects and sustainable supply chain resilience: A structural equation modeling approach. Frontiers in Environmental Science, 11, 1161437. https://doi.org/10.3389/fenvs.2023.1161437 [142] Yuniarti, R., Suparno, & Arvitrida, N. I. (2025). A scoping review and bibliometric analysis of sustainable and resilient supply chain network design. Supply Chain Analytics, 12, 100162. https://doi.org/10.1016/j.sca.2025.100162 [143] Zamani, E. D., et al. (2023). Artificial intelligence and big data analytics for supply chain resilience: A systematic literature review. Annals of Operations Research, 327(2), 605–632. https://doi.org/10.1007/s10479-02204983-y GSC Advanced Research and Reviews, 2025, 25(03), 078-126 126 [144] Zamani, E. D., Smyth, C., Gupta, S., & Dennehy, D. (2023). Artificial intelligence and big data analytics for supply chain resilience: A systematic literature review. Annals of Operations Research, 327(2), 605–632. https://doi.org/10.1007/s10479-022-04983-y [145] Zamani, E. D., Smyth, C., Gupta, S., & Dennehy, D. (2023). Artificial intelligence and big data analytics for supply chain resilience: A systematic literature review. Annals of Operations Research, 327(2), 605–632. https://doi.org/10.1007/s10479-022-04983-y [146] Zerveas, G., et al. (2021). A transformer-based framework for multivariate time series representation learning. Proceedings of KDD 2021, 2114–2124. https://doi.org/10.1145/3447548.3467401 [147] Zhang, W., Liu, J., Deng, W., Tang, S., Yang, F., Han, Y., & Liu, M. (2024). AMTCN: An attention-based multivariate temporal convolutional network for electricity consumption prediction. Electronics, 13(20), 4080. https://doi.org/10.3390/electronics13204080 [148] Zhou, H., et al. (2021). Informer: Beyond efficient transformer for long sequence time-series forecasting. AAAI 2021, 35(12), 11106–11115. https://doi.org/10.1609/aaai.v35i12.17325