Reimagine Power Electronics Design with Artificial Intelligence (AI)

ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li Reimagine Power Electronics Design with Artificial Intelligence (AI) 1University of Bath 2University of Arkansas 3Zhejiang University-University of Illinois Urbana-Champaign Institute (ZJUI) Fanfan [email protected] Peter [email protected] Xinze [email protected] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li Speaker Biographies 2/145 Xinze Li was awarded the Ph.D. degree in Electrical and Electronic Engineering by Nanyang Technological University, Singapore, 2023. He gained AI industry experience in computer vision with Singtel (Singapore) and AI research experience in natural language processing. He joined the University of Arkansas, USA, as a postdoctoral research fellow in 2024. His research interests include power converter design automation, light and explainable AI for power electronics with physics-informed systems, and fault prognosis and health management for power semiconductors. Fanfan Lin received the joint Ph.D. degree with the interdisciplinary research in power electronics and artificial intelligence (Al) from Nanyang Technological University (NTU) Singapore, and the Technical University of Denmark, Denmark, in 2023. She is now with Zhejiang University -University of Illinois Urbana-Champaign Institute as Assistant Professor. Prior to her time in academia, Dr. Lin was an AI startup founder and was part of the Sequoia Capital (Southeast Asia) Spark Fellowship Program. Her research interests include power converter design with artificial intelligence, life cycle management, and responsible Al for energy systems. Peter Wilson is currently a Full Professor of electronic engineering with the University of Bath, U.K. He is also serving as an IEEE PELS VP for Industry and Standards and Member-at-Large. He has published widely in power electronics and related fields with more than 150 articles, several patents, authored a number of books. His research interests include modeling and simulation, magnetics, design automation (specifically computational intelligence techniques including genetic algorithms and machine learning), systems engineering, robotics, open-source tools and methods, and embedded systems. He received the IEEE Standards Medallion in 2020. He was the Chair of the ITRW Roadmap and IEEE Std 1573. He is a fellow of IET, the British Computer Society, and the Higher Education Academy. ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li Seminar Outline 3/145 1. Applications of AI in Power Electronics Design: An Overview AI in the Life-Cycle Management of Power Converters Fundamentals of AI-Based Power Electronics Design 2. Hands-on Practice: Modulation Design for DAB Converters Current Stress Modeling as Regression and ZVS Modeling as Classification Combines Data-Driven Surrogate Models with Meta-Heuristic Algorithm 30-Minute Break 3. Emerging AI Trends in Power Electronics Design Physics-Informed Machine Learning in PE Design PE-GPT: A Generative AI-Based PE Design Paradigm Reinforcement Learning, AI Explainability, Graph Neural Network 4. Ethical Challenges of AI Applied in Power Electronics Significance of AI Ethics in Power Electronics Four Pillars of AI Ethics in Energy Conversion ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 1. Learning Objectives –Section 1 4/145 Overview of Artificial Intelligence (AI) in Power Electronics Explore the application of AI across the life cycle of power electronics systems, including offline modeling and design, as well as online control and maintenance. Fundamental Concepts of Power Electronics Design Explore the principles of model-based design (V-Diagram), design space, performance space, and hierarchical steps in power converter design. Basics of AI in Power Electronics Design How are machine learning and meta-heuristic algorithms applied in power electronics design. Applications of AI in Power Electronics Design Examine the present-day applications of AI in power electronics design, highlighting its impact and practical implementations. ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 1.1 AI in the Life-Cycle Management of Power Converters 5/145 Ref: [1] *Data retrieved from IEEE Xplore on 23 September 2025 using the query: ("All Metadata": AI) OR ("All Metadata": "artificial intelligence") OR ("All Metadata": "deep learning") OR ("All Metadata": "machine learning"). •Asearch across the IEEE Power Electronics Society (PELS) portfolio, shows that the number of AI-related papers published between 2020 and 2025 has increased around fourfold. ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 1.1 AI in the Life-Cycle Management of Power Converters 6/145 Ref: [1] AI in the Life-Cycle Phases of Power Converters: Design, Control, and Maintenance. Source: Shuai Zhao PE Applications FunctionAI Algorithms 9.8% 77.8% 12.4% ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 1.1 AI in the Life-Cycle Management of Power Converters 7/145 Bayesian Network: Quantify Uncertainty Reinforcement Learning: Optimization & Control Ref: [1] Most Widely Applied in Power Electronics Timeline of AI Methods in PE: Expert System and Fuzzy Logic have Limited Developments, Meta-Heuristic Algorithms Show Significant Adoption in Optimization, and Machine Learning Reflects Most Sustained and Diverse Growth. ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li Explore Design Solutions and Optimize 1.1 AI in the Life-Cycle Management of Power Converters 8/145 1. Circuit Parameter Design 2. Control Implementation 3. Feedback Experimental Data to Update Built Models Identify Design Parameters and Objectives Build Circuit Models Optimize Converter Designs Build Control Models Controller Optimization and Design Online Implementation Collect Operation Data Update Models if Outdated Evaluate the Existing Models To fine-tune or retrain models Simulation Data Performance Analysis Circuit Physics Expert Knowledge Hybrid Data-driven and Knowledge-based Offline Surrogate Models • Preprocessing • Feature Extraction • Data-driven Modeling • Fitness / Objective functions Preliminary: Offline Model Building for Circuit or Controller Trained control models Trained circuit models Generic Workflow of AI in Design and Control 2. AI as Controllers ( ) ( ) ( ) ( ) { } ( ) ( ) 123 min min , , , , 0, 0 X X lu fX fXfXfX X XX gX hX  ∈ ≤=  subject to Formulate Objectives and Constraints Evaluate and Iterate Formulate f, g, and h 1. AI as Surrogate Models ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 1.1 AI in the Life-Cycle Management of Power Converters 9/145 Ref: [1] 1. AI as Fault Classifiers 2. AI as RUL Estimators AI Functions as Fault Classifiers for Detection and Diagnosis, RUL Estimators for Predicting Remaining Useful Life, and Decision-Makers for System-Level Predictive Maintenance and Control. Generic Workflow of AI in Health and Condition Monitoring of Power Converters ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li Past: Traditional way to design power converters –Human in the loop Source: T. Hudson and M. Ametller, “How to Design aFlyback Converter in Seven Steps,” Online,https://www.monolithicpower.com/learning/resources/howto-design-a-flyback-converter-in-seven-steps?srsltid=AfmBOoo5JmyitoRlsuqjPF59bNXmkm5ZQzIjadbx-R-U2ExEVeorFjJC. Modeling Simulation Iterate Iterate x1 Iterate x5 Iterate x99 1.3 Traditional Way to Design Flyback Converters 16/145 ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li The future:A generative AI-based design paradigm ØPast: Engineers do all the work. ØFuture: With tools such as PE-GPT, design through natural language queries. Request in natural language Source: F. Lin et al., “PE-GPT: A New Paradigm for Power Electronics Design,” IEEE Trans.Ind.Electron., pp. 1–14,2024. Designer Generative AI Validate and iterate Generate •Modeling •Parameter design •Component selection •Simulation validation •Virtual prototyping … Full design details Automated! 1.3 Future Way to Design Flyback Converters 17/145 Ideally, engineers may increasingly act as system orchestrators, specifying requirements and validating. ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 1.3 Flyback Converter Design with PE-GPT PE–GPT Source: F. Lin et al., “PE-GPT: A New Paradigm for Power Electronics Design,” IEEE Trans. Ind. Electron., pp. 1–14, 2024. 18/145 Ref: [6] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 1.4 Towards Power Electronics Design Automation Advantages of AI Faster time-to-market design cycle Automate repetitive tasks, free labors More comprehensive co-designs Push SOTA performance boundaries Lower entrance barrier, generic New Tech Drive Progress, while Market Demand Shapes Adoption. New Materials and Wide Bandgap Devices Embedded Systems Time2010 2025 Technology Push AI will not replace engineers, but those who know AI will. 19/145 Ref: [4] Source:Johann W. Kolar ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 1.4 Towards Power Electronics Design Automation AI-driven modeling, optimization PE Design 4.0 - PE Autonomous Design PE Design 3.0 AI-Augmented Design Computer aided design and simulation PE Design 2.0 - Assisted Design Manual computation, analysis, and design Manual Design PE Design 1.0 Evaluated Performance New design Machine Learning Meta-Heuristic Algorithm 20/145 ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 1.4 PE Design 2.0 –a Case Study A simple design case •Design targets: inductor, capacitor •Objectives: Optimal efficiency •Constraints: size, ripple ( ) _1 _2 _ _ :{ , } , min min subject to: , % ( %) . == + ++ ≤ ∆ ≤∆ ls ls lL lC X LC LC LC lim L L lim fX P P P P Vol Vol II Pl_s1(L, C)Pl_s2(L, C) Pl_L(L, C)Pl_C(L, C) VolLC(L, C)ΔIL%(L, C) Pl_s1Pl_s2 Pl_L Pl_C VolLC ΔIL% C RL ESR L ESL High-side S1 Lowside S 2 Design LC filter for buck Formulate PE Design 2.0 21/145 ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 1.4 PE Design 2.0 – MHA for Optimization S T A R T Initialize Design Objective Function Evaluation Generate New Design Meta-Heuristically Navigate the Search Trajectory Compare, Select, and Update Candidate Pools Stop? Performance space N Optimized designs Design space Effi. Density Param. 1 Param. 2 Evolve towards Meta-heuristic algorithm (MHA) for converter optimization: New design candidates are meta-heuristically generated (like genetic mutation, crossover), objective functions guide the selection towards optima. 22/145 ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 1.4 PE Design 2.0 – Meta-Heuristic Algorithm (MHA) PE design 2.0: Boost performance with MHAs Analytical models Yt= f(Xt) Performance Evaluation New design solutions Xt+1 Design solutions Xt Efficiency, volume, ripple MHA for optimization Xt, YtXt+1 Iteration t … Design Xt Pl_s1+Pl_s2 +Pl_L+Pl_C VolLC ΔIL Design Xt+1 220 μFT80-75 core, 22 turns T106-75 core, 48 turns 2x100 μF Larger core, more winding turns, cap in parallel… Complicated? Inaccurate? � 23/145 ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 1.5 PE Design 3.0 –ML for Modeling Performance Space Y Concepts of Machine learning (ML) and their classifications Design space X Knowledge of loss breakdown f(X) Analytical approach Y = f(X) Explicitly programmed Performance space Y Machine learning Y = fθ(X) Performance space Y Design space X Identified pattern from data fθ(X) Data-driven Predict Design Space X Topology Modulation Power module EMI filter, etc. Efficiency Density Reliability Cost, etc. Evaluate PE Design 3.0 24/145 ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 1.5 PE Design 3.0 –ML for Modeling 25/145 NN for the maximum junction temperature modeling of diverse power module layouts Maximum Junction Temperature 105 ℃ Design 1 Design 2 ML for the performance modeling of diverse power modules ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 1.6 PE Design 3.0: AI in Magnetics 32/145 Automated Data Acquisition System of MagNet MagNet ML Framework for Modeling Magnetic Materials Acquired Standard Dataset Case 2: MagNet Challenge: Magnetic Materials (B-H and Loss Curves) Modeling via Machine Learning. Ref: [9] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 1.6 PE Design 3.0: AI in Magnetics MagNet Challenge 2023 Outcomes Convolutional Network Developed by Paderborn Case 2: The MagNet Challenge Advanced the “Model Size –Accuracy” towards More Accurate AI with Lighter Size. 33/145 Ref: [9] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 1.6 PE Design 3.0: AI in PCB Routing 34/145 Case 3. Convolutional Neural Networks for Thermal-Driven PCB Routing, Using 2-D Thermal Distribution Fields as Image Data for AI-Based Thermal Modeling. Thermal-Driven PCB Routing Workflow Thermal Distribution Prediction Model Ref: [10] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 1.6 PE Design 3.0: AI in PCB Routing 35/145 Hotspot Temperature 2℃↓ Thermal Modeling Error (℃) Case 3: An Automatically Routed PCB and its Thermal Distribution Ref: [10] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 1.7 Key Takeaways –Section 1 36/145 Applications of AI in Power Electronics AI in the design, control, maintenance of power converters Main functionalities of AI: Surrogate models, controllers, optimizers, fault classifiers, and remaining useful life estimators Model-Based Design in Power Electronics Top-down design: System -> Converter -> Component -> Material Design space (design variables) and performance space (performance metrics) AI in Power Electronics Design Meta-heuristic algorithm: Optimization of power converters Machine learning: Behavioral modeling of power converters Generative AI: Design automation ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li Seminar Outline 37/145 1. Applications of AI in Power Electronics Design: An Overview AI in the Life-Cycle Management of Power Converters Fundamentals of AI-Based Power Electronics Design 2. Hands-on Practice: Modulation Design for DAB Converters Current Stress Modeling as Regression and ZVS Modeling as Classification Combines Data-Driven Surrogate Models with Meta-Heuristic Algorithm 30-Minute Break 3. Emerging AI Trends in Power Electronics Design Physics-Informed Machine Learning in PE Design PE-GPT: A Generative AI-Based PE Design Paradigm Reinforcement Learning, AI Explainability, Graph Neural Network 4. Ethical Challenges of AI Applied in Power Electronics Significance of AI Ethics in Power Electronics Four Pillars of AI Ethics in Energy Conversion ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 2. Learning Objectives –Section 2 38/145 Hands-on Experience of a Data-Driven PE Design Task Learn key AI techniques via a step-by-step walk through of a modulation design case for DAB converters. You will learn basic data analysis, NN for regression and classification and how to improve the accuracy of NN, meta-heuristic algorithms for converter optimization, and easy-to-use anytime ensemble learning algorithms. Develop an Algorithm Mindset Explore insights from embedded feature space of data, learn how to adjust NN model head to adapt to different learning paradigms (regression and classification), understand some intrinsic invariants of tabular data, and integrate NN into meta-heuristic algorithms. Fundamentals of Neural Networks and Meta-Heuristic Algorithms Learn the key principles behind NN and meta-heuristic algorithms, such as NN structure, learning paradigms, and its regularization, and the balance between global exploration and local exploitation of MHAs. ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 2.1 Try Out PE Design 3.0 by Yourself! 39/145 Let’s explore a PE Design 3.0 case —on your laptop! 💻💻 Step 2: Scroll down and click “Open in Colab” https://github.com/XinzeLee/ECCE2025 Step 1: Visit the link below to start. Step 3: Sign in with your own google account! ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 2.1 Try Out PE Design 3.0 by Yourself! 40/145 Try out a PE Design 3.0 case —with 3 steps in sequence! Step 1: Exploratory data analysis Select features, gain design insights, and clean data from raw data. Step 2: Machine learning for modeling Use neural networks to build surrogate models for converter performances. Step 3: Meta-heuristic algorithm for optimization Search for optimal design parameters •Step 1: Visit the link below to start. https://github.com/XinzeLee/ECCE2025 •Step 2: Scroll down and click “Open in Colab” •Step 3: Sign in with your own google account! ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 2.2 Hands-on Design Case Background 41/145 Triple Phase Shift Modulation design for DAB converters Triple phase shift modulation for the DAB converter vi = 200 V vo∈[160 V, 240 V] fs= 50 kHz PL∈[100 W, 1 kW] Ref: [11]-[12] Dual Active Bridge (DAB) Converter Design Task: Determine modulation parameters D1, D2 Objective: Minimize current stress (ipp) and achieve all-switch zero voltage switching (ZVS) ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 2.2 Regularization in Machine Learning 48/145 Avoid overfitting in NN: Add regularization terms to loss functions ØL2 regularization: Penalize large magnitudes of weights ØL1 regularization: Zero some weights, producing sparse models ØL1 regularizationØL2 regularization Optimal solutions occur at the point of tangency! (Lagrange multiplier) ℒ′=ℒ+𝜆𝜆𝒘𝒘𝑇𝑇𝒘𝒘ℒ′=ℒ+𝜆𝜆 𝒘𝒘 1 Ref: [15] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 2.2 Regularization in Machine Learning 49/145 Try it yourself! Avoid overfitting in NN: Add regularization terms to loss functions ØL2 regularization: Penalize large magnitudes of weights ØL1 regularization: Zero some weights, producing sparse models Q: Which one uses L1 or L2 regularization? Why? L1 regularization L2 regularization *URL playground.tensorflow.org/ Ref: [16] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 2.3 Compared to Ensemble Learning Algorithms 50/145 Benchmark with easy-to-use anytime ensemble learning algorithms: XGBoost Animation of gradient boosting trees Ref: [17] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 2.3 Why XGBoost Outperforms FNN on Tabular Data? 51/145 Ref: [18] Benchmark on medium-sized datasets (numerical features only) Benchmark on medium-sized datasets (numerical and categorical) Columns := Features / Attributes Tabular data (structured) Feature 1 Feature 3 Feature D … Rows := Data samples Q: Can we switch two columns, and they still remain meaningful? Voltage Current Tabular data formats are structured, which are not invariant to feature rotation (swapping) ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 2.3 Why XGBoost Outperforms FNN on Tabular Data? 52/145 Weight-space symmetries in neural networks ØInvariant by rotation ØSign-flip symmetry ØTabular data is not invariant by rotation 0.1 -0.3 0.2 -0.3 0.1 0.2 0.3 -0.1 -0.2 … In total, 𝑀𝑀!�2𝑀𝑀combinations of the same NN. Feature 1 Feature 2 Feature 2 Feature 1 Swap the feature… The NN can still be the same by permutating the weights. Sign-flip symmetry if activation function is zero-centered and symmetrical (like tanh). Needs more efforts to find feature correlations in tabular data. Ref: [18] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 2.3 Why XGBoost Outperforms FNN on Tabular Data? 53/145 Proficiency of tree-based models on tabular data ØTree-based models are also not invariant by rotation Latent pattern (priori) aligns with tabular data. Q: Can FNNs ever win on tabular data? True or false cannot describe humidity. ØMore robust to uninformative features Test accuracy changes when removing or adding uninformative features Yes! But only with care and engineering. Good practices to develop your NN. Ref: [18] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 2.4 PSO to Optimize Current Stress and Soft Switching 54/145 Main workflow of PSO algorithms: Overall diagram START Initialize all particles Evaluate objective Update particle xi Update pbiand gb Calculate velocity vi Update hyperparameters Stop? ( ) 12 , ,, max(8 , 0) obj L out pp ZVS ZVS f PV DD in λ = +− ⋅ ipp nZVS 𝑣𝑣𝑖𝑖 (𝑘𝑘+1)=𝜔𝜔𝑣𝑣𝑖𝑖 (𝑘𝑘)+𝑐𝑐1𝑟𝑟1𝑝𝑝𝑝𝑝𝑖𝑖 (𝑘𝑘)− 𝑥𝑥𝑖𝑖 (𝑘𝑘) +𝑐𝑐2𝑟𝑟2𝑔𝑔𝑝𝑝(𝑘𝑘)− 𝑥𝑥𝑖𝑖 (𝑘𝑘) ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 2.4 PSO to Optimize Current Stress and Soft Switching Update particle xi 55/145 S T A R T Initialize all particles Evaluate objective Update pbi and gb Stop? N Optimized design(s) 𝑥𝑥𝑖𝑖 (1)𝑦𝑦𝑖𝑖 (𝑘𝑘)=𝑓𝑓𝑜𝑜𝑜𝑜𝑜𝑜 𝑥𝑥𝑖𝑖 (𝑘𝑘) 𝑝𝑝𝑝𝑝𝑖𝑖 (𝑘𝑘) 𝑔𝑔𝑝𝑝(𝑘𝑘) Calculate velocity vi 𝑣𝑣𝑖𝑖 (𝑘𝑘+1) 𝑥𝑥𝑖𝑖 (𝑘𝑘+1) Update hyperparams. Y Main workflow of PSO algorithms: Overall diagram 𝑥𝑥𝑖𝑖 (𝑘𝑘+1)=𝑥𝑥𝑖𝑖 (𝑘𝑘)+𝑣𝑣𝑖𝑖 (𝑘𝑘+1) Update current iteration k, weight inertia ω, acceleration factors c1, c2… 𝑣𝑣𝑖𝑖 (𝑘𝑘+1)=𝜔𝜔𝑣𝑣𝑖𝑖 (𝑘𝑘) +𝑐𝑐1𝑟𝑟1𝑝𝑝𝑝𝑝𝑖𝑖 (𝑘𝑘)− 𝑥𝑥𝑖𝑖 (𝑘𝑘) +𝑐𝑐2𝑟𝑟2𝑔𝑔𝑝𝑝(𝑘𝑘)− 𝑥𝑥𝑖𝑖 (𝑘𝑘) Step 1Step 2Step 3Step 4 Step 5 Step 6 ( ) 12 ,, , max(8 , 0) obj o o pp ZVS ZVS f Pv DD in λ = +− ⋅ Velocity bound applied Position bound applied ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 2.4 Global Exploration and Local Exploitation 56/145 Global exploration Local exploitation Should be better in the beginning Should be stronger in later iterations Diversity of solutions Help the algorithm avoid getting stuck in local optima Larger or random jumps in search space Refine solution quality Faster convergence Small, directed steps around high-quality designs Too much global exploration ØSlow or no convergence, waste compute Too much local exploitation ØPremature convergence to local optima 2-D Rastrigin 𝑥𝑥𝑖𝑖∈ −5, 5 Global best loc: (0, 0) vs. Iterate ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 2.5 Hardware Experimental Validation 57/145 Validate the optimized parameters in hardware prototype A hybrid NPC-DAB converter Hardware prototype Ref: [19] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.1 Challenges of Existing AI Methods in PE 64/145 Challenge 2: Explainable AI is Essential to Enhance Deployment Confidence. Input-Output Analysis: Decision Traceback Analysis: AI should be Explainable in Mathematics and Power Electronics Domains! Param. 1 Effi. ηTrends of efficiency ηw.r.t. inputs Abnormal i<0 Normal v>x Normal Fault Ref: [22] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.1 Challenges of Existing AI Methods in PE 65/145 The inductor helps to smooth out the ripple in the output current by resisting sudden changes in current flow. Lower ripple percentages typically require larger inductors. AI: For the task to design LC filter for the buck converter which has minimum current ripple, why do you recommend this inductor value? User: “Leverage PE-GPT to Provide PE Insights into a Designed Buck Converter.” PE-GPT by F. Lin et al. Challenge 2: Explainable AI is Essential to Enhance Deployment Confidence. 2. Power Electronics-Wise Explainable 1. Mathematically Explainable Training: Does it converge? Lyapunov stable? Trajectory analyzed? Infer: Are ML outputs 𝑓𝑓,∇𝑓𝑓 bounded? Controller: Is it stable? Asymptotically stable? Regret reduces to 0? Lipschitz continuity? Lipschitz ∇continuity? Β-smoothness? lim 𝑇𝑇 𝑅𝑅(𝑇𝑇) 𝑇𝑇→0? 𝑅𝑅(𝑇𝑇) 𝑇𝑇=Ο� 1𝑇𝑇 Ref: [6], [23] F. Lin et al., “PE-GPT: A New Paradigm for Power Electronics Design,” IEEE Trans. Ind. Electron., pp. 1–14, 2024. ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.1 Challenges of Existing AI Methods in PE 66/145 ØTraditional data-driven is fixed in operational settings, threatening model feasibility when settings change. ØBuilding separate surrogate models for all settings is dataintensive and computationally demanding. ØHandling multiple settings need Nt=m×n×zmodels (imaging only 3 variables), increasing complexity significantly. ØTopology, circuit parameters, control strategies, performance metrics, operation specifications, etc., could all change. #Models Required for Various Operational Settings Challenge 3: Existing AI Lacks Flexibility for Diverse Scenarios. #Modulation Strategy (MS) #Operation Specifications (OS)#Performance Metrics (PM) MS 3 MS 1 MS 2 Optimization Space Ref: [24] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.2 Next Generation of AI for Power Electronics 67/145 The Ultimate Frontier of AI for PE is: Physics-Informed AI. AI + PHYSICS Ref: [25] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.2 Physics-Informed ML in PE Design 68/145 Integrate Physics Physics Laws as Partial Differential Equations (PDEs) Machine Learning Models Guide PIML Learning to Identify Physically Consistent Solutions () () () dx t Ax t Bu t dt = + Electric Magnetic Thermal 222 222 ()Tt TTT txyz α  ∂ ∂∂∂ = ++   ∂∂∂∂  Circuit state-space equations: Heat equation in isotropic medium: E M Concept of Physics-Informed Machine Learning (PIML): ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.2 Physics-Informed ML in PE Design 69/145 Data Requirements Physics-inArchitecture Physicsin-Loss Less Data More Lightweight More Flexible More Explainable Few Data Some Data Massive Data Stronger DataDriven Capacity Physics Effectiveness Purely DataDriven ~ 100to 101~ 102 to 103≥ 104 NO Allow for Hidden Physics Well-Posed Physicsin-Init. Many Physics Some Physics NO Physics Pros and Cons of Main PIML Methods: Ref: [25] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.2 Physics-in-Loss PIML in PE Design 70/145 Solve PDEs with Hidden Dynamics Parameter Identification Identify Acquire Data Circuit Params. Embed Latent Physics-in-Loss PIML and its Applications: PDEs: Loss Terms x t PDE solution x(t)1 1 t ∂ 1 1 x∂ … Data loss Physical loss Distill into Training () 2 * 1 2 1 1() () () 1() () D PDE data data physics physics N data i i i D Ni physics i i i PDE LwL w L L xt x t N dx t L Ax t Bu t N dt = = = + = −  = −−   ∑ ∑ () () () dx t Ax t Bu t dt = + Ref: [26] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.2 PIML for Multi-Physics Modeling 71/145 Three Fin Heat Sink Heat Source (Chip) Optimize heat sink structure for minimal Thotspot min ℎ𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓,𝑙𝑙𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓,𝑡𝑡𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓 𝑇𝑇ℎ𝑜𝑜𝑡𝑡𝑜𝑜𝑜𝑜𝑜𝑜𝑡𝑡 s.t. Pressure Drop < 2.5 Pa Heat Sink Structure Variants modulus-sym by Nvidia Ref: [27] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.2 PIML for Multi-Physics Modeling 72/145 hfins lfins tfins x y z GeometryCoordinates Tsolid TFluid PSolid Optimizers Generate Structure Temperature / Pressure Fields Optimized Structure Physics-informed neural network 0 1000 2000 3000 4000 5000 OpenFOAM FEM PINN ~17 × Faster 4 V100 GPUs20 CPUs Hours Promising to Replace FEM! Thermal and Fluid Dynamics: Diffusion, Navier-Stokes Equations Ref: [27] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.2 Physics-in-Architecture NN (PANN) in PE Design 73/145 u[tk] u[tk+1]y[tk+1] x[tk+1] Physics-in-Architecture Neural Network (PANN) + ho ( u[tk+1]; θ ) h ( u[tk]; θ )Δ tk+1 States at tk+1 go ( u[tk+1]; θ ) 1+g ( u[tk]; θ )Δ tk+1 delay + States at tk ( ) ( ) () (); () (); () dxt g ut xt h ut ut dt θθ = + ( ) ( ) () (); () (); () oo yt g ut xt h ut ut θθ = + Physics-in-Architecture Neural Network (PANN), (Submitted on 21 Jun 2023) Mamba State Space Models (Submitted on 1 Dec 2023) All Roads Lead to Destination, but PANN is a Shortcut for PE https://github.com/XinzeLee/PANN Circuit Dynamics Ref: [6], [23], [24], [28]-[31] Concept of Physics-in-Architecture Neural Network (PANN): ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.2 PANN Training and Parameter Identification 80/145 PANN Training is Equivalent to Parameter Identification Neural θbefore Training: RL= 10 mΩ, Lk= 80 μH, n= 1.1 Neural θafter Training: RL= 1.8 Ω, Lk= 64 μH, n= 1.0 Manufacturing Tolerance Ambient Variations •Thermal Drift •Operation Conditions Aging and Degradation R2= 96.9% R2= 99.9% With Training Ref: [6], [23], [24], [28]-[31] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.2 PANN Training and Parameter Identification 81/145 PANN Training (“XinzeLee/PANN/DAB-inference and training.ipynb") Ref: [6], [23], [24], [28]-[31] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.2 PANN Transforms Circuit Simulators 82/145 + vo[tk+1] n ( A ) P OWER N EURAL L OOP (PNL) va[tk] vb[tk] vc[tk] vd[tk] vab[tk] vcd[tk] + - + - k oo c t CR R + ∆ + 1 () + - + - RL iL[tk] vc[tk] s1[tk] s8[tk] … k t L + ∆ 1 + - nRo × + s5-s7 nRc ×+ iL[tk+1] vc[tk+1] Ø Physics of AC-side Ø Physics of DC-side Delay c oc R RR+ Simulated waveforms over a period ΔTctrl sampled at Δtk+1 Gate driven signals over a period of ΔTctrl: [t'l+1, t'l+2] /*sampling at Δtk+1*/ s5s6 s7 s8s8 s1s2 s3 s4s4 s6s5s6 s7s8 s1s2 s3s4 ΔTctrl D1D0D2 Evaluate va(b,c,d)[tk] in (5) ( B ) C ONTROL N EURAL L OOP (CNL) Modulator vref [t'l+1] vo[t'l+1] +- kp kiΔTctrl ++ φ[t'l] φ[t'l+1] Delay Ø Proportional Ø Integral Saturation 0.5 -0.5 D0[t'l+1] D1 D2 s1~s4, s5~s8 over ΔTctrl /*sampling at ΔTctrl*/ iL[t'l+1] vo[t'l+1] t k t k+1 t k+2 Δtk+1 ΔTctrl v o [t 'l+1 ] vref [t'l+1] s1[tk],…, s8[tk], tk∈ [t' l+1 , t' l+2 ] iL[tk+1] Recurrent NeurPecs: An adaptive circuit simulator for power converters If time-series data is provided, circuit parameters will adapt to match the provided waveforms. Ref: [6], [23], [24], [28]-[31] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.2 PANN is Both Data-Light and Lightweight PANN is Lightweight (TinyML)PANN is Data-Light PANN Pioneers the Next Generation of LIGHT AI for Power Electronics in: 1. Data: Countable on ONE HAND, 2. Model Size: Deployable on EDGE DEVICE. Data Size ≥ #Params. Popular ML: Thousands, Millions, and MORE PANN: Reduce Data by 3 Orders of Magnitudes PE Physics as Data Invariant One-Layer Recurrent Net 1.72 kB of a PANN for DAB Deployable on Edge Raspberry pi TMS32 0 STM32 FPGA Arduino 1762 Strings, Approx. 160 Words 83/145 MAEs/ A (μ±std.) Data Sizes for (Train, Test, Val.) Sets (10, 90, 900) (50, 90, 860) (100, 90, 810) Piecewise 0.311±0 SVR 2.286±0 2.052±0 1.991±0 LSTM 1.560±9.5E-2 1.340±5.6E-2 1.193±9.9E-2 TCN 1.371±3.6E-1 0.726±6.1E-2 0.511±7.2E-2 PL-PINN 0.310±1.3E-2 0.259±1.1E-2 0.234±7.3E-3 LN-GRU 1.257±1.2E-1 0.991±1.2E-1 0.527±8.6E-2 PANN 0.16±4.1E-3 0.15±3.6E-3 0.13±3.3E-3 Ref: [25] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.2 Accelerate PANN’s Runtime 84/145 2400 1200 0 Average Simulation Time / ms 2000 Faster than Plecs when bs ≥ 60 Batch Size bs (L unfold = 4) 20 60 100 140 180 300 35 14 7 Memory Usage (RAM) /MB If bs ≥ 200, reduce sim. time by ≥ 73% (3.8 times faster) 0 220 260 21 28 1600 800 Memory Avg. Sim. Time Plecs: 932 ms 400 Ø Multi-core Ø Batching RAM From scalar to vector Condition set 1 Condition set 2 … Condition set bs Multi-port RAM addr1 addr2 Single Distribute Quad-core CPU 1 2 1 3 4 Data and task-level parallelism Impact of batch size on average simulation time Neuralizing circuit state-space equations facilitates parallel computing Ref: [6], [23], [24], [28]-[31] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.2 Accelerate PANN’s Runtime 85/145 200 100 0 LUnfolded length Lunfold 1 2 4 5 10 20 40 50 Average Simulation Time / ms 150 50 Optimal Lunfold: bs=100 26 ms per 10-ms sim. of DAB bs=200 bs=300 bs=500 bs=400 bs x[tk] u[tk] x[tk+1] y[tk+1] PANN ... ... ONNX Runtime Ø Unrolled length over time Lunfold Optimize Lunfold Higher Lower Large I/O latency High memory usage & more read/write TinyML methods also helps for runtime the optimization of PANN ONNX runtime optimization Impact of Lunfold on average simulation time Ref: [6], [23], [24], [28]-[31] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.2 PANN Achieves Out-of-Distribution Generalization 86/145 Total Number of Traditional ML Models: MS×PM×OS×TP #Modulation Strategy (MS) #Operation Specifications (OS)#Performance Metrics (PM) [OS 1 , PM 3 , MS 3 ,] [OS 3 , PM 3 , MS 2 ,] OS 3 OS 2 OS 1 PM 1 PM 2 PM 3 MS 3 MS 1 MS 2 Diverse Operating Conditions #Topologies (TP) TP 1 TP 2 TP 3 ONE PANN for Diverse Conditions & Topologies Training-FREE ScenarioSpecific PANN Defines the Next Level of AI Flexibility: Training-Free Generalization x[t k ] u[t k ] θ θ ∆t k+1 x[t k+1 ] PANN at t k Expert System •vi •vo •PL, etc. •PSM •PWM •Hybrid •Efficiency •ZVS •ZCS, etc. •NonResonant •Resonant •Multi-Port Operating Conditions Performance Metrics Modulation Strategy Circuit & Topology 01 02 03 04 Ref: [6], [23], [24], [28]-[31] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.2 PANN Achieves Out-of-Distribution Generalization 87/145 PANN’s Flexibility across Conditions (XinzeLee/PANN/DAB-Operational Transfer.ipynb) Source: vi= 200 V; vo∈[160 V, 240 V] Target: vi= 300 V; vo=220 V Target: vi= 300 V; vo=380 V Target: vi= 200 V; vo=140 V; D1_cycle = 50%, D2_cycle = 35% Target: vi= 200 V; vo=260 V; D1_cycle = 68%, D2_cycle = 50% Source: TPS Modulation Ref: [6], [23], [24], [28]-[31] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.2 PANN Achieves Out-of-Distribution Generalization 88/145 Piecewise Linear Sinusoidal PANN’s Flexibility across Topology (XinzeLee/PANN/DAB-Operational Transfer.ipynb) Target: Resonant Lr= 63 μH, fr= 1.1·fs (Capacitive) Target: Resonant Lr= 63 μH, fr= 1.0·fs (Resistive) Target: Resonant Lr= 63 μH, fr= 0.9·fs (Inductive) Source: Non-resonant DAB with Single L,Lk= 63 μH Ref: [6], [23], [24], [28]-[31] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.2 PANN Achieves Out-of-Distribution Generalization 89/145 7 4 1-6 -4 -2 0 2 CASE III : Boost CASE II : Unit-gain CASE I : Buck T-sne feature 1 T-sne feature 2 Val.: Robustness to operational diversity Train Test CASE II : Unit-gainCASE I : Buck CASE III : Boost NUMBER OF DATA SAMPLES Train: Test: Val: 0 20 40 60 80 100 108, 100% 108, 100% 93, 86.1% 10, 9.3% 5, 4.6% 0 0 0 0 * * Intentionally Biased Data Partition 3 2 1 0 4 LSTM CNN TST PANN Figure Legends Mean Squared Loss L(θ) Good OOD Generalization to Case I and III Train-Unit Gain: Test-Unit Gain: NO OOD Capability Test-Unit Gain: Val.-Buck: Val.-Boost: Data-light Statistics of Condition Transferability Statistics of Modulation Transferability R2=99.68% R2=99.75% Statistical Study of PANN’s Flexibility Ref: [6], [23], [24], [28]-[31] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.3 Techniques to Customize LLM Agents 96/145 LLM Customization Methods Prompt Engineering FineTuning RAG Chain-of-Thought: Easily implemented, but lack of scalability and knowledge accuracy (hallucinate) PE-specific dataset Fine-Tuning: PE-specific corpus and specialize in PE tasks, but hard to implement and no explainability RAG: Easily implemented, flexible across diverse PE tasks, and the knowledge retrieval accuracy is high Freeze most layers Vectorized knowledgebase Text-based PE knowledge base Ref: [6] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.3 Steps of Retrieval Augmented Generation (RAG) 97/145 Query Vector User Embedding Power Electronics Knowledge Base Chunking Split Text Document Vector Database Embedding Chunks Chunks Chunks Retriever Similarity Search Relevant Chunks Text Query Text Query ① ② ③ ② PE-GPT Agent i i AB AB ⋅ = 1. Chunking 2. Embedding 3. Retriever 4. Query Ref: [6] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.3 RAG Performance 98/145 Chunk size Top-k 256 512 1024 3 7 11 0 1 0.5 MRR Ø Complete HR Rel RS FF Ø Efficient Ø Precise PE-GPT's Response is: Comparisons among Human Experts, Stateof-the-Art LLMs, and PE-GPT Linguistic Performances of RAG w.r.t. Different Hyperparameters Mean Reciprocal Rank MRR Faithfulness FF Relevancy Rel Average Correctness of 77.2%, which Outperforms Experienced Human Experts by 22.2%. Against other Leading LLMs, PE-GPT Improves Correctness by 35.6% and Consistency by 15.4%. ØPrecise, Complete, and Efficient Response Hit Rate HR Response Speed RS Ref: [6] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.3 Experimental Validation 99/145 Efficiency η 100% 50 100 300 500 700 P L /W 91% 97% Current stress i pp /A 30 24 18 12 6 SPS-η TPS-i pp SPS-i pp 900 85% 79% 73% 5DoF-i pp 5DoF-η 97.53% 4.9% 1.8% TPS-η Ø Optimal i pp ZVS 5DoF TPS 46666888888 888888 88888 50 100 200 300 400 500 600 700 800 900 1000 P L /W ZCS 5DoF TPS 0 0 002220000 22000000243 Ø Extended ZVS & ZCS ranges V 2 = 160 V v s i L v p v p : [200 V / div] v s : [200 V / div] i L : [5 A / div] Time: [5 μs / div] V 2 = 160 V, P L = 300 W, (D 1 , D 2 , φ 1 , φ 2 ) = (0.71, 0.88, 57%, 56.8%) s 2 ZCS s 1 , s 2 , s 3 , s 4 ZVS q 1 , q 2 , q 3 , q 4 ZVS q 3 ZCS q 4 ZCS q 1 ZCS v s i L v p V 2 = 160 V, P L = 600 W, (D 1 , D 2 , φ 1 , φ 2 ) = (0.78, 1, 50%, 51%) s 2 , s 1 ZCS s 1 , s 2 , s 3 , s 4 ZVS q 1 , q 2 , q 3 , q 4 ZVS v s i L v p V 2 = 160 V, P L = 1 kW, (D 1 , D 2 , φ 1 , φ 2 ) = (0.88, 1, 50%, 50%) s 1 , s 2 , s 3 , s 4 ZVS q 1 , q 2 , q 3 , q 4 ZVS i L : [10 A / div] Efficiency, Current Stress (CS), and Soft Switching of CSOptimized TPS and 5-DoF Modulation “Improves Efficiency, ZVS, and ZCS under Light Load Conditions.” Waveforms under Voltage Step-Down Scenarios Ref: [6] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.3 Key Techniques in PE-GPT 100/145 PE–GPT FunctionTool Function calling and routing; Define tools to handle power electronics design tasks. RAG Integrate external knowledge bases to augment power electronics domain expertise. by F. Lin, X. Li et al. Ref: [6] Agentic RAG XinzeLee/PE-GPT ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.3 LLM Orchestration to Build PE-GPT 101/145 External Knowledgebase Orchestration Framework (LlamaIndex, LangChain) Available LLMs RAG User Query Synthesized Response Pipeline Manager Load & Chunk Retrieved Documents/ Invoked Functions FunctionTool Vectorize Define Parse & Route & Invoke Simulate Validate the optimized D1and D2 in simulation Optimize Optimize D1and D2 for EPS of DAB with specified Vin, Vout, PL Embed Modulation of DAB q4q3 q1q2 s4s3 s1s2 DAB Modulation Knowledge Multi-Modal: Process PE Data Types (Waveforms) Ref: [6] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.4 Reinforcement Learning for Control 102/145 Reinforcement Learning for Control Action (Control Dynamics) Memory Replay Environment (Power Converters) Reward Function Reinforcement Learning (RL) Optimizes Control Trajectories or Solves Optimization Problems by Reinforcing Positive Actions: Trial-Reward-Learn. Critic (Twin of Environment) Controller ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.4 Reinforcement Learning for Control 103/145 RL-Based MPPT Control Ref: [32] Online Learning & Deploying A Reinforcement-Learning (RL)-Based Controller for MPPT in Wind Energy Systems ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.4 Reinforcement Learning for Control 104/145 The RL-Based MPPT Controller Tracked Optimal Power More Closely, Indicating Faster Transient Response and Reduced Power Dip. Conventional P&O MPPT RL-Based MPPT Closer to Optimal Power vs Ref: [32] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 3.4 Reinforcement Learning for Optimization 105/145 Reinforcement Learning for Topology Optimization Action (Add/delete node/edge) Memory Replay Environment (Circuit Simulator) Reward Function Critic (Twin of Environment) Topology Generator Reinforcement Learning (RL) to Optimize the Topology of Multiport DC-DC Converters (Optimization Tasks) ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 4. AI Ethics in Energy Conversion Engineering Practice 112/145 WHY Does It Matter? WHAT Are the Key Ethical Challenges in Practice? HOW Can We Responsibly Engage? ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 4.1 Ethics Is Not Optional 113/145 AI Usage is Growing. Ref: [35] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 4.1 Ethics Is Not Optional 114/145 Utilities using AI applications by category, 2024 AI Usage is Growing. Ref: [36] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 4.1 Ethics Is Not Optional 115/145 AI Incidents: Why Ethics Can’t Be an Afterthought Ref: [35] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 4.1 Ethics Is Not Optional 116/145 AI Incidents: Why Ethics Can’t Be an Afterthought Ref: [35] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 4.1 Ethics Is Not Optional 117/145 Moreover, The Rise of AI Regulations Is Driving Responsible AI Development The AI Act is a European regulation on artificial intelligence (AI) –the first comprehensive regulation on AI by a major regulator anywhere. Rules for general-purpose AI models are set to take effect in August 2025, while most other rules of the AI Act are scheduled to begin to apply in August 2026. The OECD AI Principles are a set of guidelines adopted by the Organisation for Economic Co-operation and Development (OECD). First adopted in 2019 and updated in 2024, not legally binding regulations but rather a framework for governments and AI actors to follow ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 4.2 Four Pillars of AI Ethics in Energy Conversion 118/145 •Energy System Stability •Energy System Security •Cybersecurity Security & Safety 01 02 03 04 Interpretability & Transparency Energy Sustainability Evolving Roles of Engineers •Intrinsic Interpretability •Post-Hoc Interpretability •Growing Carbon Emissions •Be Energy Aware •New Roles and Opportunities Ahead •Upskilling and Reskilling ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 4.2.1 Security & Safety of Energy System 119/145 AI Risk in Energy System Stability & Energy System Security •The ethical responsibility to ensure that AI technologies deployed in energy systems do not compromise the energy system’s stability and security. Use Case Vulnerability Consequence Sudden irradiance drops (e.g., cloud shading) cause erratic behaviour that the AI model wasn’t trained for. AI-based MPPT in PV inverter DC bus fluctuation and grid disturbance. AI-Driven Control of Energy Storage Systems Fails to enforce thermal or SoC constraints under edge conditions (e.g., summer peaks). Overheating or overcharging could trigger thermal runaway or equipment fire. Sudden rise in wind power occurs outside the training distribution (e.g., strong gusts ). Frequency rise, Inverter tripping due to power quality violations. AI in Wind Power Forecasting ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 4.2.1 Security & Safety of Energy System 120/145 Metrics Note Accuracy How closely AI predictions match actual values. Robustness The system’s ability to maintain performance under disturbances, noise, and fault conditions (e.g., thermal spikes, grid events). Generalization AI must perform reliably in unseen or rare edge cases (e.g., extreme weather, load surge). Safety Constraints Predefined operational limits within which the AI system must operate to avoid safety risks or violations of grid codes, such as frequency / current / thermal limits, etc. Latency AI decisions must be made within strict timing requirements, especially for real-time control. Safety Risks Translated to Engineering Metrics ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 4.2.1 Security & Safety of Energy System 121/145 Safety Risks Translated to Engineering Metrics:An Example IEC 61850-5: or Class P1 (protection), the end-to-end latency requirement is ≤ 10 milliseconds. When AI is used for a circuit breaker command, its worst-case inference time must comply the ≤ 10 ms latency requirement for P1 and 3 ms for P2/P3. •Use Lightweight AI Models •Optimize Deployment Environment •Integrate Fallback Logic Engineering Insights Ref: [37] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 4.2.2 Interpretability & Transparency 128/145 Trade-off Between Interpretability and Model Complexity Post-Hoc Tools Intrinsic Interpretability Ref: [39] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 4.2.2 Interpretability & Transparency 129/145 Inference Stability Analysis X. Li, F. Lin, H. A. Mantooth, and J. J. Rodríguez-Andina, “Explainable Physics-in-Architecture Neural Networks for Power Electronics: From a Lipschitz Continuity Perspective,” 2025, arXiv. 𝒙𝒙 𝑡𝑡𝑘𝑘+1 =1− 𝐴𝐴Δ𝑡𝑡 −1 1− 𝐴𝐴Δ𝑡𝑡 −1𝐵𝐵Δ𝑡𝑡 𝒙𝒙 𝑡𝑡𝑘𝑘 𝒖𝒖 𝑡𝑡𝑘𝑘+1 Params. 𝜽𝜽are Defined in Matrices A and B PANN: Formulations of PANN Definitions of Lipschitz Continuity To Mathematically Explain AI: From a Lipschitz Continuity Perspective 𝑓𝑓 𝒙𝒙1− 𝑓𝑓 𝒙𝒙2≤ 𝐿𝐿1𝒙𝒙1− 𝒙𝒙2,∀𝒙𝒙1,𝒙𝒙2∈ 𝑅𝑅𝑛𝑛 Validation of Inference Stability Validation of Training Convergence ‖∇𝜽𝜽𝑓𝑓(𝜽𝜽)‖≤ 𝐿𝐿1𝜽𝜽 𝑦𝑦𝑖𝑖𝑦𝑦𝑦𝑦𝑦𝑦 � ⎯ ⎯ � ‖𝑓𝑓(𝜽𝜽𝟏𝟏)− 𝑓𝑓(𝜽𝜽𝟐𝟐)‖≤ 𝐿𝐿1𝜽𝜽‖𝜽𝜽𝟏𝟏− 𝜽𝜽𝟐𝟐‖ ‖∇ 𝑧𝑧 𝑓𝑓(𝒛𝒛)‖≤ 𝐿𝐿 1𝑧𝑧 𝑦𝑦𝑖𝑖𝑦𝑦𝑦𝑦𝑦𝑦 � ⎯ ⎯ � ‖𝑓𝑓(𝒛𝒛 𝟏𝟏 )− 𝑓𝑓(𝒛𝒛 𝟐𝟐 )‖≤ 𝐿𝐿 1𝑧𝑧 ‖𝒛𝒛 𝟏𝟏 − 𝒛𝒛 𝟐𝟐 ‖ 0 Bounded 20 40 60 80 100 1 0.9995 0.999 Lipschitz Const. of Jacobian L 1z Theoretical L1z = 0.999996 MC Evaluation �∇ 𝒛𝒛𝑦𝑦𝑑𝑑𝑝𝑝 𝒙𝒙 � 𝑦𝑦𝑑𝑑𝑝𝑝 (𝒛𝒛 𝑦𝑦𝑑𝑑𝑝𝑝 )� Proof: Jacobian Matrix Norm is Bounded. Proof: Gradient Matrix Norm is Bounded. PANN has Bounded Jacobian Matrix Norm 𝜵𝜵𝒛𝒛𝒇𝒇(𝒛𝒛) Ref: [23] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 4.2.3 Energy Sustainability 130/145 AI Advancements Come with Growing Carbon Emissions Ref: [35] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 4.2.3 Energy Sustainability 131/145 Energy Usage Breakdown in AI Model Development Ref: [38] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 4.2.3 Energy Sustainability 132/145 Primary Energy Usage: 1. LLM customization; 2. LLM inference; 3. Design workflow Case Study: Energy Usage of PE-GPT Ref: [6] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 4.2.3 Energy Sustainability 133/145 Case Study: Energy Usage of PE-GPT Energy Consumption Breakdown For a DAB Converter Modulation Design Case. Assumptions: LLM (LLaMA 65B) customization: once LLM inference: 8-round interactively Note: A token is defined as the smallest data unit processed by LLMs (e.g., words or characters). Comparison of three LLM customization techniques Ref: [6] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 4.2.3 Energy Sustainability 134/145 Only Use AI when We Truly Need it Optimize Energy Efficiency with PE Minimize the Computation of AI Establish Regulations and Foster Collaboration Reconsider whether AI is the best tool to address specific PE tasks. Energy consumption can be reduced through algorithms, architectures, and hardware. Advanced power electronics can help reduce energy consumption in return. Establish regulations to promote awareness and encourage opensource collaboration. Image Credit to Open AI GPT-4o: Trade Off Between AI Gains and Energy Consumption. Ref: [35] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 4.2.4 Evolving Roles of Engineers 135/145 With that said, AI may reduce electrical engineers’ work in theoretical analysis, calculation, documentation, etc. … ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 4.2.4 Evolving Roles of Engineers 136/145 Or maybe even hardware prototyping? Image Credit to GPT 4o ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 4.2.4 Evolving Roles of Engineers 137/145 AI opens doors to new roles and opportunities Ref: [40] ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li 6. Conclusions 144/145 AI in Power Electronics Design: What’s the Present AI is omnipresent in PE design: including system-level, converter-level, and component-level designs, where ML algorithms are mainly used for modeling, and MHAs are for optimization. Hands-on practice of AI-augmented modulation design for DAB converters, including current stress modeling (regression), zero voltage switching modeling (classification), MHA for modulation parameter optimization. Emerging Trends of AI in PE Design PIML, which supports the data-driven capacity of AI by integrating PE physics, is leveraged to solve PDEs and identify parameters, paving the ways for the next-generation PE-level interpretability and light and flexible AI. PE-GPT marks new PE design paradigm with multi-modal generative AI directing interactivity at a high level. Reinforcement learning and graph neural network provide substantial benefits. Ethics of AI Applied in Energy Systems Four main pillars of AI ethics ØSecurity and safety ØInterpretability and transparency ØEnergy sustainability ØEvolving roles of engineers ECCE2025 Tutorial · Reimagine Power Electronics Design with AI ·Peter Wilson, Fanfan Lin, Xinze Li Thank You! Q&A 1University of Bath 2University of Arkansas 3Zhejiang University-University of Illinois Urbana-Champaign Institute (ZJUI) Fanfan [email protected] Peter [email protected] Xinze [email protected]