Framework with Digital Twins and Federated Learning for Decision Support in Multi-plant Schemes for Electric Steelmaking and Beyond

Framework with Digital Twins and Federated Learning for Decision Support in Multi-plant Schemes for Electric Steelmaking and Beyond Petri Kannisto*, Bernd Kleimt, Akhilesh Chandgude (1) Nicolás García Crespo, Irene Guardiola Luna (2) Stefano Dettori, Valentina Colla (3) Bárbara Fernandez (4) (1) BFI (VDEh-Betriebsforschungsinstitut), Düsseldorf, Germany (2) EVIDEN, Madrid, Spain (3) Scuola Superiore Sant’Anna, Pisa, Italy (4) Celsa Group, Barcelona, Spain * Corresponding author & presenter; [email protected] DTMI_019 Contents •ALCHIMIA project introduction •ALCHIMIA framework •Digital Twins •Electric Arc Furnace •Ladle Furnace •Kubernetes •Message-oriented data integration •Results from Federated Learning •Lessons learned •Conclusions ALCHIMIA project “Data and decentralized Artificial intelligence for a competitive and green European metallurgy industry” Duration: 9/2022-11/2025 (39M) Horizon Europe grant no. 101070046 https://alchimia-project.eu/ This work has been supported by the project "Data and decentralized Artificial intelligence for a competitive and green European metallurgy industry" (ALCHIMIA), which has received funding from the European Union's Horizon Europe research and innovation programme under grant agreement No. 101070046. Views and opinions expressed are however those of the authors only and do not necessarily reflect those of the European Union. Neither the European Union nor the granting authority can be held responsible for them. ALCHIMIA project themes •(Steel production) process optimization with decision support •Tools: •AI models •»Non-AI» models •Federated Learning, Continual Learning •Human factors •Life Cycle Assessment (LCA) Use cases Celsa •Electric steelmaking route •Case plants: France, Spain, Poland Fonderia di Torbole (FdT) •Foundry; automotive components •Case plant: Italy Celsa: electric steelmaking route Figure copyright © Celsa Group This presentation ALCHIMIA‘s Celsa use case Federated Learning and Continual Learning •Federated Learning •Central server generates global model weights from local models •Local models not available to other local clients -> preserve privacy •Continual Learning •Monitor model performance •Re-train model when appropriate Picture from doi:10.5281/zenodo.14741719 ALCHIMIA framework •Aiming at Federated Learning (FL) between production plants •Improve performance esp. if limited data coverage in some plant •Each plant operates: •Its own optimization models •Its own data infrastructure •Connectivity between plants and FL through the local client •The models provide decision support to the operators/engineers •Human in the Loop Kubernetes infra Developer computer Kubernetes •“Infrastructure as Code” (IaC) •Configure your system with manifest files; the infra installs components for you Image registry Local Docker kind: Deployment volumes: ... containers: ... image: ... --- kind: Service ... --- kind: PersistentVolume ... Software Manifest (YAML) Software as „pod“ Manifest file Message-oriented communication •Excels at event-based schemes •We created “Smart Data Model” for common JSON structures •Open standard; https://smartdatamodels.org •Subject “IndustrialProcess” under domain “Smart Manufacturing” Message broker (Apache Kafka) ExtractTransformLoad (ETL) pipeline EAF model LF model { "id": "urn:ngsi-ld:xxx", "type": "ProcessChemicalAnalysis", "dateObserved": "2025-07-17T14:07:00Z", "processName": "eaf_1", "heatNumber": 10021, "chemicalAnalysis": { "sampleNumber": 2, "chemicalConcentration" : [ { "substance": "Fe", "percentage": 0.9873 }, { "substance": "Cu", "percentage": 0.0052 }, { "substance": "Mn", "percentage": 0.0023 } ] }} Raw data Example message Federated Learning results: EAF model (temperature; offline) •How is Federated Learning possible here? •EAF model is physical/analytical, not AI! •Solution: apply FL to modelparametrization-related knowledge •Seems to work, but the data must be OK •Future work: fix the problematic dataset One dataset has problems Problematic dataset removed MAE: Mean Absolute Error Federated Learning results: LF model (temperature; offline) Model Features Plant 1 Plant 2 Plant 3 R2 Score MSE R2 Score MSE R2 Score MSE Local FinalTemperature 0,278 137,33 0,580 36,65 0,158 59,10 Federated FinalTemperature 0,446 98,88 0,579 45,56 0,299 32,77 •Plants 1 and 3: improvement in both R2 and MSE •Plant 1: smallest dataset •Plant 3: worst data quality; largest dataset •Plant 2 results are maintained, especially R2 •Possible reason: imbalanced datasets R2 score: regression performance MSE: Mean Squared Error •Conclusion: FL likely improves results •Future experiments •Compensate imbalance in datasets •Discard bad data points Lessons learned •Kubernetes •Infrastructure as Code -> reduce manual pain (however, someone must install + maintain it) •Requires some learning •Sometimes surprisingly primitive wrt. Docker •Federated Learning •Lots of potential •More plants -> more challenges •Data quality, availability and symmetry? •Last step in chain -> will suffer if problems elsewhere •Event-based communication •Important: functional interoperability: *what* happens *when* •Any message missing -> cannot calculate Conclusions •Digital Twin framework reached its goals •Data integration •Decision support online •Federated Learning •Federated Learning has potential in process industry •Promising results reached (offline in this presentation) •Data must be first-class citizen •Develop concepts further in future projects Thank you for your attention! Petri Kannisto BFI (VDEh-Betriebsforschungsinstitut) Düsseldorf, Germany [email protected] linkedin.com/in/petri-kannisto alchimia-project.eu All authors: Petri Kannisto, Bernd Kleimt, Akhilesh Chandgude BFI (VDEh-Betriebsforschungsinstitut), Düsseldorf, Germany Nicolás García Crespo, Irene Guardiola Luna EVIDEN, Madrid, Spain Stefano Dettori, Valentina Colla Scuola Superiore Sant’Anna, Pisa, Italy Bárbara Fernandez Celsa Group, Barcelona, Spain This work has been supported by the project "Data and decentralized Artificial intelligence for a competitive and green European metallurgy industry" (ALCHIMIA), which has received funding from the European Union's Horizon Europe research and innovation programme under grant agreement No. 101070046. Views and opinions expressed are however those of the authors only and do not necessarily reflect those of the European Union. Neither the European Union nor the granting authority can be held responsible for them.