The project ALCHIMIA, where R&D Spain has acted as the coordinator and leader of WP1 (Project management) and WP3 (Federated Learning and Continual Learning), ended successfully on November after 39 months of work involving colleagues from the Trustworthy AI and HPC software products teams.

To close it, a final review was organized by the EC and the consortium, held online on 29 January, in which our team presented the WP3 main outcomes, including a demo on federated learning, continual learning, and transfer learning.

The mission of the ALCHIMIA project was to provide sustainable and competitive EU metalworking industries with a platform to support the transition to high-quality, competitive, efficient, and green production processes with the guarantee of high-quality products in the steel-making industry.  The project has achieved most of its expected results. However, some challenges have been addressed. In particular, data availability and harmonization across the different pilots required several data-collection campaigns to obtain higher-quality data.

Our contribution has been identified as one of the main key outcomes, because the Federated Learning platform was the cornerstone of the project. 

ALCHIMIA has served as the basis to the conceptualization, design and first implementation of the Federated Learning framework developed by R&D Spain (and now renamed “FeLiX”). It has been also vehicular to increasing its maturity, incorporating key features towards industrialization, such as continual learning and transfer learning.

By coupling FeLiX, with continual learning capabilities, we are now able to deploy automated, distributed systems for performance monitoring and automatic retraining when statistical deviations in the input data are detected. Additionally, by coupling FeLiX with transfer learning features, its usability is extended to cross-domain applications, enhancing its applicability.

Furthermore, the prototype has been successfully validated through ALCHIMIA’s use-cases, where the framework has been used to federate a wide range of models, and the continual learning system has been successfully tested in a real industry scenario, together with CELSA.

Some of the main results of this project have been the following:

• Platform for Federated Learning, Continual Learning and Transfer Learning     
• Dynamic process model for Monitoring and Control of the Electric Arc Furnace process    
• Scrap mix characterisation and optimisation for the Electric Arc Furnace
• Industry 5.0 Tool Kit for human-centric technology development and insertion
• Adaptable and scalable data transformation process for large scale use     
• Ladle furnace model and optimisation for product quality improvement and environmental impact minimization based on LCA
• Prediction and classification of defects in foundry process
• IAM4SDG methodology

With federated learning being the cornerstone of this project, using it to optimize manufacturing processes and help large metalworking industries become more environmentally friendly and efficient, ALCHIMIA has been vehicular for the R&D team in Spain to conceptualize, design and improve our patented Federated Learning Framework (newly named “FeLix”), including advanced capabilities such as continual learning (CL) and transfer learning (TL), leveraging its modularity and scalability.