Reproduce “Machine learning for structure-guided materials and process design” on OpenResearch

About this paper

Lukas Morand, Tarek Iraki, Johannes Dornheim, Stefan Sandfeld, Norbert Link, Dirk Helm

This research presents an integrated machine learning framework for materials and process design that addresses the entire process-structure-property chain. It uses a two-stage approach to first identify a diverse set of near-optimal microstructures for desired properties and then find an optimal manufacturing path for the best reachable microstructure, demonstrating its capabilities in crystallographic texture optimization.

Key insights

• Leveraging the inherent non-uniqueness of inverse design problems, the framework identifies a diverse set of near-optimal microstructures, providing flexibility for subsequent process design.
• The two-stage machine learning approach allows for a holistic optimization across the entire process-structure-property chain, linking materials design with manufacturability.
• The process design step (MEG-SGGPO) demonstrates high data efficiency, achieving successful texture reproduction with significantly fewer samples than the initial materials design training.

Summary by alphaXiv.