A Physics-Aware Materials Discovery Platform for Stable High Entropy Alloys

High entropy alloys (HEA) have become a topic of significant interest due to their combinatorial nature and their potentially superior properties in application spaces such as catalysts and high temperature aerospace and manufacturing. However, materials discovery pipelines for HEA still lag behind other materials classes such as polymers and molecules. To address this, we take the approach of combining both discovery and design into a single process, where our discovery pipeline also learns the underlying physics behind structure-property mappings. This combination creates an informative feedback loop where new materials can be down-selected from the search space based on physis-awareness and not purely statistical similarity. We showcase this pipeline for the HEA system Co_xMo_70-xFe₁₀Ni₁₀Cu₁₀, which has been studied extensively due to its reported superiority as a catalyst for ammonia decomposition. We show that by combining density functional theory (DFT) with physics-aware graph neural networks (GNN), we can learn properties such as mixing free energy, providing a basis for solid solution stability of the HEA. We also show that by learning the mixing free energy with our GNN framework we can autonomously rank geometric and chemical HEA descriptors based on their importance towards the mixing free energy, a natural output of our GNN architecture. This framework provides a set of design rules and property predictions, creating an informative search criterion for new HEA systems.