Unraveling the Mechanisms of Stability in Co_xMo_70-xFe₁₀Ni₁₀Cu₁₀ High Entropy Alloys via Physically Interpretable Graph Neural Networks

In recent years high entropy alloys (HEA) have become a topic of significant interest due to their combinatorial nature, showing promise for hypersonics and catalysts. In particular, the HEA system Co_xMo_70-xFe₁₀Ni₁₀Cu₁₀ has been studied experimentally and computationally due to its reported superiority as a catalyst for ammonia decomposition. However, such catalytic reactions take place at elevated temperatures, leading to potential HEA instability and eventual phase separation at catalytically active temperatures. To this end, we combine density functional theory (DFT) calculations of mixing free energies, that include mixing and vibrational entropy terms, with physics-inspired graph neural networks (GNN) and consider binary (A ↔ B + C), ternary and quaternary decomposition routes. We show that by learning the mixing free energy with our GNN framework we can rank geometric and chemical HEA features to better understand which features are more important than others at stabilizing HEA stability at catalytically active temperatures.