LCO-sensitive graph neural network framework for high-accuracy energy mapping in NiCoCr medium-entropy alloys
Machine learning (ML) techniques have become pivotal in material design, yet accurately capturing the subtle energy shifts associated with local chemical order (LCO) in multi-principal element alloys remains a significant challenge. This study establishes a graph convolutional neural network (GCNN) framework specifically engineered to map the potential energy landscape of NiCoCr medium-entropy alloys with high sensitivity to LCO transitions. Utilizing hybrid Monte-Carlo molecular dynamics simulations as a systematic evaluation benchmark, we evaluate the GCNN’s capacity to learn the non-linear relationship between atomic arrangements and structural stability across varying thermal regimes. The atomic configurations are transformed into graph representations, where nodes incorporate atom types and absolute velocities, and edges are established with the 12 nearest neighbours to encapsulate the local chemical environment. The model’s fitting and predictive fidelity was assessed through three distinct case studies: individual thermal datasets, combined temperature ranges, and unseen configurations. The GCNN demonstrates exceptional performance, achieving coefficient of determination R2 values up to 0.98 and a mean absolute error as low as 1.04 meV atom−1 on entirely withheld thermal trajectories (550 K), effectively tracking the energy variations correlated with the system’s LCO evolution. This research offers a comprehensive graph-based modelling approach for understanding the configuration-to-energy mapping relationships in complex multi-principal element alloys, providing a baseline structural architecture that can be extended toward ML interatomic potential workflows.