Machine Learning-Enhanced Multiscale Material Mechanics Modeling for Complex Alloys

Conspectus Structural materials at the frontier of performance, such as high-entropy alloys (HEAs) and metallic glasses (MGs), derive their exceptional properties from disorder. Chemical disorder in HEAs and structural disorder in MGs both generate rugged energy landscapes that confound conventional multiscale modeling, where the transfer of parameters across scales often erases the essential local fluctuations that govern strength, ductility, and failure. Over the past decade, our group has sought to tame disorder through the integration of machine learning (ML) and physics-based mechanics, building data-driven frameworks that remain anchored in mechanism and interpretability. We addressed this challenge in chemically complex refractory HEAs by developing machine learning interatomic potentials (ML-IAPs) that reproduce quantum-mechanical accuracy for atomistic simulations of dislocation processes. These potentials revealed how local chemical ordering, lattice distortion, and diffuse antiphase boundaries jointly control the screw–edge slip discrepancy and the strength–ductility trade-off. To extend these insights to mesoscale thermodynamics, we designed a physics-informed graph neural network (PIGNN) that embeds lattice symmetry into Monte Carlo sampling, accelerating the exploration of chemical ordering phenomena and further influencing their mechanical behaviors. In parallel, we examined structurally disordered MGs, where a mechanical response arises from the activation of nanoscale shear transformation zones (STZs). Through ML-based outlier detection, we identified flow defects directly from thousands of thermally activated events, providing an unbiased view of atomic-scale plasticity. Coupling these insights with mesoscale STZ dynamics modeling and surrogate neural networks, we uncovered how elastic heterogeneity governs the organization of STZs into shear bands and how microstructural design, such as a dual-phase amorphous architecture, can suppress strain localization and enhance ductility. Across these two classes of disordered alloys, a unifying philosophy emerged: ML is most powerful not when replacing physics but when revealing the hidden structure of disorder that physics alone cannot resolve. Integrating data-driven models with mechanistic frameworks enables a predictive, interpretable, and transferable approach to materials design. This Account reflects our research journey, its motivations, conceptual lessons, and evolving perspective on how data-driven and physics-based approaches can operate synergistically to accelerate the discovery of complex alloys in extreme environments.