Data driven discovery and optimization of high entropy alloys and their hydrides

Data-driven modeling in recent years has ushered in a new paradigm for rapid discovery of useful materials across a plethora of domains in the physical and materials sciences. These methods become particularly invaluable when investigating applications of high-entropy materials, where the combinatorial growth of explorable chemical space makes brute-force experimentation or first-principles simulation intractable. To assess the primary figure(s) of merit for 10,000s of possible materials for a given application and down-select only top candidates for more rigorous examination, accurate and efficient machine learning and data-driven techniques are required. This talk will survey a variety of data-driven (high entropy) hydride discovery exemplars representing efforts between Sandia and a group of international collaborators. These range from traditional machine learning approaches for direct hydride thermodynamic property prediction to modern graph neural networks, trained as DFT surrogate models, that feed sampling-intensive first principles simulations. These modeling strategies can rapidly screen hydride thermodynamic properties and therefore experimentally target new materials or rationally tune existing ones for various hydrogen storage or compression use cases. Coupled with additional simple objectives (i.e., raw material costs), multi-dimensional Pareto optimal materials can therefore be identified, targeted, and synthesized.