Solving industrial materials problems by using machine learning across diverse computational and experimental data

Many materials characteristics of broad industrial relevance, such as in-service degradation, lack satisfactory physics-based models. Nonetheless, we often have access to more fundamental simulations, mechanistic theories, and experiments that we might expect to correlate in some way with a highly applied property of interest. In these situations, we suggest that machine learning can serve as an effective integrator of physical signal from many different sources, ranging from first-principles calculations to analytical models to experimental observations, in service of predicting a complex property that has traditionally resisted accurate modeling. This data-driven approach is advantageous because it utilizes our existing understanding of materials physics to the greatest extent possible, while still enabling us to extend beyond well-understood regimes. Further, it allows us to partially offset the need for very expensive real-world tests with databases of simpler experiments, or plentiful computational data.