Computational approaches play an increasingly important role in materials science, but improvements are much needed in their accuracy, speed, automation, and scaling to realistic systems and large computer resources. This talk will highlight challenges in atomistic modeling and illustrate how fundamental physics, chemistry, mathematics, and computer science can be combined to develop modeling methods for understanding and designing materials for energy technologies. In the domain of semiconductors, first-principles quantum calculations can extract the full details of interactions between electrons and phonons and quantitatively predict thermal and electrical transport properties of complex crystals. Once validated against experiment and automated, these computational methods have enabled discovery several previously nonexistent alloy compositions, that achieved record thermoelectric efficiency in devices. In the domain of polymer and crystalline ion conductors, molecular dynamics simulations enabled discovery of new classes of solid electrolytes and surprising anomalous phenomena in correlated liquids. Modeling of heterogeneous and reactive dynamics has long remained beyond reach due to the combined requirements of high accuracy and large scale of simulations. New classes of methods, such as Bayesian active learning and deep equivariant neural networks, when trained on accurate quantum computations, have opened promising avenues for bringing near-quantum accuracy to systems of billions of atoms, reaching the scales needed for describing realistic materials. In the domain of heterogeneous catalysis and surface science, simulations of surface reconstruction and reaction phenomena approach the “digital twin” capability of correctly capturing what expensive experiments measure. A major remaining bottleneck is the fidelity of the underlying quantum calculations, where semilocal density functional approximations qualitatively miss key aspects of many-body electron exchange and correlation. Development of new types of orbital-dependent nonlocal functional approximations holds promise in breaking through the accuracy barrier, especially important for capturing charge transfer in defects, surfaces, and battery materials.
Boris Kozinsky is the Thomas D. Cabot Associate Professor of Computational Materials Science at the Harvard School of Engineering and Applied Sciences and Principal Scientist at Bosch Research. He studied at MIT for his B.S. degrees in Physics, Mathematics, and Electrical Engineering and Computer Science, and received his PhD degree in Physics also from MIT. He then established and led the atomistic computational materials design team at Bosch Research in Cambridge MA. In 2018 he started the Materials Intelligence Research group at Harvard that works at the intersection of fundamental materials physics, computational chemistry, and data science. His group develops and combines atomistic and electronic structure computations with machine learning for understanding and predicting quantum-level microscopic effects, particularly ionic, electronic and thermal transport and transformations in materials for energy storage and conversion. His work on the development and application of computational methods led to computation-driven design of materials for thermoelectrics, batteries, catalysts, and functional polymers.
AP/MSME Colloquium Series