Machine Learning Chemical Compound Space from first Principles - Towards Self-Driving Laboratories

by Anatole von Lilienfeld (University of Toronto & Vector Institute)

Monday Jun 19, 2023, 9:30 AM → Monday Jun 26, 2023, 10:30 AM America/Chicago

Many of the most relevant observables of matter depend explicitly on atomistic and electronic structure, rendering a first principles approach to chemistry and materials mandatory. Unfortunately, due to the combinatorial nature of chemical compound space, the set of all conceivably possible materials and molecules, gaining a holistic and rigorous understanding through exhaustive quantum and statistical mechanics based sampling is prohibitive --- even when using high-performance computers. Accounting for explicit and implicit dependencies and correlations, however, will not only deepen our fundamental understanding of chemical compound space but also benefit the efficiency of computational as well as experimental exploration campaigns within self-driving lab settings. I will discuss insights into such relationships gained thanks to supervised machine learning on quantum data obtained from supercomputing. Numerical results indicate promising performance in terms of efficiency, accuracy, scalability, and transferability (EAST).