Title : The Role of First Principles Methods in a Data-Driven World

Abstract: Two Nobel prizes were just awarded on machine learning topics, reflecting the broad enthusiasm for data-driven methodologies in the physical sciences. The public facing view on machine learning—and also what is taught in the classroom—emphasizes the powerful algorithms that enable learning through deep neural networks and related models. In contrast, I will present my view on the less visible counterpart to the algorithm: the data, upon which all machine learning models stand or fall. Two examples in this talk will show how first principles techniques can provide insight into data-starved areas where machine learning cannot reach. The first case will involve computational techniques for predicting stereoselective reactions, and the second will build bridges from wavefunctions to electron densities. In all, I will highlight the importance of physical interpretability of these models as well as their potential to be combined with machine learning techniques to expand their overall impact.