"Chemistry by design with physics-informed machine learning"

Our ability to discover and design new materials and molecules is currently being revolutionized by machine-learning (ML) methods. Many existing approaches use ML to predict a variety of molecular properties. This has enabled computational molecular property prediction within vast chemical compound spaces and high-dimensional parametrization of energy landscapes for the efficient molecular dynamics simulation of measurable dynamic observables. However, as all properties fundamentally derive from the quantum mechanical (QM) wave function, an ML model that can predict said wave function also has the potential to predict any derived properties. ML models of the electronic wave function also provide new opportunities to consider deeper integration of data-driven and physics-led approaches. In this talk, I will describe ML approaches that directly represent wave functions and Hamiltonians and their role in developing new computational materials design approaches.

Using example systems from heterogeneous catalysis, organic electronics, and small molecule drug design, I will discuss the challenges associated with encoding physical symmetries and invariance properties into ML models of electronic structure. Once these challenges are overcome, integrated ML-QM methods offer the combined benefits of data-driven and quantum chemical methods. I will discuss several opportunities associated with building ML-augmented quantum chemical methods, including inverse chemical design approaches and the development of efficient and accurate surrogate models that can benefit scientific discovery across the chemical sciences.