Matthew Carbone, Columbia University

Photo of Matthew Carbone

The connection between the atomistic and electronic structure of atoms, molecules and materials is of paramount importance in many fields of physical science. Unfortunately, the difficulty of solving the Schrödinger equation and predicting quantum observables scales exponentially with particle number due to the nature of the many-body problem, making analytic solutions impossible in all but the simplest cases, and numerical solutions challenging - and costly - at best. Even some of the fastest approaches, such as Density Functional Theory, scale no better than cubically with particle number, making calculations on large systems prohibitively expensive. That said, we are now in the age of big data, and as such machine learning offers a cheap and highly scalable alternative to conventional quantum chemistry methods. It may even be preferable in some circumstances. Critically, it has two advantages over the old guard: first, it eliminates the need for human feature engineering; second, once a model has been trained, new results may be predicted in what is effectively constant time, allowing for high throughput. Here we present results demonstrating the efficacy and predictive power of machine-learning algorithms on both forward (predicting spectra from structure) and inverse (predicting structure from spectra) problems, and motivate why machine learning should be a part of the quantum chemist's toolbox.

Abstract Author(s): Matthew R. Carbone, Mehmet Topsakal, Deyu Lu, Shinjae Yoo