Adrian Roitberg
Affiliation
University of Florida, USA - Roitberg Group
10+ years of General Machine Learning Interaction Potentials. Ubi eramus? Ubi sumus? Quo imus?
Abstract
When one wants to work in a field where it is important to sample molecular conformations according to a distribution (e.g. Boltzmann), the standard view has been that one can either sample properly (i.e keep many structures, for which you need cheap computational methods) OR have accurate weights (i.e. Good energy calculations, which are expensive), but never both.
This general idea goes back to Dirac, who in 1929 wrote that we (theoretical physicists) need to design methods that are computationally tractable, but as accurate as possible.
In the early 2000’s, a idea came to life asking: “Can a machine learning method learn quantum chemistry?” More precisely, can it learn and predict energies and forces for molecular systems with the same accuracy as actual QM calculations ? Surprisingly, the answer was YES, for a given molecular system. This enables the breaking of the problem described above. One could calculate energies very accurately, but at a very low cost, which enabled sampling !
In 2016 my group asked the follow up question: Can an ML method learn to predict energies and forces for ANY substance, given enough data ? The answer was again, yes! We showed that one can create a Neural Network that can learn from a large dataset of diverse structures, and predict E and F with high accuracy for compounds outside the training dataset.
I will present some of the history of the field (Ubi eramus? ), show some our more recent results (Ubi sumus? ), and discuss some ideas as to where we might be going (Quo imus?”).