ML for many-particle physics

In principle, the equations governing the behavior of atoms in solids (such as metals) or molecules are perfectly known -- they just fall out of quantum mechanics. However, when the size of the system grows beyond a handful of atoms, even on today's hardware, solving the equations directly becomes next to impossible. On the other hand, understanding such phenomena as elasiticity and fracture in crystals, protein folding or docking to receptors requires simulating systems involving hundreds or thousands of atoms.

Classically, most molecular dynamics simulations solve this problem by employing fairly rough, sometimes heuristic, approximations to the atomic laws of motion. The nascent field of ML/MD attempts to get around this problem by using actual "first principles" quantum mechanics computations to first learn the potentials (or forces) that smaller groups of atoms exert on each other, and then use these learned potentials to simulate the full system. As a result of a long collaborative relationship with Gábor Csányi's group at the University of Cambridge, we can now learn state-of-the art potentials for various atomic systems [P1]. From the Machine Learning point of view, a key element of this work is how to represent atomic environments in a way that also respects the natural physical symmetries. Our paper [P2] introduced the bispectral and the so-called SOAP representations, which are widely used in the field.


[P1]  A. P. Bartok, M. C. Payne, R. Kondor and G. Csanyi:  Gaussian Approximation Potentials: the accuracy of quantum mechanics, without the electrons (Physical Review Letters 104, 2010)  [PRL] [arXiv]

[P2]  A. P. Bartok, R. Kondor and G. Csanyi:  On representing chemical environments (Phys. Rev. B 87, 2013)  [PRB] [arXiv]


DARPA Young Faculty Award for Multiresolution Machine Learning for Molecular Modeling ($500,000, Sep 2016-Sep 2018) [DARPA]