Teaching a New Dog Old Tricks: What ML Potentials can learn from Conventional Force Field Training

Conventional molecular mechanics force fields try to do the impossible: provide accurate simulations of observable phenomena across chemical space with typically five simple terms (stretches, angles, torsions, electrostatics, and lennard-jones), combined with even simpler mixing rules, and fit through the herculean effort of fifty years of grad student descent. Such FFs rely on falsehoods like the existence of bonds, rigid hydrogens, and the complete disregard of many-body interactions.

But just because Bonds Aren’t Real, doesn’t mean we should abandon fifty years of valuable wisdom and hard work. Developers of ML potentials have recognized the importance and elegance pairwise FF terms can bring, shown via the increasing adoption of both explicit long-range charge models and short-range analytic exchange-like repulsion terms in ML potentials. In our talk, we will present another piece of “lore” worth bringing into ML vogue: how to train for the outcomes you really care about.