Machine Learning for Physics-Based Force Fields Using the Alexandria Chemistry Toolkit

Many popular force fields rely on compensation of errors between terms, including the electrostatic energy due to the charge distribution in molecules, dispersion, exchange and induction. This means that errors in the description of electrostatic interactions are compensated by errors with the opposite sign in the other terms. By using symmetry adapted perturbation theory it is possible to decompose the interaction energy from high level quantum chemistry calculations into the terms mentioned above. Force fields can then be trained to reproduce these components rather than the total interaction energy, and in this manner stay closer to the underlying physics.

In the seminar I will describe our new software, the Alexandria Chemistry Toolkit for training force fields and show preliminary applications, for instance on modeling the exchange interaction in hydrogen-halides and water in J. Phys. Chem. Letters 15 (2024) 9974-9978.

Lecturer

David van der Spoel got his Ph.D. 1996 in the group of Herman Berendsen, one of the pioneers in the field of molecular simulations. During and after his Ph.D. work he was one of the main developer of the GROMACS MD software, which is now used by thousands of researchers. After moving to Uppsala University, Sweden he worked on a number of topics, including proteins in the gas phase, interaction of biomolecules with X-rays and simulations of liquids. Over the last 15 years his focus is developing force fields for molecular simulation from scratch.