Supplemental: Overview of the ML Force Fields

Overview

Teaching: 0 min
Exercises: 0 min

Questions

• What are the categories of molecular dynamics force fields?

• What force fields are available?

• What systems they work best with?

Objectives

• Be able to recognize the strengths and weaknesses of different types of force fields.

• Find out which force fields are available and which systems they are most suitable for.

• Identify the original papers that introduced force fields.

• Machine Learning Force Fields
  • ANAKIN-ME (Accurate NeurAl networK engINe for Molecular Energies)
  • SO3LR
  • MACE
  • ViSNet

Machine Learning Force Fields

ANAKIN-ME (Accurate NeurAl networK engINe for Molecular Energies)

ANI-1 neural networks are trained on molecules up to up to 8 atoms (H, C, N, O)

• ANI-1x Trained on 5 million molecular conformations, ωB97X/6-31G(d).
• ANI-1ccx Trained on 500,000 molecular conformations, CCSD(T)/CBS.

ANI-2 neural networks added F, Cl, S.

• ANI-2x Trained on 8.9 million molecular conformations, ωB97X/6-31G(d).

SO3LR

Trained on 4 million molecular conformations, PBE0+MBD. SO3krates neural network for semi-local interactions combined with a pairwise force field for short-range repulsion, long-range electrostatics, and dispersion interactions. Scales to 200,000 atoms on a single GPU. Implemented in JAX-MD.

MACE

The MACE architecture combines Atomic Cluster Expansion framework with Equivariant Message Passing Neural Networks. Several pre-trained models are available.

ViSNet

Equivariant geometry-enhanced graph neural network. Developed by Microsoft Research. A Graph neural Network with Vector-scalar interactive message passing. Avaialble in Pytorch Geometric and GitHub.

Key Points

• There are different types of force fields designed for different types of simulations.

• Induction effects are not accounted for by fixed-charge force fields.

• Using more accurate and diverse target data allows MD force fields to be improved.