NequIP and Allegro
NequIP (Neural Equivariant Interatomic Potentials) is the original equivariant message-passing graph neural network model for force fields. Allegro is a more recently developed equivariant model that is strictly local and therefore massively parallelizable to enable fast and accurate simulations. Both models are based on PyTorch and integrated with ASE and the LAMMPS molecular dynamics code.
Many-body NequIP extensions (e.g. MACE): https://arxiv.org/abs/2206.07697
Phoebe is a framework for ab-initio prediction of electrical and thermal transport properties by solving phonon and electron Boltzmann equations. For electron-phonon interactions, it integrates with the Quantum ESPRESSO code and for anharmonic phonon interactions it works with ShengBTE and Phono3py, which interface with most DFT packages. Phoebe is written in C++ and is designed for use on modern HPC infrastructures through hybrid MPI/OpenMP parallelization, distributed memory computation via ScaLAPACK, and support for GPU acceleration using Kokkos.
FLARE (Fast Learning of Atomistic Rare Events) is a platform for autonomous active learning of Bayesian force fields. It relies on Gaussian process regression to provide accurate and fast predictions of forces, energies and stresses and to quantitatively estimate the principled uncertainty at every step of an MD simulation.
0.5 Trillion atom surface reaction dynamics: https://arxiv.org/abs/2204.12573
News 6/21/23: FLARE is powering the "Active Learning" component of the Azure Quantum Elements, introduced by Microsoft CEO Satya Nadella: "Just imagine speeding up chemistry simulation by as much as 500,000 times so that you can rapidly screen tens of millions of candidates to find new compounds"