Building Custom Atomistic Simulation Workflows for Chemistry and Materials Science with NVIDIA ALCHEMI Toolkit

NVIDIA has released the ALCHEMI Toolkit, a GPU-native, PyTorch-based framework for building custom atomistic simulation workflows in chemistry and materials science. It bridges the gap between machine learning interatomic potentials (MLIPs) and classical simulation infrastructure by providing modular, composable building blocks including batched dynamics engines, geometry optimizers (FIRE/FIRE2), molecular dynamics integrators (Velocity Verlet, Langevin, NPT), model wrappers for MACE/TensorNet/AIMNet2, and GPU-resident data management. Key features include a FusedStage API for single-GPU pipelines with torch.compile support, a pipe operator for distributing stages across multiple GPUs/nodes, and a hook system for custom dynamics like simulated annealing. Partners including Orbital, MatGL, and Matlantis have already integrated the toolkit, reporting speedups of up to 33x for batched simulations. Installation is via pip with Python 3.11+ and CUDA 12+.