Data management keeps simulation data resident on the GPU to avoid transfers, using AtomicData and Batch objects, with native ASE and Pymatgen interfaces and Zarr-based storage for efficient batch writing.

Key integration partners include Orbital (OrbMolv2 with PME electrostatics and MTK integrator), MATGL (TensorNet integration for faster property predictions), and Matlantis leveraging Toolkit-Ops for high-throughput workloads.

The toolkit sits between domain-specific GPU kernels and deep learning models, enabling composable simulation workflows with features like geometry relaxation, molecular dynamics, and multi-stage pipelines.

It supports building end-to-end workflows with customizable dynamics classes, model wrappers, and advanced data management to minimize CPU-GPU data transfers and maximize GPU residency for large-scale simulations.

Early adopters report meaningful performance gains, including roughly 1.7x acceleration for large systems and up to 33x speedups for batched smaller systems with GPU-accelerated graph construction.

System requirements include Python 3.11–3.13, PyTorch 2.8+, CUDA 12+, Linux or macOS, NVIDIA GPUs with compute capability 7.0+, at least 4 GB RAM (16 GB recommended), along with installation steps and links to the GitHub repository and documentation.

Two scaling approaches are offered: FusedStage for single-GPU deployment and a distributed multi-GPU pipeline, demonstrated with eight GPUs for geometry optimization and eight for Langevin dynamics.

The toolkit is available on GitHub at NVIDIA/nvalchemi-toolkit, with JAX support planned for version 0.2.0.

Core capabilities include batched dynamics kernels, JAX support, and integration with neighbor list construction, DFT-D3 dispersion, and long-range electrostatics via Toolkit-Ops.

The framework supports custom model integration with wrappers and built-in support for MACE, TensorNet, and AIMNet2 architectures.

NVIDIA ALCHEMI Toolkit provides a GPU-accelerated framework to build custom atomistic workflows, bridging MLIP models with physics-based calculations for quantum-like speed and accuracy.

Usage patterns demonstrated include combining FIRE2 optimizers with VelocityVerlet dynamics, both on single GPUs and distributed multi-GPU setups, with API-oriented code examples for assembling and running batched workflows.