NVIDIA ALCHEMI Toolkit#

GPU-first Python framework for AI-driven atomic simulations.

NVIDIA ALCHEMI Toolkit gives you a unified, composable API for machine-learned interatomic potentials — from single-GPU prototyping to distributed high-throughput production. Wrap any MLIP, assemble multi-stage simulation pipelines with Python operators, and let inflight batching keep your hardware fully utilized.


Who Is This For?#

Computational Chemists

Run batched geometry/cell optimization, molecular dynamics (NVE/NVT/NPT), or multi-stage relaxation-equilibration-production workflows — all from a single Python script.

Data structures →

ML Researchers

Plug your own potential into the framework with BaseModelMixin, compose it with existing force fields, and generate training data through GPU-buffered trajectory capture.

Model interface →

HPC Engineers

Scale from one GPU to an entire node with DistributedPipeline. Inflight batching and size-aware sampling handle load balancing automatically.

Dynamics pipelines →

Get started →


Highlights#

  • Bring your own model — wrap MACE, AIMNet2, or any PyTorch MLIP in a few lines with a standardized ModelCard interface.

  • Compose, don’t configure — fuse stages on one GPU with +, distribute across GPUs with |, and inject behavior at nine hook points per step.

  • GPU-native dataAtomicData and Batch are Pydantic-validated, jaxtyping-annotated graph structures that live on-device.

  • Inflight batching — converged samples are replaced mid-run so the GPU never idles during high-throughput screening.

  • Zarr-backed I/O — write trajectories with zero-copy GPU buffering; reload through a CUDA-stream-prefetching DataLoader.

Read the introduction →


User Guide#

Models#

Examples#

Change Log#

API#