NVIDIA ALCHEMI Toolkit-Ops

GPU-accelerated primitives for atomistic simulations and atomic-scale modeling.

Get started →

Key Capabilities

• Neighbor lists: naive, cell-list, and cluster-tile searches, with batching, selected-atom queries, differentiable distances and vectors, and matrix or COO outputs.

• Dispersion corrections: two-body DFT-D3(BJ) with environment-dependent coefficients, including real-space and particle-mesh Fourier implementations.

• Electrostatics: direct Coulomb, Ewald, PME, and DSF calculations, with slab corrections and PyTorch multipole electrostatics.

• Molecular dynamics: velocity Verlet and Langevin integration, thermostats, and constant-pressure integration.

• Geometry optimization: FIRE, FIRE2, and L-BFGS methods for atomic relaxation and supported variable-cell calculations.

• Segment operations: reductions and transformations of grouped data, including sums, means, dot products, and matrix-vector products.

• Framework integration: NVIDIA Warp kernels with PyTorch and JAX interfaces for supported operations.

• Repeated calculations: batching, reusable setup and storage, and compilation or CUDA Graph execution on supported routes.

Backend, differentiation, and compilation support varies by operation. See the component guides for details.

Explore the toolkit

• User Guide — installation and component guides.

• Examples — runnable examples of toolkit operations.

• Benchmarks — performance results and methodology.

• API — function and class reference.

• Changelog — release history.