NVIDIA ALCHEMI Toolkit Powers AI Materials Simulations

NVIDIA has detailed how its ALCHEMI Toolkit, a PyTorch-native library released in early 2026, can pair with AI coding agents like Claude Code to automate materials science simulations. By providing specialized agent skills, the toolkit allows LLMs to write correct code for Machine Learning Interatomic Potentials using the MACE-MPA-0 foundation model. This setup bridges natural-language prompts and GPU-accelerated code, running on hardware like NVIDIA H200 GPUs.

To test this, researchers ran a benchmark of 45 simulation pipelines across five prompt levels. The tests evaluated three workflows: silicon equation of state, oxygen adsorption on copper, and liquid lithium self-diffusion. For silicon, the pipelines achieved a lattice constant of 5.4661 angstroms and a bulk modulus of 88.15 GPa. The copper adsorption tests found the fcc hollow most stable at -4.799 plus or minus 0.004 eV. For lithium self-diffusion, Langevin dynamics damped diffusion by 3 to 5 times compared to NVE ensembles. The L3 recipe prompt yielded 2.91 x 10-4, 5.38 x 10-4, and 7.22 x 10-4 cm2/s, while the L5 contract prompt yielded 3.62 x 10-4, 4.32 x 10-4, and 6.49 x 10-4 cm2/s at 600 K, 800 K, and 1000 K respectively.

Higher prompt specificity increases code complexity and token costs without changing physical correctness. A basic Sketch prompt processed approximately 2.4 million tokens (generating 32,000) to produce a 498-line script over 44 iterations, while a Contract prompt processed 10.0 million tokens (generating 107,000) to produce 1,168 lines over 84 iterations. Running these workflows requires CUDA 12 or 13, driver version 570 or higher, an NVIDIA GPU of RTX 20xx or newer with Compute Capability of at least 7.0, and at least 4 GB of RAM, with 16 GB recommended.

For practitioners, this automation shifts the bottleneck from writing boilerplate code to scientific validation. While agents successfully generated functional code, they did not verify physical plausibility, such as attempting to simulate lithium-ion diffusion in pristine crystalline lithium fluoride. NVIDIA emphasizes that human judgment remains essential to validate MLIP model accuracy against experimental data. Industry partners are already adopting this paradigm, with Matlantis releasing a public skills library and Dassault Systemes integrating ALCHEMI into its virtual companion, MARIE.