Skip to content
AI.info

The Pulse

Mat3ra adds NVIDIA ALCHEMI to browser-based materials simulations

Mat3ra users can run batched atomic simulations using NVIDIA's ALCHEMI toolkit from a browser notebook. A platform benchmark reported faster crystal-structure relaxation than a one-at-a-time approach, with results limited to one configuration and one run per case.

Mat3ra adds NVIDIA ALCHEMI to browser-based materials simulations

Mat3ra users can now run batches of atomic simulations through a browser notebook using NVIDIA's ALCHEMI toolkit, rather than installing the software or buying a GPU. The company detailed the integration and its initial benchmark in a November 12 post.

The workflow uses NVIDIA GPUs to process multiple candidate structures together. Researchers can substitute their own structures and compositions, or use scripts for dynamics and reaction paths, within the same notebook workflow.

Benchmark and its limits

In its first platform benchmark, Mat3ra reported that one NVIDIA H100 NVL GPU relaxed crystal structures 40–60 times faster than a conventional one-at-a-time approach. The post put the time for 4,000 structures at about three minutes, compared with close to three hours for the conventional route.

The test used strained, randomly displaced 64-atom silicon supercells, relaxed to 0.03 eV/Å. It compared ALCHEMI Toolkit 0.3.0 in batches of 10 to 800 structures against the Atomic Simulation Environment, or ASE, processing one structure at a time. The model was MACE-MPA-0 medium, accelerated with NVIDIA cuEquivariance.

At a batch size of 400, the post reported throughput of 44 structures per second during optimization and 21 per second end to end. The corresponding one-at-a-time rates were 1.1 and 0.34 structures per second. End-to-end throughput was measured using a different total number of structures for each batch size.

Those figures are geometry-relaxation measurements from one configuration, with one run per case. Throughput varies with the operation, structures, model, settings and hardware, the post cautioned.

From structures to results

ALCHEMI is NVIDIA's open-source framework for AI-powered atomic simulation. The toolkit can evaluate hundreds of structures in one pass, and cuEquivariance accelerates the MACE machine-learned interatomic potential used underneath, according to the post.

Alongside structural relaxation, the same batched engine can run molecular dynamics at a set temperature and pressure. Version 0.3.0 adds faster optimizers, reaction- and diffusion-path searches, enhanced sampling and a direct pipeline from generative AI models into simulation.

The benchmark workflow is contained in a Jupyter notebook. The platform builds structures from its materials library, saves them as a set and submits a job for each method to a GPU queue. It then installs the toolkit and GPU dependencies on the node and returns the results as a table.

Access and collaboration

The notebook can run in a browser through Mat3ra's built-in JupyterLite or locally through the Mat3ra API. Users can replace the benchmark structures with their own, or substitute a dynamics or reaction-path script while keeping the rest of the workflow.

Mat3ra CEO Timur Bazhirov described the workflow as modular: “our users plug in their own structures, compositions and use cases” using the setup from NVIDIA's ALCHEMI team. NVIDIA worked on the toolkit and benchmark, while Mat3ra handled platform and workflow engineering.

Mat3ra says its platform has more than 40,000 registered users. Structures, workflows, jobs and results belong to a team's organization account, where members can work from the same data. The company says the notebook will be public in its API examples repository for users to rerun or adapt.

Sources

Related stories

Explore

More articles