AI for Science

The slowest steps in
simulation, learned.

SciLM builds open datasets and generative models that replace the most expensive steps in computational science, from transition state search in materials chemistry to 3D reservoir modeling of the subsurface.

Get in touch Browse the code

34,135,597

converged transition states in MaterialsSaddles

904 GB

of open simulation data on Hugging Face

0.05 eV/Å

saddle convergence across the full dataset

Research

Two domains, one method

Where physics-based simulation is too slow to search, we generate the data at scale, then train generative models to skip the search entirely.

Materials kinetics

Transition states at scale

Reaction rates in materials are set by transition states, and each saddle point search costs thousands of force evaluations. SaddleMill runs those searches at high throughput across GPU clusters. The results became MaterialsSaddles, 34 million converged transition states released in the open. And SaddleFlow trains on them: a flow matching model that proposes the saddle structure directly from a reactant and product pair.

Subsurface geology

Reservoir models from sparse wells

Decisions in carbon storage, groundwater, and energy hinge on 3D subsurface models built from a handful of wells. ResMill generates geologically grounded training volumes: turbidite lobes, fluvial channels, deltas. ResFlow learns from them, generating facies volumes conditioned on layer type and real well observations, the basis of the upcoming SiliciclasticReservoirs dataset.

Open source