Scope

Property-conditioned inorganic crystal generation and candidate triage
Local CUDA workflows, hosted Azure AI Foundry inference, adapter-based extension
Generation, evaluation, and adapter fine-tuning only — DFT-level physics validation is out of scope and should be handed to computational chemistry teams

Prerequisites

Windows users: run inside a WSL2 distro (Ubuntu recommended) with NVIDIA drivers installed on the Windows host (not inside WSL). The MatterGen/MatterSim prototype was developed and tested on WSL2/Ubuntu; native Windows Python/PowerShell is not supported for the local CUDA path. See Microsoft's WSL install guide and NVIDIA's CUDA-on-WSL guide. The Azure AI Foundry hosted path still requires WSL2 for the local CLI tooling used to call it.
Python 3.10 (3.11+ breaks torch_cluster wheels) with pip and uv
CUDA-capable GPU (16 GB VRAM recommended) — or Azure AI Foundry access for the hosted path
Git LFS installed before cloning the MatterGen repo
MatterSim installed separately for evaluation (pip install mattersim)