Installation

SCGO is on PyPI. Use exactly one MLIP extra per environment ([mace], [uma], or [upet]); the stacks conflict.

Prerequisites

• Python 3.12+

• SQLite with JSON1 (pysqlite3-binary if your build lacks it)

• CUDA for GPU MLIPs

PyPI

pip install "scgo[mace]"    # or [uma] / [upet]
# UPET only:
# pip install 'vesin==0.6.0' --force-reinstall --no-deps

For MACE/TorchSim pip installs, ensure nvalchemi-toolkit-ops is available and avoid vesin / vesin-torch unless you are on the UPET extra.

Unsupported calculator device values raise SCGOValidationError (no silent CPU fallback).

From source

git clone https://github.com/rlaplaza-lab/scgo.git
cd scgo
conda env create -f environment.yml   # [mace,dev]; or:
# pip install -e ".[mace,dev]"         # or [uma,dev] / [upet,dev]
conda activate scgo
pre-commit install

environment.yml is MACE-oriented. Use a separate env for UMA or UPET (vesin==0.6.0 for UPET).

Optional: Sella (needs a C toolchain). scipy>=1.14,<3 for fairchem UMA.

Parallel jobs and HPC

Run folders are unique (run_YYYYMMDD_HHMMSS_ffffff), so parallel jobs under the same parent usually write different *.db files. Prefer one output directory (or scratch) per job when sharing a filesystem.

SQLite defaults to WAL off (fewer -wal/-shm issues on Lustre/GPFS/NFS). Discovery may write .scgo_db_registry.json / .scgo_db_registry.lock at the *_searches root. If flock is unreliable, avoid concurrent registry updates.

Set SCGO_LOCAL_DEV=1 or configure_logging(..., hpc_mode=False) for noisier local logs.

Direct ga_go / bh_go knobs (not accepted in run_* go_params): db_enable_expression_indexes, ga_adaptive_retry_enabled, ga_fast_prefilter_enabled.

Publishing releases

Maintainers: GitHub Actions Publish to PyPI (workflow_dispatch, confirm=publish). Configure trusted publishing for the pypi environment (and testpypi if used).