Gaussian Process accelerated optimization for computational chemistry. ChemGP provides GP surrogate models for energy surface exploration, reducing expensive oracle (electronic structure) evaluations by 3-22x.

Paper: R. Goswami, "Bayesian Optimization with Gaussian Processes to Accelerate Stationary Point Searches," arXiv:2603.10992, Mar. 2026. doi:10.48550/arXiv.2603.10992

Two kernel types are provided: MolInvDistSE for molecular systems (operates on inverse interatomic distances, providing rotational and translational invariance) and CartesianSE for arbitrary smooth surfaces (operates directly on coordinates).

Methods

• Minimization: GP-guided local optimization with FPS subset selection, EMD trust regions, and LCB exploration
• Dimer: GP-accelerated saddle point search with L-BFGS translation
• NEB: All-image (AIE) and one-image (OIE) evaluation with per-bead FPS and RFF approximation
• OTGPD: Adaptive threshold GP dimer with HOD training data management

All methods share four mechanisms: FPS subset selection, EMD/Euclidean trust regions, RFF approximation for scalable prediction, and method-adapted LCB exploration. The PredModel enum dispatches between exact GP and RFF uniformly across all optimizers.

Benchmark Results

Surface	Method	Oracle calls	Speedup
Muller-Brown	GP minimize	7	4.9x vs direct GD
LEPS	GP minimize	9	22x vs direct GD
LEPS	GP-Dimer	13	3.5x vs standard
LEPS	OTGPD	13	3.5x vs standard
LEPS	GP-NEB AIE	62	2x vs standard
LEPS	GP-NEB OIE	49	2.6x vs standard

Building

cargo build --release
cargo test -p chemgp-core

Examples

# 2D analytical surface (CartesianSE kernel)
cargo run --release --example mb_minimize

# Collinear H + H2 reaction (MolInvDistSE kernel)
cargo run --release --example leps_minimize
cargo run --release --example leps_dimer
cargo run --release --example leps_neb

Each writes a .jsonl file with per-step convergence data.

Figure Reproduction

All manuscript figures are generated via a Snakemake pipeline in scripts/figures/. See scripts/figures/README.org for the full workflow.

Benchmark Harness

The repository now includes a dedicated benchmark subproject under benchmarks/ for minimization, dimer, and NEB comparisons. The top-level development entrypoints are:

pixi run -e dev bench-smoke
pixi run -e dev bench-minimize
pixi run -e dev bench-dimer
pixi run -e dev bench-neb
pixi run -e dev bench-all

The compact production slice is designed to stay useful within roughly 45 minutes on rg.cosmolab.

A rehydration archive is deposited on Materials Cloud Archive for exact reproduction without recomputation.

Documentation

pixi run -e docs docbld

Or visit chemgp.rgoswami.me.

License

MIT License. See LICENSE for details.