cazorazitora/benchmark-dg-ddg-for-garden

Ni-catalysed alkyne carboxylation: ΔG and ΔΔG models

Trained weights and the descriptor pipeline for three models that predict the reaction free energy (ΔG) and the regioselectivity gap (ΔΔG) of Ni-catalysed alkyne carboxylation with CO₂ over bipyridine ligands.

This repository exists to serve the artifacts to a Garden deployment: Garden rebuilds its container from a single app file, so the weights and the descriptor code have to be fetchable from a public URL at image-build time. Everything here is downloaded by wget during that build.

Contents

File	Size	What it is
egnn_best.pt	48 MB	EGNN checkpoint (weights + target normalisation + hyperparameters)
rf_dg.joblib	322 MB	RandomForestRegressor for ΔG, 800 trees
rf_ddg.joblib	145 MB	RandomForestRegressor for ΔΔG, 800 trees
rf_src.tar.gz	39 KB	The descriptor pipeline (geometry → 118 descriptors)
PROVENANCE.txt	—	Source commit and per-file transformation record