Reference Geometry Stability¶
Purpose¶
This benchmark assesses the MLIP’s capability to preserve the ground-state geometry of organic small molecules during energy minimization, ensuring that initial X-ray or DFT-optimized structures remain accurate and physically consistent.
Description¶
This benchmark performs energy minimization using the BFGS optimizer with alpha=70, maxstep=0.03 and a convergence threshold of 0.05 eV/Å.
After minimization, structural fidelity is assessed by computing the root-mean-square deviation (RMSD) of all heavy atoms relative to the initial geometry, using the RMSD implementation provided by mdtraj.
Classification of energy minimization results based on RMSD values.¶
Dataset¶
This benchmark draws its test set from the OpenFF [1] industry dataset, which contains 73,301 conformers of thousands of different molecules, collected by industry partners of the OpenFF consortium. To keep the benchmark conductable in a reasonable amount of time on standard hardware,we extracted 200 geometries of neutral molecules and 20 structures of charged molecules from the dataset. We ensured chemical diversity in this reduced dataset by applying the rdkit MaxMinPicker to the Morgan fingerprints of the original datasets.
Interpretation¶
The reference geometry stability benchmark can be understood as a sanity-check benchmark to verify that the MLIP recognizes a true geometrical minimum of a molecule as such. Often, true ground state structures, with forces on all atoms being essentially zero, are absent from training data. In this case the ability of the MLIP to correctly interpolate between near-minimum structures is tested. In any case, the ground-truth geometry should be retained and not deviate from the QM geometry. Since the QM geometry is also the starting structure, the geometry should ideally not be changed at all. We apply a strict 0.3 Å criterion for a good RMSD here, to reflect this demand.