36 configurations: three solutes × three parent snapshots × 1, 3, 6 and 12 waters. Each size uses the same solute and nested neighbour list for its parent. The original response training contains both three- and four-water environments. Six and twelve waters exceed that maximum.
Paper: Fig. S17a,b. Full paper map.
| Open | Contents |
|---|---|
| Geometries | ExtXYZ coordinates, solute identities, water counts and configuration registry. |
| QM references | Full per-configuration reference potentials, probe coordinates and dipoles. |
| Predictions | Frozen outputs, grouped by method, on the same probe coordinates. |
| Results | Recomputed per-configuration response errors. |
| Solute results | Configurations averaged within solute. |
| Summary | Equal-solute aggregate metrics. |
python scripts/reproduce/score_experiment.py --experiment shell_sizeRun from the repository root. The script reads the raw reference and prediction arrays and writes a fresh result under build/recomputed/shell_size/. Units and NPZ fields.