GLIDER adapts information from MACE-POLAR, whose pretraining includes energies and forces. Its response head learns the interaction-induced potential and dipole directly.
- Extract geometry-dependent MACE-POLAR-1-M features.
- Predict complex and isolated-fragment sites with frozen M and L checkpoints. Subtract fragments and average the two resulting responses.
- Learn local corrections to the atom-centred charges and dipoles, plus a global response-dipole correction.
- Remove net response charge and distribute the mismatch between site moment and global prediction over atomic dipoles.
- Evaluate the resulting potential at any requested off-site points.
The global constraint is exact for the predicted moment. It does not make either prediction exact relative to QM, and it does not enforce dissociation.
The original 48 configurations are listed in training. The final ensemble uses seeds 2026081300–2026081304, 64 epochs, AdamW at 0.002 with weight decay 0.00002, cosine scheduling and gradient clipping at 5. The objective combines response ESP with dipole weight 2 and auxiliary fitted-site weight 0.05. Architecture and transformations are retained in the frozen implementation.
The response head was selected through nested leave-one-chemistry-out development. All 14 chemistries enter final fitting. Panels I–III never enter that final fitting operation. Panel I later trained a separate unsuccessful model family, as explained in data lineage.
The geometry wrapper assembles the official frozen M/L models and the released five-member head. The model supports the compact neutral systems evaluated here. The dissociation test demonstrates a residual distant response. The current checkpoint should not be treated as a general molecular-dynamics force field or as a validated fragment-separable response model.