MOFSeq-2: A Space-Group-Aware MOF Sequence Representation for Free Energy Prediction

Predicting the thermodynamic likelihood of metal–organic frameworks (MOFs) at scale requires representations that are both compact and structurally informative. We introduce MOFSeq-2, a sequence representation that augments the local chemical and global topological information in MOFSeq-1 with an explicit space-group token. Using the same LLM-Prop backbone and data splits as MOFSeq-1, MOFSeq-2 reduces strain-energy MAE from $0.623$ to $0.470$ $\\mathrm{kJ/mol_{MOFatom}}$ during pretraining and reduces free-energy MAE from $0.789$ to $0.645$ $\\mathrm{kJ/mol_{MOFatom}}$ after fine-tuning, corresponding to relative reductions of 24.5\% and 18.3\%, respectively. For downstream evaluation, MOFSeq-2 achieves an F1 score of 97.77\% and a ROC AUC of 98.49\% for threshold-based synthesizability classification, and an overall polymorph-selection accuracy of 79.68\%. The largest polymorph-selection gain occurs for nearly degenerate structures, where the free-energy difference is $0.16$ $\\mathrm{kJ/mol_{MOFatom}}$. These results show that explicitly encoding crystallographic symmetry can provide a useful inductive bias for sequence-based MOF property prediction. We have made checkpoints, code, and data publicly available for easy reproducibility.