<scp>SigmaFormer</scp> : Augmenting transformer encoders with <scp>COSMO</scp> sigma profiles for pure component property prediction

Abstract Transformer‐based molecular models pretrained on SMILES strings demonstrate strong performance in property prediction. However, these model often lack explicit integration of molecular surface charge distributions that govern intermolecular interactions such as hydrogen bonding and polarity. SigmaFormer addresses this limitation by augmenting a pretrained encoder with a 53‐dimensional descriptor derived from COSMO quantum‐chemical calculations, including the σ‐profile, cavity surface area, and cavity volume. Benchmarking across 25 thermophysical and environmental properties indicates that SigmaFormer achieves the highest average (0.836) and the best in 9 of 25 properties among five models. Performance improvements are most pronounced in properties governed by intermolecular interactions, with phase transition and environmental/toxicity categories exhibiting average of +2.30% and +1.03%, respectively. Monte Carlo dropout uncertainty is reduced in 19 of 25 properties. Data‐efficiency experiments identify three distinct contribution modes, and gradient‐based interpretability analysis demonstrates that the model leverages σ‐profile regions in a thermodynamically consistent and property‐specific manner.