A Physics-Informed Fingerprint for Generalizable Prediction of Supramolecular Stability

The rational design of supramolecular systems is a central challenge in materials science, yet predictive modeling is often hampered by a trade-off between the black-box complexity of high-dimensional descriptors and the domain-specific limitations of specialist models. Brute-force featurization strategies inevitably suffer from informational dilution, obscuring the critical, localized physics of non-covalent interactions. This study introduces a definitive solution to this long-standing problem: the Kulkarni-NCI Fingerprint (KNF), a compact, 9-feature, physics-informed descriptor engineered to be both informationally dense and interpretable. On its native domain of 2,600 Deep Eutectic Solvent complexes, the KNF demonstrates robust and reliable predictive accuracy with R2 = 0.793, representing a 47% relative improvement over state-of-the-art structural descriptors. A particularly notable outcome is the KNF’s demonstrated capacity for generalization, suggesting promise for broader supramolecular applications. By training a single “Universal Model” on a pan-chemical dataset uniting DES complexes with the chemically diverse S66x8 benchmark, this work demonstrates that one model can master the distinct physics of both hydrogen-bond- and dispersion-dominated domains simultaneously, with R2 for DES ≈ 0.69 and R2 for S66 ≈ 0.96. SHAP analysis suggests that the KNF provides a clear signature, allowing the model to align with domain-specific physical rules and supporting a form of implicit domain adaptation in chemical AI. Finally, the KNF’s universal applicability is confirmed on the S30L benchmark of large host–guest systems, where a domain-specific model achieves notable performance with R2 = 0.803. This research provides a framework for developing generalizable and interpretable models, positioning the KNF as a promising tool for rational materials design.