A Physics-Informed Fingerprint for Generalizable Prediction of Supramolecular Stability

The predictive design of supramolecular systems remains constrained by a tension between the complexity of high-dimensional descriptors and the prohibitive cost of first-principles simulations. From a physical chemistry perspective, supramolecular stability arises from the statistical distribution of noncovalent interactions, not from single geometric or thermodynamic parameters. Here, the Kulkarni-NCI Fingerprint (KNF), a physics-informed, nine-feature descriptor, is introduced that captures the essential signatures of noncovalent interactions in a compact and interpretable form. Applied to a data set of 2,649 complexes, the KNF achieves state-of-the-art predictive performance (R² = 0.793), representing a 47% improvement over existing structural descriptors while reducing dimensionality by 3 orders of magnitude. A single ″universal model″ trained on the KNF exhibits strong cross-domain generalization, accurately capturing both hydrogen-bond- and dispersion-dominated regimes (R² = 0.955 ± 0.027 on S66, R² = 0.688 ± 0.069 on DES). To overcome the computational bottleneck of quantum-mechanical KNF generation, a 3D-enhanced Graph Attention Network (GAT) surrogate is developed, capable of predicting the entire nine-dimensional KNF vector with near-perfect fidelity on seven critical features (R² > 0.78), delivering a 416-fold acceleration (6.9 s → 16.6 ms per complex). This enables exhaustive screening of 34,776 binary combinations in 2.6 min versus 18.5 h. This end-to-end framework─from descriptor design to AI acceleration─offers a scalable, interpretable, and high-throughput pathway for the rational discovery of complex molecular materials.