Experimental dipole moment prediction with a progressive 2D-3D hybrid framework and twin-pair-based diagnostic evaluation.

Accurate prediction of molecular dipole moments () is essential for modeling electrostatic interactions in solvation, molecular recognition, and materials. While modern 3D learning models can approach saturation on low-noise quantum-chemical benchmarks, performance often degrades on experimentally compiled datasets where measurement noise, stereochemical ambiguity, and conformer variability introduce a substantial gap between idealized theory and practical data. Here, we present a progressive 2D-3D hybrid framework for experimental dipole-moment prediction under heterogeneous labels. The framework is constructed in two stages. First, a strong 2D tabular CatBoost model based on SMILES-derived multi-view fingerprints is enriched with physicochemical, dipole-related, and conformer-dependent 3D shape descriptors, introducing a first level of 2D+3D integration and improving performance over the 2D representation alone. Second, this enriched tabular predictor is linearly fused with a geometry-aware 3D graph neural network operating on conformer-derived molecular graphs, yielding a further gain through complementary structural learning. On the experimental test set, the resulting framework achieves with , outperforming a nonparametric fingerprint k-NN baseline () and remaining close to the similarity-conditioned diagnostic reference range estimated from twin-pair dispersion (-0.91) under Tanimoto similarity thresholds of 0.95, 0.97, 0.98, and 0.99. SCIENTIFIC CONTRIBUTION: The main scientific contribution of this study is the development of an experimentally grounded progressive 2D-3D hybrid learning framework that integrates enriched tabular chemical representations, geometry-aware graph learning, and twin-pair-based diagnostic evaluation to predict dipole moments under heterogeneous experimental labels. To contextualize performance in this noise-affected experimental setting, we further use twin-pair analysis within a test-to-train similarity-stratified evaluation protocol, while baseline comparisons and component-wise interpretability analyses help clarify the roles of representation diversity and learner complementarity. Relative to three recently published dipole-prediction studies, the proposed framework remains competitively positioned, although these comparisons are intended only for cautious literature contextualization because the underlying datasets, label sources, and noise regimes are not directly matched. In contrast to prior studies focused mainly on lower-noise theoretical benchmarks or narrower molecular domains, the present work targets the comparatively underexplored Stenutz experimental compilation.