ResNet-KAN for Small-Sample Chemical Formulation Prediction: Hybrid Architecture and Interpretability Analysis

Formulation property prediction in industrial chemistry is often hindered by severe data scarcity due to high experimental costs and confidentiality constraints, which leads to overfitting and unreliable evaluation. Here, we propose a hybrid ResNet-KAN framework for small-sample formulation modeling and demonstrate it on proprietary emulsion oil formulations (32 real formulations, 11 anonymized components) for predicting pH and defoaming performance. To improve data efficiency while preventing metric inflation, we integrate fold-wise CTGAN augmentation into a repeated nested cross-validation protocol, where synthetic (X, y) samples are generated exclusively within outer training folds and are never used for performance evaluation. Under this rigorous setting, the proposed approach achieves strong and stable generalization across both tasks, reaching R ² = 0.785 ± 0.098 for pH and R ² = 0.755 ± 0.068 for defoaming on outer test folds and consistently outperforming standard baselines and architectural variants. Beyond accuracy, interpretability is provided through complementary analyses: SHAP identifies globally influential components, while KAN-based functional and layer-level visualizations reveal explicit nonlinear response patterns and potential operating windows, enabling actionable formulation guidance without disclosing proprietary identities. Overall, this work presents a leakage-controlled augmentation and interpretable modeling strategy for data-scarce industrial formulations in precision chemistry.