A dendritic neuron-inspired nonlinear regression framework for flash point modeling across diverse liquid mixture systems

Accurate prediction of physicochemical properties for complex liquid mixtures remains challenging because molecular structure, composition, and intermolecular interactions can jointly affect property variations. In this study, a dendritic neuron regression (DNR) framework was developed to establish nonlinear structure-property relationships for flash point prediction in diverse binary liquid mixtures. A SiRMS-based molecular representation strategy was employed to characterize component-contribution and mixture-specific structural information, while composition-dependent mixing rules were used to construct mixture descriptors. The dendritic neuron architecture was adapted to continuous regression by retaining nonlinear dendritic information integration and introducing a trainable regression mechanism to capture complex relationships between molecular descriptors and flash points. The proposed framework was evaluated using four binary liquid mixture categories, including organic miscible, organic partially miscible, water-miscible, and water-partially miscible systems. Under a single 80/20 Point-out split, the DNR models achieved external predictive coefficients (Q_EXT²) of 0.9798, 0.9781, 0.9884, and 0.9957 for the four datasets, respectively, with external RMSE values ranging from 1.169 to 2.040 K. Across 25 repeated Point-out experiments, DNR achieved the highest mean Q_EXT² and the lowest mean RMSE among six models for all four datasets, with statistically significant improvements in 15 of 20 pairwise comparisons after Holm correction. Under strict Mixtures-out (M-out) validation, DNR achieved Q_M-out² values of 0.8447 and 0.9382 for Datasets 1 and 2, respectively. Compounds-out (C-out) validation revealed substantially larger prediction errors (RMSE: 8.0303-23.0532 K), including a negative Q_C-out² for the water-miscible dataset. Distance-based applicability domain analysis and Y-randomization tests further assessed model reliability. Overall, the SiRMS-DNR framework demonstrated favorable composition-level predictive performance, while generalization to unseen chemical components remained challenging.