Multimodal graph-based deep learning for predicting heat of combustion as a key hazard indicator under GHS/CLP standards

We report a multimodal graph-based deep learning framework that integrates molecular graph topology with physicochemical descriptors to predict the heat of combustion (HoC) directly from chemical structure. The model employs a neural network convolution (NNConv) to capture bonding and topological information, while molecular descriptors provide complementary global features. Trained on 4,516 experimental data points, the model achieved a mean absolute error (MAE) of 351.4 kJ/mol (1.89 kJ/g), corresponding to approximately 6.35% of the dataset mean (5529.65 kJ/mol), and a coefficient of determination (R²) of 0.942 on the test set. When benchmarked against linear regression, random forest, and graph-only GNN models, the corresponding MAE and R² values were 376.0 and 0.925, 381.7 and 0.914, and 1608.8 and 0.484, respectively, highlighting the benefit of integrating descriptors with graph-based learning using NNConv. The predicted HoC values enable reliable classification with respect to the 20 kJ/g GHS/CLP decision threshold, except for compounds within an ambiguity band of ± 1.89 kJ/g. Residual analysis showed low bias and an approximately normal error distribution. These results demonstrate that multimodal graph-based learning can provide decision-relevant HoC predictions, supporting chemical hazard screening while reducing reliance on experimental measurements.