Chemomile: Explainable Multi-Level GNN Model for Combustion Property Prediction

Measuring the combustion properties of potentially hazardous chemical compounds is critical to preparing safety guidelines or regulations but is often challenging and costly. Developing precise prediction models for the combustion properties is, therefore, an issue of importance in both industry and academy. Previous studies reported promising models based on graph neural networks (GNNs) and message-passing architectures. However, these models often neglect the hierarchical and three-dimensional structure of chemical compounds and do not provide chemical information like which fragments of the compound contribute to the combustion properties. In this study, we introduce Chemomile, an explainable geometry-based GNN model specifically designed for combustion property prediction. Chemomile constructs multiple graphs for each chemical compound using its molecular geometry: molecule-level, fragment-level, and junction-tree-level graphs. We employ multiple AttentiveFP layers for multiple graphs to make the final prediction of the combustion properties. Chemomile is optimized using particle swarm optimization (PSO) and benchmarked against five combustion properties (flashpoint, autoignition temperature, enthalpy of combustion, and upper and lower flammability limits). We use a perturbation-based explanation method to quantify the atom-wise contribution to the properties, thus providing valuable information on how the chemical structure and each atom influence the overall combustion properties.