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

Area of Science:

• Computational chemistry
• Chemical engineering
• Materials science

Background:

• Predicting chemical combustion properties is vital for safety but challenging and expensive.
• Existing graph neural network (GNN) models lack detailed structural and fragment-based insights.

Purpose of the Study:

• To develop an explainable GNN model, Chemomile, for accurate prediction of chemical combustion properties.
• To incorporate molecular geometry and provide atom-wise contributions to property predictions.

Main Methods:

• Chemomile constructs multi-level graphs (molecule, fragment, junction-tree) from chemical geometry.
• AttentiveFP layers and particle swarm optimization (PSO) are used for model training and prediction.
• A perturbation-based method explains atom contributions to combustion properties.

Main Results:

• Chemomile demonstrates accurate predictions for five key combustion properties: flashpoint, autoignition temperature, enthalpy of combustion, and upper/lower flammability limits.
• The model provides explainability, identifying specific chemical fragments and atoms influencing combustion behavior.

Conclusions:

• Chemomile offers a novel, explainable approach to predicting combustion properties, improving safety analysis.
• The geometry-based GNN provides valuable chemical insights beyond simple property prediction.