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

Area of Science:

• Computational Chemistry
• Machine Learning
• Chemical Engineering

Background:

• Accurate prediction of heat of combustion (HoC) is crucial for chemical safety and process design.
• Traditional methods often rely on extensive experimental data, which can be time-consuming and costly.
• Integrating diverse chemical information sources can improve predictive model performance.

Purpose of the Study:

• To develop a multimodal deep learning framework for predicting heat of combustion (HoC) directly from chemical structures.
• To evaluate the framework's performance against existing machine learning models and experimental data.
• To assess the utility of predicted HoC values for regulatory classification, such as GHS/CLP.

Main Methods:

• A graph-based deep learning model incorporating neural network convolution (NNConv) for molecular graph topology.
• Integration of physicochemical descriptors to capture global molecular features alongside topological information.
• Training and validation on a dataset of 4,516 experimental heat of combustion data points.

Main Results:

• The multimodal model achieved a mean absolute error (MAE) of 351.4 kJ/mol (1.89 kJ/g) and R² of 0.942 on the test set.
• Outperformed linear regression (MAE 376.0, R² 0.925), random forest (MAE 381.7, R² 0.914), and graph-only GNN models (MAE 1608.8, R² 0.484).
• Predicted HoC values enabled reliable classification for GHS/CLP hazard assessment, with a small ambiguity band of ±1.89 kJ/g.

Conclusions:

• Multimodal graph-based deep learning effectively predicts heat of combustion (HoC) by combining structural and descriptor information.
• The developed framework offers a reliable alternative to experimental measurements for chemical hazard screening.
• This approach supports efficient and accurate safety assessments in chemical development and regulation.