Comparative Evaluation of Deep Learning Graph Neural Networks and Classical Machine Learning Models for Predicting...

Area of Science:

• Computational Chemistry
• Materials Science
• Chemical Engineering

Background:

• Accurate prediction of heat of combustion (HoC) is crucial for fuel development and chemical safety.
• Machine learning (ML) offers promising avenues for predicting molecular properties.

Purpose of the Study:

• To systematically evaluate four ML models (XGBoost, MLP, GCN, NNConv) for predicting HoC.
• To compare the performance, efficiency, and applicability of descriptor-based versus graph-based models.

Main Methods:

• Utilized a dataset of 4516 compounds for HoC prediction.
• Implemented and compared XGBoost, Multilayer Perceptron (MLP), Graph Convolutional Network (GCN), and Neural Network Convolution (NNConv) models.
• Assessed models based on training time, test error, coefficient of determination, generalization, and predictive fidelity.

Main Results:

• Descriptor-based models (XGBoost, MLP) showed faster training and lower errors (MLP R²=0.942).
• Graph-based models (GCN, NNConv) had longer runtimes but better generalization and predictive fidelity (NNConv).
• All models reliably classified flammability under GHS/CLP criteria, with NNConv showing minimal false negatives.

Conclusions:

• A trade-off exists between computational efficiency and representational richness in molecular property prediction.
• Both descriptor- and graph-based ML models are effective high-throughput screening tools for regulatory applications.
• NNConv demonstrated superior predictive fidelity and classification performance for GHS/CLP flammability.