Prediction model for chemical explosion consequences via multimodal feature fusion

TL;DR

This paper introduces a new model to predict chemical explosion consequences using molecular and scenario features, offering improved accuracy for risk assessment.

Contribution

The novel Bayes-Transformer-SVM model with multimodal feature fusion improves prediction accuracy for chemical explosion consequences.

Findings

The model achieved an R2 of 0.9475 and RMSE of 0.1139 on test data.

It outperforms existing models in predicting explosion consequences.

The model supports safety-by-design by enabling risk assessment for new chemicals.

Abstract

Chemical explosion accidents represent a significant threat to both human safety and environmental integrity. The accurate prediction of such incidents plays a pivotal role in risk mitigation and safety enhancement within the chemical industry. This study proposes an innovative Bayes-Transformer-SVM model based on multimodal feature fusion, integrating Quantitative Structure–Property Relationship (QSPR) and Quantitative Property-Consequence Relationship (QPCR) principles. The model utilizes molecular descriptors derived from the Simplified Molecular Input Line Entry System (SMILES) and Gaussian16 software, combined with leakage condition parameters, as input features to investigate the quantitative relationship between these factors and explosion consequences. A comprehensive validation and evaluation of the constructed model were performed. Results demonstrate that the optimized Bayes-Transformer-SVM model achieves superior performance, with test set metrics reaching an R2 of 0.9475 and RMSE of 0.1139, outperforming alternative prediction models. The developed model offers a novel and effective approach for assessing explosion risks associated with both existing and newly developed chemical substances. The model enables rapid explosion consequence assessment for chemical storage or transport scenarios, supporting safety-by-design frameworks.

This study constructed a Bayes-Transformer-SVM model for predicting the consequences of hazardous chemical explosions. The model utilized SMILES encoding and Gaussian16 quantum chemical descriptors, combined with leakage condition scenario parameters, achieving excellent performance. Its core lies in the establishment of a multimodal fusion theoretical framework, breaking through the limitations oftraditional cross-modal correlation analysis; the development of an optimized architecture that combines Transformer feature extraction and SVM regression; highlighting the potential application of the model in chemoinformatics; and enabling the prospective assessment of the explosion risks of unknown chemicals, supporting a safety-oriented design concept.

The online version contains supplementary material available at 10.1186/s13321-025-01060-x.

Figures