Predicting fire consequences with the transformer model based on multimodal feature fusion

Abstract

Accurate prediction of fire consequences is fundamental to process safety management and quantitative risk assessment in the chemical process industries. Traditional empirical and computational fluid dynamics models often struggle to balance computational efficiency with predictive accuracy, especially when considering the complex interplay between molecular properties and process conditions. To address this challenge, this study proposes a high-precision Transformer-based prediction model for estimating the radiation effect distance, a critical parameter for defining safety zones. The model is founded on a multimodal feature fusion approach, integrating SMILES encoding, molecular descriptors, and three-dimensional molecular structures to comprehensively represent chemical identity and hazard potential. A dataset was generated using industry-standard PHAST software simulations, covering diverse release scenarios for 40 flammable chemicals. The proposed multimodal Transformer model significantly outperforms traditional machine learning models. Compared with the random forest (RFR) benchmark model, its coefficient of determination (R²) has increased by approximately 15%, rising from 0.84 to 0.98. More importantly, it performs better in regression performance. The mean squared error (MSE) on the test set is as low as 0.0082, the root mean squared error (RMSE) is 0.0959, and the mean absolute error (MAE) is 0.0047. SHAP analysis confirmed the model's decision-making aligns with safety principles, identifying leak size and pressure as key hazard drivers. This confirms the model's reliability. This work successfully applies the Transformer architecture to fire consequence modeling within a process safety framework, achieving an optimal balance between computational performance and prediction accuracy. It provides a reliable and interpretable tool for enhancing industrial hazard assessment, supporting safety design, and informing emergency planning, thereby underscoring the transformative potential of deep learning in advancing chemical safety practices.

Scientific contribution

This study proposes a Transformer-based deep learning framework with multimodal feature fusion for predicting firethermal radiation distances, establishing an innovative bridge between molecular informatics and process safety engineering. The model integrates SMILES sequences, three-dimensional molecular structures, quantum-chemical descriptors, and key process parameters into an end-to-end architecture, moving beyond traditional QSPR/QPCR methods that rely solely on handcrafted descriptors. Systematic comparison demonstrates that incorporating 3Dmolecular spatial information significantly improves prediction accuracy, achieving an 18.3% reduction in meanabsolute error for late pool fires. The proposed model substantially outperforms conventional machine learning benchmarks, attaining an R² of 0.98 for jet fires—a 15% improvement over random forest—with a test set MSE aslow as 0.0082. SHAP-based interpretability analysis confi rms that the model's decision-making aligns with established safety principles and molecular orbital theory, identifying leak size and pressure as primary hazarddrivers while revealing how specific molecular features influence distinct fire types. Furthermore, a two-stage transfer learning strategy with pre-training on 130,000 molecules enables eff ective generalization to the fi reconsequence prediction task. The framework establishes a clear applicability domain, and with open access to dataand code, provides a fast, accurate, interpretable tool for industrial risk assessment, safety zone design, and emergency planning.