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

Area of Science:

• Process Safety Engineering
• Chemical Engineering
• Artificial Intelligence in Chemistry

Background:

• Traditional fire consequence models struggle to balance accuracy and computational efficiency.
• Accurate prediction of fire radiation distances is crucial for process safety management and risk assessment in the chemical industry.
• Existing methods often fail to capture the complex relationship between molecular properties and fire behavior.

Purpose of the Study:

• To develop a high-precision Transformer-based prediction model for estimating fire radiation effect distance.
• To integrate multimodal data, including molecular structure and process parameters, for enhanced hazard prediction.
• To provide a reliable and interpretable tool for industrial hazard assessment and safety design.

Main Methods:

• Developed a multimodal feature fusion approach using SMILES encoding, molecular descriptors, and 3D molecular structures.
• Utilized a Transformer deep learning architecture for prediction.
• Generated a dataset using PHAST software simulations for 40 flammable chemicals under diverse release scenarios.
• Employed SHAP analysis for model interpretability.

Main Results:

• The Transformer model achieved a coefficient of determination (R²) of 0.98, a 15% improvement over the random forest benchmark.
• Achieved low regression errors: MSE of 0.0082, RMSE of 0.0959, and MAE of 0.0047.
• SHAP analysis confirmed model decisions align with safety principles, identifying leak size and pressure as key hazard drivers.

Conclusions:

• The Transformer-based model offers an optimal balance between computational performance and prediction accuracy for fire consequence modeling.
• This deep learning framework provides a reliable, accurate, and interpretable tool for enhancing industrial hazard assessment and safety planning.
• The study demonstrates the transformative potential of deep learning in advancing chemical safety practices.