Explosion pressure and duration prediction using machine learning: A comparative study using classical models with Adam-optimized neural network

The application of machine learning (ML) for the prediction of gas explosion pressure remains limited, and the prediction of the explosion duration is nearly non-existent. A series of dispersion and subsequent explosion computational fluid dynamics (CFD) simulations were conducted to determine explosion pressure and duration values. These results were used to train classical ML models, that is, support vector regression (SVR), random forest (RF), and decision tree (DT) models. Additionally, a multi-output Adam-optimized artificial neural network (ANN) model was employed for performance comparison. All the models demonstrated respectable predictions for both parameters, while the RF model demonstrated the highest performance based on the metrics analyzed, followed by the DT model. The proposed gas volume and gas volume blockage ratio (gas-VBR) emerged as the most crucial feature for predicting explosion pressure, while the monitoring point and gas-VBR was the most important feature for explosion duration. It is recommended to consider the gas-VBR feature in future studies rather than solely focusing on blockage ratio or obstacle location. The model proposed was compared with models from previous studies for predicting explosion pressure. The findings conclusively demonstrate that the multi-output model outperforms the compared models, offering a notable advantage in its ability to predict both gas explosion pressure and duration.