机理模型与数据驱动模型交互的爆炸后果快速预测

This study aims to propose a mechanism and data-driven model fusion method for rapid consequence forecast of explosions in specific chemical clusters. Based on the premise of specific scenes, the Machine Learning (ML) technique is used for explosion modeling. To solve the issue of lacking blast data for ML learning, numerical modeling validated by an explosion experiment is employed to generate benchmark data. In the data-driven modeling process, a partition modeling and progressive learning scheme are designed to ensure the model's predictive accuracy and efficiency. Since numerical simulation relies on Partial Differential Equations (PDEs), which governs the explosion process, this paper refers to the validated numerical model as a mechanism model. During the development of the data-driving model, a zoning modeling strategy is designed to improve the training efficiency and enable the model to be deployed in small devices. Meanwhile, a self-learning strategy is adopted to continually improve the model's accuracy. Furthermore, the model's predictions are combined with existing blast damage criteria to estimate the potential consequences of explosions occurring in this specific scene. Taking Vapor Cloud Explosion (VCE) as an example for illustration, a data-driven model for predicting peak overpressure at any position was developed, using the Improved Generalized Regression Neural Network (IGRNN) by an enhanced sparrow search algorithm. Explosive parameters and the spatial position of the observation point were selected as the model's inputs. The self-numerical program was used to provide high-fidelity data for IGRNN learning. To validate the numerical method, an experiment of gas explosion in the underground pipe gallery was conducted. After validating the reliability of the numerical model, extensive numerical simulations of VCEs in typical scenes of chemical clusters were conducted, thus obtaining the IGRNN model. The results from numerical data show that the IGRNN model can predict the full-field overpressure accurately and with an inference time of 7. 53 s by using one common CPU. Finally, a case study was presented to demonstrate the feasibility of the proposed method.