Abstract
The prediction of blast loading for complex structures using deep learning requires extensive training data from field experiments or numerical simulations. However, the destructive nature of explosions complicates the collection of adequate field data, and traditional simulations are often time-consuming. To address these challenges, a Bayesian deep learning approach is proposed that quantifies prediction uncertainty. This method utilizes an uncertain selection strategy to actively choose high-quality samples, enhancing the simulation process and iteratively expanding the training dataset. The experimental results demonstrate that this Bayesian deep active learning method achieves a mean absolute percentage error (MAPE) of 6.1% for peak overpressure predictions. Additionally, more than 73.1% of confidence intervals include true values, with prediction times under 20 ms for single-point blasts. Notably, only 60% of the training data is required to achieve the same accuracy as conventional deep learning methods. This approach facilitates rapid and reliable predictions of blast loading for complex structures while significantly reducing training costs.