CGAN-based prediction on underwater explosion response of hull side panel structures

To tackle the problems of inefficient traditional numerical simulations and high-cost experimental tests in evaluating underwater explosion (UNDEX) damage of ship side grillages, this study proposes a predictive framework integrating acoustic-structure coupled finite element simulation with machine learning. Parametric modeling and computation were first performed using the acoustic-structure coupled method, a dataset comprising 1,920 cases covering varied combinations of standoff distance, charge weight, hull side structural parameters, and attack angles was generated. The Random Forest (RF) algorithm is employed for classifying the damage mode to two categories, i.e., with and without crack, and an accuracy of 94.62% of the classification model is attained. A conditional-generative adversarial network (CGAN) model is then introduced to construct a direct mapping between input parameters (including explosion and structural parameters) and structural responses. The spatial distribution contour plots of stress, strain, and deformation were directly generated. CGAN predictions for spatial distribution patterns of deformation, stress and strain show strong consistency with original images. Furthermore, Gradient Boosting Regression Trees (GBRT) are employed to predict key quantitative indicators (maximum stress, maximum strain, maximum deformation, and crack area) of the panel structures subjected to UNDEX, achieving the coefficient of determination R²>0.98 and mean absolute percentile error (MAPE) <4%. The results demonstrate that the explosion attack angles and charge weight are the key factors that dominate the structural responses. The CGAN framework reduces the prediction time for a single scenario from hours to seconds, providing an efficient and reliable tool for ship blast-resistant design and rapid damage assessment.