A Neural Network Prediction Model for Concrete Cratering Effect Incorporating Prior Knowledge

To improve prediction accuracy of concrete cratering effects, this study developed a neural network model incorporating prior knowledge. Experimental data from concrete target cratering tests were collected. Anomaly sample detection and composite data augmentation methods were applied to filter and expand the dataset. The Buckingham Π theorem was introduced to reduce input feature dimensions and optimize neural network input representation. Optimal loss functions were selected for different prediction tasks to establish the prior knowledge-integrated neural network model. Results show that the data enhancement strategy improved prediction accuracy by 0.4% for crater depth and 0.9% for crater diameter. The Huber loss function outperformed MSE (Mean Squared Error) and MAPE (Mean Absolute Percentage Error) in depth prediction by reducing outlier impacts, while MSE achieved optimal performance in diameter prediction. Compared with traditional empirical models and data-driven neural networks, the prior knowledge-integrated model demonstrates higher prediction accuracy under wide velocity range conditions and better handling of complex nonlinear relationships.