Machine Learning-Driven Prediction Method for Blast Performance of Irregular Charge Structures

To enable rapid prediction of the fragment power characteristics of variable-section elliptical charge structures, a prediction model based on a neural network method was established. The major axis length, shell thickness, and shell height were selected as input parameters. Experimental data were used to establish and validate the numerical simulation method, and numerical simulations of parameter-derived cases were further conducted to construct the training sample set for the model. The model performance was evaluated using a test set. Results show that the model can effectively predict the shell velocity distribution of variable-section elliptical charge structures within the parameter range covered by the training cases, and still exhibits a certain generalization capability when the input parameters exceed the maximum range of the training data by approximately 40%. Based on the model prediction results, the effects of various structural parameters on the response regulation of variable-section charge structures in both experiments and numerical simulations were analyzed. The study demonstrates that the proposed approach can rapidly obtain fragment velocity fields while maintaining satisfactory prediction accuracy, providing an effective data-driven tool for velocity assessment and parameter optimization of variable cross-section charge structures.