Impact performance of an annular shaped charge designed by convolutional neural networks

The design optimization of an annular shaped charge (ASC) is a complicated task. The traditional design parameters of ASC are mostly determined empirically. Even with the emergence of numerical simulation technology, we cannot easily find the best scheme as it is impossible to conduct all the simulation for any given point in the design space. To address this issue, in this work, a framework combining the results calculated by finite element method with convolutional neural networks (FEM-CNN) for design optimization of an ASC is proposed. First, the finite element software AUTODYN is used to simulate of the formation of ASCs. The training and testing data are generated by numerical simulation of 720 ASCs with different liner configuration, the number of which is further expanded to 72000 dues to data augmentation technology. Then, the collected data is employed to train and test a CNN with sixteen convolutional layers, five max-pooling layers and three fully connected layers, and the well-trained one is used to predict the optimal parameters of annular liner. The well-trained CNN with various ideal projectiles generates the same predicted values of annular liner, that is, b = 0.79 mm, e = 1.23 mm, and f = 1.02 mm. The numerical simulation and experimental results indicate that the ASC designed by CNN has a good performance of penetration into stiffener targets, and the targets tested in this work are successfully penetrated through by ASCs. •A framework integrating finite element method with convolutional neural networks methodology for design optimization of an annular shaped charge is proposed.•The CNN architecture designed for optimization of ASCs is stable and converged reliably in training and testing.•The numerical simulation and experimental results indicate that the ASC designed by CNN has a good performance of penetration into stiffener targets.