Automated topology design to improve the susceptibility of naval ships using geometric deep learning

Abstract The survivability of a naval ship is defined as its ability to evade or withstand a hostile environment while performing a given mission. Stealth technology, which reduces the probability of detection by enemy detection equipment using a highly advanced detection system, is one of the most important technologies to improve the survivability of naval ships. Moreover, radar cross-section (RCS) reduction is a very important factor in stealth technology because a small RCS, which is the main parameter determining susceptibility, improves the ability of ships to evade enemy detection equipment. In this study, an automated topology design for reducing susceptibility was developed by combining geometric deep learning and topology optimization. A convolutional neural network model was used as the geometric deep-learning model, and the triangular meshes of the naval ship models and equipment models were used as datasets. To compensate for the lack of training data, randomly generated meshes were additionally used as datasets. To express the feature data of the mesh as a matrix, points at equal intervals were projected orthogonally and the distance between the plane and point was set as a matrix value. The label data were defined as the highest RCS values excluding the cardinal points. After realizing the topology design for reducing susceptibility using the developed system, verification was performed through RCS analysis of the original model and the topology-designed model.