Data-Driven Physics Informed Neural Network of Submarine Hull Shapes for Active Sonar Evasion

Navies around the world are increasingly using active sonar to detect underwater craft. Therefore, there is interest in optimizing submarine hull shapes to evade active sonar. While it is known that faceted shapes are ideal for scattering acoustic signals, developing the ideal hull shape to suit multiple needs remains difficult. Conventional simulations for hull shape optimization are often time-consuming. In this research, we developed a physics-informed neural network (PINN) for shape optimization, combining the acoustic wave equation PDE with a data-driven loss component. For the supervised portion of our PINN’s training, we generated ground truth data by running 80 active sonar simulations, where each simulation utilized a different faceted submarine hull shape. The method behind the simulation was the second order Runge Kutta method. For the unsupervised portion of our PINN’s training, we encoded the acoustic wave equation directly into the neural network. Ultimately, we trained an ensemble of three PINNs for approximately 100 epochs each, reaching a simulation MSE of 3E-05. After training, our PINN was able to evaluate 4,320 faceted submarine hull shapes in 24 hours. This was 120 times faster than the second order Runge Kutta. Our validation tests showed that the PINN was able to distinguish between shapes with good and bad active sonar scattering. With this substantial improvement, our approach shows great promise in advancing sonar evasion strategies and optimization in naval applications.