A Cummins-equation-based physics-informed neural network for response prediction and system identification of wave energy converters

This study evaluates the accuracy of a Physics-Informed Neural Network (PINN) framework for the hydrodynamic analysis of wave energy converters (WECs). While time-domain solvers efficiently predict forward hydrodynamic responses of floating systems, extracting unknown parameters traditionally requires coupling solvers with iterative optimization or purely data-driven machine learning. To bridge this gap, the proposed model embeds the governing Cummins equation directly into the network’s learning process. The convolution integral associated with fluid memory effects is resolved by linearizing the impulse response function (IRF) via the Prony approximation. Verified against WEC-Sim solutions, the PINN’s accuracy is assessed under free-decay, regular, and stochastic wave scenarios. The framework effectively solves inverse problems by identifying unknown parameters, including quadratic drag, power take-off damping, and spring stiffness, directly from input time–history data. While increasing the number of exponential IRF terms improves overall accuracy, this sensitivity is more pronounced in inverse applications compared to forward problems. Furthermore, the PINN functions as a noise-filtering tool, and the model maintains high accuracy even when trained on data containing 20% noise. This unified architecture successfully bridges predictive simulation and data-driven system identification, providing a robust foundation for the advanced performance analysis and design of marine energy systems.