Physics-guided graph attention surrogate modeling for wave-structure interactions of large-scale floating photovoltaic arrays

Hydrodynamic simulations of multi-body floating photovoltaic (FPV) arrays become increasingin size and can consist of thousands of floaters. High-fidelity simulations, such as computational fluid dynamics (CFD), may require many hours of computation for a single operating condition, thus limiting their use in iterative design and risk assessment. In this study, we propose a physics-guided graph attention network (Physics-GAT), a graph attention surrogate that accurately predicts heave and pitch responses whilst providing attention-based diagnostic interpretability. The model conceptualizes an FPV array as a graph, with floating bodies represented as nodes and their coupling relationships as edges. Additionally, it incorporates three types of physics-guided priors into the attention mechanism: geometric topology, wave-propagation direction, and phase-related propagation encoding. Besides, this approach guides neighborhood aggregation toward patterns aligned with the imposed wave-propagation prior, thereby reducing the likelihood of propagation-inconsistent shortcut-like aggregation in conventional data-driven graph attention surrogates. On the selected dataset, Physics-GAT achieves the lowest global RMSE under independent and identically distributed (IID) settings while maintaining competitive high-response accuracy. Moreover, under moderate wave-parameter out-of-distribution (OOD) shifts defined by q = 0.8, it maintains good predictive performance, and its attention remains clearly aligned with the imposed directional prior according to the proposed diagnostic metrics. It is worth noting that stronger physics-guided attention priors may reduce model flexibility under narrower training distributions or configuration shifts, reflecting a trade-off between physics-guided inductive bias and OOD generalization capability. After training, Physics-GAT predicts the hydrodynamic response of a floating solar array in a wave condition within milliseconds using an NVIDIA RTX 4070 GPU, which potentially supports future real-time implementation requirements such as digital twins.