Bayesian Hyperparameter Optimization of Physics-Informed Neural Networks for Marine Current Turbine Wake Prediction
Year: 2026
Type: conference-paper
Abstract: Predicting the turbulent wake behind marine current turbines from limited velocity measurements is critical for optimization of the layout of the array. Physics-informed neural networks (PINNs) can reconstruct such flow fields by coupling sparse data with governing equations. However, their accuracy depends heavily on several interacting hyperparameters whose joint effects are not well understood for high-Reynoldsnumber three-dimensional wakes. Tree-structured Parzen Estimator (TPE) based Bayesian optimization is used to search over nine hyperparameters of a PINN that solves the non-dimensional RANS equations for the wake of a 1:40 scale NREL Reference Model 1 marine current turbine at$R e \approx 5.23 \times 10^{5}$. The search covers three activation functions (tanh, swish, SIREN), network depth and width, learning rate, velocity component weights, and batch sizes. A functional analysis of variance (fANOVA) decomposition of the 25-trial optimization reveals that the activation function alone explains more than 60% of the variance in the quality of the reconstruction. After retraining each activation's best configuration for 25000 epochs, swish achieves the highest streamwise velocity coefficient of determination$(R^{2}=0.88)$, while cross-flow components remain below$R^{2}=0.15$regardless of the configuration, pointing to a limitation rooted in the RANS closure formulation rather than in the choice of hyperparameters.
Institution: Indian Institute of Technology Madras
Language: English
Subfield: Statistical and Nonlinear Physics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7201836635