Integration of data-driven and physics-based modeling of wind waves in a shallow estuary

Numerical models solving the wave action balance equation have been widely used to simulate wind waves. In-situ measurements, albeit sparse, are crucial to the calibration and validation of numerical wave models. In this study, a novel hybrid approach was developed by integrating a physics-based Simulating WAves Nearshore (SWAN) model with machine learning algorithms to predict wind waves in a shallow estuary. Two machine learning methods, bagged regression tree (BRT) and artificial neural network (ANN), were employed. It was found that the hybrid approach (BRT-SWAN) could be an efficient tool for modelers to identify sources of error and calibrate parameters in physics-based models. In this study, the wind direction and bottom friction coefficient were determined as the main factors causing errors in SWAN-simulated significant wave height and peak wave period, respectively. Furthermore, it turned out that BRT-SWAN-ANN (ANN trained with BRT- SWAN results) could achieve a similar level of accuracy to OBS-ANN (ANN trained with field observations of wind waves). Thus, the hybrid approach can be applied to estimate wave parameters, removing the limitation of using scarce observations in developing a predictive ANN model.

Automated summary

A novel hybrid approach was developed by integrating a physics-based SWAN model with machine learning algorithms to predict wind waves in a shallow estuary. The study found that the hybrid approach (BRT-SWAN) can be an efficient tool for modelers to identify sources of error and calibrate parameters in physics-based models. Furthermore, the study showed that the hybrid approach achieved a similar level of accuracy to the approach trained with field observations of wind waves.