Abstract
Characterizing the internal airflow of an oscillating water column (OWC) wave energy converter normally requires repeated wave-flume experiments, and every additional chamber geometry or wave condition adds to an already substantial testing cost. This paper examines whether machine learning surrogates can stand in for part of that campaign by reconstructing the full time-varying duct air velocity waveform, not merely a scalar performance index. Seven surrogates—Gaussian process regression, Kriging, radial basis function interpolation, random forest, support vector regression, an artificial neural network (ANN), and a voting ensemble—were trained on 75 experimental cases measured on a 1:10-scale OWC model spanning three chamber geometries, five wave periods, and five wave heights. A 21-dimensional feature set combining harmonic phase encodings with physics-informed descriptors was paired with an interpolation-based split, so the test set contained wave heights never seen in training. Random forest gave the strongest global accuracy (RMSE 3.23 m/s,MAE2.24 m/s, R2 0.87, NRMSE 0.08) at 0.02 ms per sample, whereas group-level errors varied widely, from an NRMSE of 0.04 to 0.28, depending on how geometry and wave period interact. Waveform inspection tied the difficult cases to attenuated extrema, local phase shifts, and unresolved within-cycle fluctuations—errors that a validation-based linear calibration, by construction, could not repair. The results support fast surrogate-based screening of OWC operating conditions while showing why waveform-level evaluation is indispensable for energy assessment.