Rip currents are a leading cause of beach-related drownings worldwide, motivating the development of reliable prediction tools that can support timely public warnings. Yet prediction skill and transferability are often difficult to simultaneously achieve because rip currents depend on both hydrodynamic forcing and local morphology, and observational constraints frequently limit the availability of training and validation data. The machine learning (ML) model implemented by the National Oceanic and Atmospheric Administration (NOAA) uses logistic regression trained on lifeguard observations, wave height, wave direction, and water level to estimate rip current likelihood. The pre-simulation (PS) model, previously used by the Korea Hydrographic and Oceanographic Agency (KHOA), incorporates additional parameters such as wave period and spectral spreading derived from FUNWAVE simulations. The ML model, although potentially less accurate at specific locations due to its lack of morphological detail, captures general trends well and can therefore be applied broadly across diverse coasts. By contrast, the PS model explicitly reflects the morphology of a specific coast and is potentially more accurate, but its applicability is limited to similar settings. This study compares the two models for rip current prediction using measured flow velocity data from Duck, North Carolina, and proposes and discusses integrated PS–ML approaches that combine their respective strengths. • Compare NOAA ML with a physics based FUNWAVE pre simulation (PS) model. • Validate both with surf zone velocity data from Duck, North Carolina. • Physics based PS captures morphology and spectral effects; ML is portable. • A shallow neural network (SNN) improves rip current probability estimates. • PS and ML integration improves agreement with observed rip current speeds.
Choi et al. (Wed,) studied this question.