High-resolution evaluation of the wide-area distribution of airborne sea salt flux using numerical fluid dynamics simulations and machine learning

Corrosion of steel structures in atmospheric environments is a critical issue in infrastructure maintenance. Among various corrosive environmental factors, sea salt flux, which correlates with the amount of sea salt deposited on surfaces, has a particularly pronounced impact in regions where large portions of land are adjacent to the sea. The sea salt effect varies with distance from the coast, as well as with sea-surface winds, topographic undulations, and surface roughness. While computational fluid dynamics (CFD) simulations offer high spatial resolution and are valuable for estimating the spatial distribution of sea salt flux, their substantial computational cost presents a challenge for wide-area assessments. In this study, low-resolution (1km) CFD simulations were combined with machine learning (ML) to reconstruct high-resolution (200 m) airborne sea salt flux distributions across a wide area of approximately 1500 2 km 2 (encompassing most of Japan). Comparisons with low-resolution CFD data and observational datasets yielded the following key findings: (a) ML enhances the reproducibility of wind velocities that influence sea salt flux and improves consistency with observations in mountainous regions, while also significantly reducing the computational burden of CFD; (b) for efficient ML application, the optimal spatial size of input data is approximately 5 to 10 km (when using 200 m-resolution CFD data for training), and inclusion of ground-surface information, specifically elevation and surface roughness, further improves predictive performance.