Physics-Guided Neural Forecasts of Nearshore Harmful Algal Blooms in the California Current System

Harmful algal blooms (HABs) are increasing in frequency, duration, and extent along the California coast, driven by climate variability, nutrient enrichment, and complex physical–biogeochemical interactions. Forecasting HAB development and spread remains a challenge, especially in Eastern Boundary Upwelling Systems where advection, stratification, and episodic river inputs strongly shape bloom dynamics. Existing approaches often trade physical realism for statistical flexibility, limiting generalization across bloom regimes. We present a physics-informed deep learning framework for nearshore chlorophyll-a forecasting in the California Current System, integrating multi-sensor satellite products, atmospheric and ocean reanalysis fields, and static geospatial predictors. Three architectures are evaluated: a convolutional long short-term memory network (ConvLSTM), a Temporal Fusion Transformer (TFT), and a physics-informed ConvLSTM (PINN) incorporating the two-dimensional advection–diffusion equation as a soft training constraint. A multi-year, 4 km-resolution dataset (2003–2021) is processed via a tailored feature engineering pipeline with quality-controlled gap-filling, rolling statistics, lagged predictors, and climatology-based anomalies. Models are assessed with strict spatiotemporal cross-validation, emphasizing spatial fidelity, bloom footprint representation, and predictor interpretability. Post-hoc explainability analyses identify key environmental drivers consistent with known upwelling–bloom linkages in the region. We present comparative skill assessments, spatial bias analyses, and predictor attribution results, highlighting the advantages and trade-offs of adding physical constraints to coastal HAB forecasting models. This work delivers a scalable and transferable methodology with direct implications for ecosystem management, fisheries, and public health.