A modular deep learning surrogate model for simulating harmful algal blooms in complex process-based systems

Process-based models (PBMs) are widely used for simulating harmful algal blooms (HABs) but are constrained by high computational costs and parameter calibration challenges, limiting their efficiency for large-scale applications. This study develops a modular deep learning surrogate model to approximate PBM outputs while significantly improving computational efficiency and predictive accuracy. Applied to bloom-prone Daecheong Lake in South Korea during the calibration (2022) and validation (2023) periods, the framework emulates hydrodynamic (FLOW), water quality (WAQ), and phytoplankton dynamics (BLOOM) processes through a sequential structure, where outputs from FLOW serve as inputs for WAQ, and WAQ outputs feed into BLOOM, preserving key environmental interactions while reducing model complexity. Performance comparisons indicate that integrating surrogate model-generated data with probabilistic parameter optimization enhanced model performance. Surrogate model-based parameter optimization (SM-PO) achieved higher predictive accuracy than trial-and-error-based calibration (TE-PC) and TE-PC with data augmentation (DA) across all modules. For total cyanobacteria cell counts, SM-PO improved Nash-Sutcliffe Efficiency (NSE) from 0.644 (TE-PC) and 0.782 (TE-PC with DA) to 0.930 in 2022, and from 0.520 to 0.719 to 0.867 in 2023. Additionally, chlorophyll-a predictions achieved an RMSE reduction of approximately 40 % compared to TE-PC, demonstrating the effectiveness of integrating surrogate modeling with data augmentation and probabilistic parameter optimization. Furthermore, temporal dimensionality reduction significantly accelerated parameter optimization, reducing computation time by 87.5 % for hydrodynamic simulations and 96.4 % for water quality and phytoplankton modules, without sacrificing model accuracy. The modular structure enables targeted module updates, reducing retraining requirements and enhancing flexibility for different environmental conditions. Beyond accelerating parameter optimization, the trained surrogate model enables near real-time HAB forecasting. By leveraging daily observed environmental inputs, it generates one-day-ahead predictions without requiring full Delft3D simulations. The proposed framework provides a scalable and computationally efficient tool for HAB simulation, with broad applicability to various aquatic systems and potential for integration into operational water quality management. By bridging PBMs with deep learning, this approach offers an advanced framework for water resource management, ecological forecasting, and eutrophication mitigation in freshwater ecosystems.