Abstract
Process based marine ecosystem models such as Atlantis ecosystem model provide high fidelity simulations of species interactions and environmental dynamics. However, their computational complexity makes them expensive to run, limiting their use for real time forecasting, rapid scenario testing, and ecosystem management decisions. This creates a critical gap between model accuracy and practical usability, particularly in dynamic systems such as the Northeast U.S. Large Marine Ecosystem, where timely predictions are essential for fisheries management. The goal of this thesis is to develop a deep learning based surrogate based model capable of approximating Atlantis biomass simulations while significantly reducing computational cost for ecosystem forecasting and fisheries scenarios analysis. This study proposes a deep learning based surrogate modeling framework that learns spatio temporal patterns directly from Atlantis simulation outputs. Biomass data spanning 1964 – 2020 across multiple ecological guilds and spatial polygons are structured into temporal sequences, augmented with environmental drivers such as temperature and salinity. A Bidirectional Long Short-Term Memory (Bi-LSTM) architecture is used to capture both forward and backward temporal dependencies and model nonlinear ecosystem behavior. Unlike traditional surrogate approaches that rely on simplified statistical approximations, this method leverages sequence learning to preserve complex temporal dynamics across ecological spatial dimensions. The proposed surrogate model achieves strong predictive performance, with an R2 score of approximately 0.94 on held out test data. The model demonstrates stable performance across multiple ecological guilds and spatial regions, indicating its ability to generalize across heterogenous ecosystem components. This framework provides a practical tool for researchers and policymakers involved in fisheries management. By replacing computationally expensive simulations with a fast and scalable surrogate, the model enables rapid scenario evaluation, real time forecasting, and improved decision support. The approach offers a pathway towards integrating data driven methods with models, making complex ecosystem simulations more accessible and operational for management and planning.