Surrogate Modeling of Atlantis: Deep Learning Approaches for Predicting and Forecasting Biomass
Faculty Supervisor:
Dr. Gokhan Kul, Computer & Information Science/Data Science
Committee Members:
Dr. Gavin Fay, SMAST / Fisheries Oceanography
Dr. Ashokkumar Patel, Computer & Information Science/Data Science
Dr. Asif Turzo, Computer & Information Science/Data Science
Abstract:
Process based marine ecosystem models such as Atlantis Provide high fidelity simulations of species interactions and environmental dynamics, but their computational cost limits their use in real time forecasting and large scale scenario analysis. This study proposes a deep learning based surrogate modeling framework to emulate Atlantis simulations and enable efficient prediction of marine framework to emulate Atlantis simulations and enable efficient prediction of marine biomass dynamics in the Northeast U.S. Large Marine Ecosystem. Unlike traditional approximation approaches, the proposed method focuses on learning spatio temporal dependencies directly from simulation outputs. Atlantis data spanning 1964 - 2020 across multiple guilds and spatial polygons are structured into temporal sequences incorporating environmental drivers such as temperature and salinity. A Bidirectional Long Short Term Memory (Bi-LSTM) architecture is employed to capture both forward and backward temporal relationships and model nonlinear ecosystem behavior. The surrogate achieves strong predictive performance, with an R2 score of 0.90 on held out test data, while maintaining consistency across ecological groupings. In addition to accuracy, the model significantly reduces computational overhead compared to full Atlantis simulations, enabling rapid multi step forecasting and scalable exploration of management scenarios.
Online
Dr. Gokhan Kul
gkul@umassd.edu
https://teams.microsoft.com/meet/27173518080619?p=89Z8ItRNZFdiQC8x2x