headgraphic
loader graphic

Loading content ...

Advancing Gulf of Mexico Operational Forecasting
with Application to Fisheries, Industry Safety, and Natural Hazards (GOFFISH)

P.F.J. Lermusiaux, P.J. Haley, Jr.,
C. Mirabito, V.A. Rodriguez

Massachusetts Institute of Technology
Center for Ocean Engineering
Mechanical Engineering
Cambridge, Massachusetts

Lead PI:
E. Chassignet

Florida State University

Project Summary
Ongoing MIT-MSEAS Research
Additional Links
MSEAS GOFFISH-supported Publications
Background Information

 

This research is sponsored by the The National Academies of Sciences, Engineering, and Medicine.

Project Summary

This project, under lead PI Eric Chassignet (FSU), aims to achieve measurable and significant improvements in short- to medium-range (1- to 10-day), subseasonal, and long-range (3- to 6-month) prediction skill of sustained continuous operational forecasts of Gulf of Mexico (GoM) ocean dynamics. The specific objectives are to i) Improve the prediction skill for Gulf of Mexico operational forecast models run within NOAA, U.S. Navy, and industry environments by increasing and enhancing model and DA capabilities and by maximizing the use of existing and new near-real-time observations, including those adaptively deployed; ii) Develop new derived products and tools from observations and numerical models to provide skillful operational forecasts and knowledge of full-water-column currents (including upper ocean circulation and near-bottom currents); and iii) Apply improvements in hydrodynamic forecasting capabilities to hurricane forecasting and fisheries management.

The specific objectives of the MIT-MSEAS component of the research are described below.

Background information is also available below.

Top of page

Ongoing MIT-MSEAS Research

Specific Objectives:

Real-Time Sea Exercises

We will be conducting real-time sea exercises in the Yucatan Strait and Gulf of Mexico during February 2024. We provide varied data processed by our group, the MSEAS-PE ocean field forecasts (plus interactive maps), and more.

MASTR Real-Time Sea Exercise page

Publications

MSEAS GOFFISH-supported Publications

Top of page


Additional GOFFISH Links

Top of page


Background Information

Our project team plans to develop and apply our data-assimilative probabilistic ocean modeling and ML tools to quantify uncertainties, predictability, and forecast risks for end-users, to complete adaptive sampling studies, and to augment Navy and NOAA prediction systems.

Probabilistic ocean modeling, forecasting, and analysis

To forecast uncertainties and quantify risks, we plan to start with our large-ensemble error subspace statistical estimation (ESSE) ensemble of simulations with perturbed initial conditions, boundary conditions, and stochastic forcing (Lermusiaux et al., 2007, 2017a).

DO-PEs. We then expect to use differential dynamical-orthogonal (DO) equations for optimal reduced-order stochastic predictions (Sapsis and Lermusiaux, 1999; Ueckermann et al., 2013; Feppon and Lermusiaux, 2018a,b). DO systems can provide predictions of LC probabilities with a massive number of realizations (e.g., the equivalent of 105–107 Monte Carlo forecasts), allowing accurate computation of probability density functions (pdfs) due to uncertainties in the forcing and initial and boundary conditions. As a result, we can then compute accurate predictability limits and predictive capabilities based on information-theoretic metrics. Several of these metrics employ mutual information (Lermusiaux et al., 2017b) that naturally handles non-Gaussian effects, the relations between vectors instead of scalars, and the expected observation errors.

Risks. Using the forecast pdfs, we plan to assess future risks for end-users (e.g. oil and gas industry), using the users’ risk profiles (e.g. Subramani and Lermusiaux, 2019). We plan to complete such risks assessments for several GoM background states, in accord with end-users needs.

Principled adaptive sampling

We intend to develop and test principled adaptive sampling schemes for physical dynamics in the GoM, with an emphasis on mutual information-based criteria (Lolla, 2016; Lermusiaux et al., 2017a,b). For ocean platforms, we plan to first predict the regions that observing assets could reach in the future, e.g. for several days to weeks in the GoM, using our path planning and reachability forecasts (Lermusiaux et al., 2016, 2017). Within these reachable regions, we then predict the locations and paths that provide the most information about the ocean variables, regions and times selected by end-users, using dynamic programming and information-theoretic methods (Lermusiaux et al., 2017; Lolla, 2017; Lin, 2021). We also expect to investigate the impact of deep data on improving the longer-terms (1-3 months) forecasts. We can also include acoustical and biogeochemical dynamics in such studies. Adaptive sampling questions that we can answer include: i) What, where, and when to observe to maximize the predictive capability of a LC extension event or eddy formation that is relevant to end-users? ii) How many in-situ platforms are needed for a given region, dynamics, and accuracy? iii) What is the ideal combination of remote and in-situ sensing for given forecasting or learning objectives?

Bayesian learning, neural closure modeling, and machine learning

Bayesian Learning. We plan to develop and use principled data-driven Bayesian learning and deep neural networks for improving forecasts and learning model formulations themselves. Specifically, we expect to use dynamically-based Bayesian learning (Lu and Lermusiaux, 2014, 2021; Lin and Lermusiaux, 2021; Gupta and Lermusiaux, 2021b) to learn model formulations and parameterizations from sparse data, integrating over state variables and parameter values, and enabling a principled Bayesian update of model distributions. This learning method can select the most accurate parameterizations among those used by existing ocean modeling systems, for example, submesoscale closures, biogeochemical-fish interactions, or oil spill model parameterizations.

Neural closure modeling. We also plan to use novel delay-differential-equations-based ML framework to learn time-delayed closure parameterizations for low-fidelity models using data from high-fidelity simulations or high-resolution observations (Gupta and Lermusiaux, 2021a). For this project, we intend to employ and extend our neural closure modeling schemes to study and improve the surface parameterizations in the existing ocean modeling systems, learning from high-resolution surface data sets and high-resolution ocean simulations.

ML for transports. New ML schemes, e.g. LSTM and graph neural networks (GNNs), are also planned to be used to predict transport features and coherent structures using SST or SSS fields, without a priori knowing the underlying flow field (Kulkarni, 2021; Kulkarni and Lermusiaux, 2021). To achieve this, we plan to combine SST, SSH, and SSS data with trajectory data to infer and predict surface transport properties for end-user applications in the GoM (e.g., surface pollution, search and rescue).

Machine Learning with ROMS and DA. We also intend to augment the ROMS DA system by applying and developing new XAI and ML techniques to improve the fidelity and computational efficiency of the DA system, to extract information from both observations and model output, and to post-process system output for feature detection and extreme event prediction. Neural networks are planned to be used in DA packages to assimilate LC and associated eddies position from satellite ocean color images.

Top of page


Go to the MSEAS home page