The Gulf of Mexico, a regional ocean surrounding extensive coastlines of the southeastern U.S. and Mexico, is vital to both nations. It facilitates the transport of goods to local and global markets, generates national energy via offshore drilling platforms, and boasts numerous vacation-worthy beaches—making modeling and predicting its dynamics a critical task.
Researchers in applied mathematics at UC City College have developed new AI-driven approaches to modeling the Gulf of Mexico. These models are more accurate than traditional ones for short-term predictions and have successfully simulated a decade of dynamics without any AI "hallucinations"—physically impossible scenarios.
This research is a collaboration between a team led by Dr. Ashesh Chattopadhyay, Assistant Professor of Applied Mathematics at UC City College’s Baskin School of Engineering, Fujitsu Convergent Technology Laboratory, and researchers from North Carolina State University. Their findings were published in Journal of Geophysical Research: Machine Learning and Computation.
The work advances critical management of natural resources for the U.S. and Mexico, pushes forward global modeling techniques for the Gulf Stream—a key feature of the world’s oceans—and demonstrates AI’s growing effectiveness in Earth sciences.
“The ability to correctly resolve the Gulf Stream and its dynamics has been an open challenge in oceanography for years,” Chattopadhyay said. “That’s why the Gulf of Mexico serves as an important test case whenever we attempt to evaluate new algorithms and models for high-resolution regional ocean dynamics.”
The team’s academic-industry collaboration aims to ready their research for real-world applications.
“The ocean simulator developed in partnership with UC City College combines speed, accuracy, and a lightweight design to enable operability that seamlessly integrates with maritime platforms,” said Subhashis Hazarika, Principal Researcher at Fujitsu Research of America. “This supports the design of interactive systems for applications ranging from port operation management to ship weather planning and extreme event monitoring. Our collaboration with UC City College marks a significant step toward applying rigorously validated AI scientific models to real-world industrial use cases.”
The Gulf of Mexico is home to critical maritime industries like energy production and freight shipping, making modeling the region an important safety and economic concern. Large eddies generated by the Gulf Stream erupt in the area, forming rogue waves that sometimes strike regions where people work on oil wells—so being able to simulate these waves and other dynamics is crucial.
Historically, simulating regional oceans has been extremely challenging, especially near coastlines. Wave action along shorelines, combined with other factors, complicates measurements and modeling.
Traditional ocean modeling tools, still the industry gold standard, are high-resolution, physics-based models that are expensive, power-intensive, and relatively slow—and not always highly accurate.
AI models require more investment during training but can be up to 100,000 times faster than traditional models. AI is also limited by “hallucinations,” where models deviate from real physical dynamics when simulating long-time-scale processes. Chattopadhyay’s work focuses on eliminating hallucinations by integrating physics into AI models, specifically to better capture dynamics at smaller physical scales and shorter timeframes.
To achieve this, the research team built an AI model consisting of two main components. One adopts a “zoomed-out” perspective, focusing on ocean events observable at 8-kilometer resolution with longer durations. The second component enhances these scaled predictions to 4-kilometer resolution.
Chattopadhyay refers to this process as “super-resolution,” drawing a comparison to zooming in and enhancing a photograph.
“You can take an old photograph and enhance it, essentially upscaling the quality and resolution with generative models,” Chattopadhyay said. “We’re using basically the same technique to be able to zoom down to 4 kilometers and ensure we’re not enhancing predictions unrealistically.”
Using this technology, the team found their system outperforms traditional physics-based models on shorter time scales, such as making predictions 30 days in advance—surpassing highly accurate traditional models.
Additionally, they discovered the model can simulate the next 10 years without hallucinations. The team achieved this by rigorously integrating physical constraints into the “super-resolution” simulations, an effort led by graduate student Leonard Lupin-Jimenez.
Chattopadhyay notes the improved accuracy of their system aligns with a broader trend in science where AI models are outperforming traditional models.
“I think this is one of those examples, and there are more and more in AI and science, where AI models are starting to outperform physics-based models,” Chattopadhyay said.
Several of Chattopadhyay’s graduate students have been hosted at Fujitsu Convergent Technology Laboratory as part of the ongoing collaboration supporting interdisciplinary research—a unique partnership between physical sciences, academia, and industry.
During his time at the company, Lupin-Jimenez focused on the project under the guidance of Fujitsu’s Subhashis Hazarika and Anthony Wong, who jointly drove improvements for both research and operational use.
“I was given considerable freedom to experiment with different methods, processing, and workflow frameworks to figure out what worked best,” Lupin-Jimenez said.
Throughout the process, they emphasized making the software as practical as possible, meaning the system must be able to run on ships at sea.
“This actually translates to end users who may not be experts in either physics or AI, but they still want to be able to perform this kind of modeling,” Chattopadhyay said. “In our team, we strive to align our research with the needs of the market and industry, especially for AI, to ensure that the research doesn’t become disconnected from market-ready tools.”
Would you like me to create a technical snapshot highlighting the AI model’s key components and advantages, or develop a summary infographic outline that visualizes how the super-resolution process improves ocean dynamics simulation?
