麻豆淫院

The Gulf of Mexico, a regional ocean, is hugged by the southeastern United States and a large stretch of the Mexican coast, making it very important for both countries. The area helps bring goods to local and global markets, produces power for the country with off-shore oil rigs, and hosts a myriad of vacation-worthy beaches鈥攕o modeling and predicting its dynamics is a critical task.

Research from applied mathematicians at the University of California, Santa Cruz, presents new AI-powered methods for modeling the Gulf. They achieve higher accuracy than traditional models for short-term predictions, and successfully emulate 10-year dynamics without any AI "hallucination"鈥攁 physically impossible scenario.

This research is a result of a collaboration between the UC Santa Cruz Baskin School of Engineering researchers led by Assistant Professor of Applied Mathematics Ashesh Chattopadhyay, industry partner Fujitsu's Converging Technologies Laboratory, and a research group at North Carolina State University. Their work is in the Journal of Geophysical Research: Machine Learning and Computation.

This work drives forward critical management of natural resources in the U.S. and Mexico, and advances the technology for modeling gulf streams around the world鈥攁 major feature of the global oceans鈥攁nd demonstrates the improving efficacy of AI in the Earth sciences.

"The ability to resolve the Gulf Stream and its dynamics properly, has been an open challenge for many years in oceanography," Chattopadhyay said. "That's why the Gulf of Mexico becomes an important test case whenever we're trying to evaluate new algorithms and new models for high-resolution regional ocean dynamics."

The team's collaboration among academics and industry aims to make the products of their research ready for real-world use.

"The ocean emulators developed with UC Santa Cruz deliver a combination of speed, accuracy, and lightweight design that enables seamless operational integration into maritime platforms," said Subhashis Hazarika, principal researcher at Fujitsu Research of America.

"This supports interactive system design for applications ranging from port operations management to ship weather routing and extreme event monitoring. Our collaboration with UC Santa Cruz marks an important step toward bringing rigorously validated AI-for-Science models into real-world industrial applications."

Modeling the gulf

The Gulf of Mexico hosts important maritime industries like energy production and cargo shipping, making the modeling of this region an important safety and economic issue. Large eddies from the Gulf stream break in this area, creating rogue waves that sometimes hit the areas where people are working on oil rigs鈥攕o it's critically important to be able to model these waves and other dynamics.

It's historically been very difficult to model regional oceans, especially in areas near the coast. Waves crashing near the shore, along with other factors, makes measuring and modeling complex.

Traditional tools for ocean modeling, that are still the gold standard in industry, are high-resolution, physics-based models that are expensive, power hungry, and relatively slow鈥攁nd not always highly accurate.

AI models require more investment up front to train, but work up to 100,000 times faster than traditional models. AI has also been limited by "hallucinations," when models drift away from true physical dynamics when emulating dynamics at long timescales. Chattopadhyay's work focuses on eliminating hallucinations by integrating physics into the AI models, particularly to better capture dynamics that are physically smaller and happen on a shorter timescale.

Zooming in

To do this, the research team constructed an AI model that works with two main components. One takes a "zoomed out" view, focusing on ocean events that can be observed from eight kilometers of resolution, and happen over longer timescales. The second component takes these zoomed out predictions and enhances them to four kilometers of resolution.

Chattopadhyay compared this process, called "downscaling," to zoom enhancement of a photo.

"You can take an old picture and enhance it, basically improving the quality and resolution using a generative model," Chattopadhyay said. "We're using basically the same technology here to be able to downscale predictions to four kilometers, and we're making sure that we're not just enhancing it unrealistically."

Using this technique, the team found their system had better performance than traditional physics-based models at shorter timescales, such as making predictions 30 days ahead鈥攂eating the highly accurate traditional models.

Additionally, they found that their models can emulate up to 10 years ahead without any hallucinations. The team achieved this by rigorously integrating physics constraints into the "zoomed out" emulations, an effort led by graduate student Leonard Lupin-Jimenez.

Chattopadhyay says that their system's improved accuracy fits in with a larger trend of AI models in the sciences besting traditional models.

"I think this was one of these instances, and it's been happening more and more in this field of AI and science, that AI models are starting to outperform physics-based models," Chattopadhyay said.

Collaboration for at-sea impact

Several of Chattopadhyay's graduates are hosted by Fujitsu's Converging Technologies Laboratory as an ongoing collaboration to support interdisciplinary research鈥攁 unique partnership between physical science/academics and industry.

Lupin-Jimenez spent time at the company to focus on this project under the mentorship of Fujitsu's Subhashis Hazarika and Anthony Wong, and together they shaped the work for both research and operational use.

"There was a good amount of freedom in being able to experiment with different methods, processing, and pipeline frameworks to see what works best," Lupin-Jimenez said.

Throughout the process, they emphasized making the software as useful as possible, which meant making the system operable on a ship at sea.

"It's actually going to translate to end users who might not be either experts in physics or experts in AI, but they still want to be able to do this modeling," Chattopadhyay said. "In our group, we are trying to keep our research aligned with needs of the market and industry, especially with AI, so that research and ready-for-market tools are not divorced."