How can we gather more insight from data to solve problems in the world today? This is one of the most fundamental questions driving the artificial intelligence boom. It’s also one that GSO Assistant Professor of Oceanography Dr. Nick Pizzo is no stranger to posing, and recently it led him and his collaborators into the pages of Nature Geoscience.
One of the ways that physical oceanographers like Dr. Pizzo can remotely observe ocean currents is using data from satellite altimeters. These satellites send signals down to Earth’s surface as they orbit, and they can infer the elevation of the surface based on the time it takes for the signal to bounce back. At large spatial scales, there is a simple relationship between the tilt of the sea surface and the surface currents, so velocities can be backed out from the altimetry data. However, satellite altimeters are only able to determine the surface elevation below them in this way and they take weeks to return to the same location. Their spatial resolution is also relatively low, which naturally limits the resolution of the currents you can infer from them.
While looking at infrared readings from geostationary weather satellites, Dr. Pizzo and his collaborators began to wonder if they could leverage that data to derive surface currents instead. These satellites work together to provide snapshots every 5 minutes at ~2-km resolution around the globe, so if they could somehow tease out velocities from the data, that would allow them to resolve features at substantially smaller scales than before, all without a single in-situ measurement or the deployment of additional hardware. However, infrared readings map back to sea surface temperature, not velocities, so extracting velocities from that data is still quite a hurdle.
“You can look through cloud covered regions and start to see that there are structures underneath,” Dr. Pizzo explained. “We’re coming from the fluid mechanics world, where if you have texture, and you have a bunch of frames…you want to do a correlation between the frames.” In other words, one might attempt to programmatically identify connections between structures present in subsequent frames in order to infer the average velocities required for those structures to change locations. Dr. Pizzo continued, gesturing to a sea surface temperature plot on his computer screen, “But if you do that in this scenario, there are all these interweaving braids – so that it’s very difficult to correlate across frames.” What is one to do in the year 2026 when you have a lot of data, a targeted question to ask, and traditional methods just aren’t cutting it? Enter machine learning!
Dr. Pizzo and his collaborators decided to attack this problem by developing a U-Net, a type of neural network designed for image segmentation. In order to train their U-Net, they first needed to generate a massive amount of training data, which they accomplished by using a run of the MIT Global Circulation Model at approximately the same spatial resolution of the geostationary satellite data: a very computationally-expensive undertaking. They also needed ample access to high-performance GPUs in order to train and run their machine learning model. For this study, computing resources were provided by the Expanse cluster at the San Diego Supercomputer Center. As is often the case at high-performance computing clusters nowadays amid the rise of AI and machine learning, hardware demand can be a major issue to navigate. “These are tools people want to use, so you have to get in line. And that’s okay!” Dr. Pizzo acknowledged. “I think we are so fortunate to live in a world where you have access to these types of resources.”
That sense of gratitude is unsurprising given both Dr. Pizzo’s personality and the wild success of the project. After a couple of years of hard work, the team’s effort and electricity paid off in the form of their Geostationary Ocean Flow (GOFLOW) product, a deep-learning product capable of deriving high-resolution surface currents from geostationary satellite data. The resolution and favorable comparison with altimetry data – see Figure 3 in the paper – was enough to make my jaw drop. “It’s like putting on glasses, right?” laughed Dr. Pizzo. The high-resolution velocity fields generated by GOFLOW also allow for the computation of quantities including vorticity (i.e. rotation of the fluid) and the notoriously difficult to compute horizontal divergence (i.e. spreading of the fluid), which can in turn give more insight into the movement of different water masses, including dynamics below the surface (e.g. upwelling or downwelling).
—
As it turns out, similar deep learning methods potentially valuable for solving a wide variety of different problems. Andy Goering, a graduate student working with Dr. Pizzo, is currently applying the method at very small scales in order to get at some of the fundamental physics relating to the interactions between the wind, waves, and currents. For example, how exactly does the wind generate the initial millimeter-scale surface wrinkles that ultimately grow into full-sized waves? A major hurdle to clear en route to answering this question is to determine flow velocities around the air-sea interface. Andy and Dr. Pizzo see room to apply machine learning to this problem as well, except this time using lab images of tiny particles injected into wind blowing over waves rather than sea-surface temperature fields, and using training data generated from Oceananigans, a GPU-ready software for fluid dynamics simulations, rather than MITgcm. There are some outstanding challenges, such as particles being more sparse near the surface and some uncertainty around how much physics they’ll need to put into the training data generation to get reasonable answers, but it is another promising application of this technique.
Although he also began his research computing efforts on Expanse, Andy has since moved over to Unity, a high-performance computing cluster housed at the Massachusetts Green High-Performance Computing Center in Holyoke, MA. URI is a partnering institution at the Unity cluster, which means that Unity is free to use for URI researchers like Dr. Pizzo and Andy. This removes a common logistical hurdle for studies that require substantial computing resources and can allow researchers’ grant dollars to go further. Andy cosigned: “It is definitely nice not having to worry about requesting more resources!” He has also found that the highest-performing GPUs, which he needs for his research, are more readily accessible on Unity and that the system has been reliable. Like Dr. Pizzo, Andy feels that his biggest research computing challenge is the competition for limited resources, and he also has a positive outlook on the inevitable bottlenecks. “Of course, we wouldn’t have access to these resources if they weren’t getting used!” he chuckled.
—
When asked for advice for others who would want to get into machine learning, or other technical research computing, Dr. Pizzo was as encouraging as ever: “It’s never been easier – if you have the data, and you feel like you need just one additional piece of information to make sense of things, we’re now in a time where you can learn things that you couldn’t in the past… There are more people to help, and there are fewer barriers. It’s just another tool for us to use to learn about how things work.” He added that there’s less stigma now than ever around doing research with models and software tools; he considers himself an observationalist, but acknowledges that these tools can allow you to extract insight from that data. “If these models are the way of the future they need legitimate data. Now we need to focus on that part of the problem.”
Dr. Pizzo seems to have a knack for seeing which way the currents are going, so I think we can take him at his word! There really has never been a better time to jump into research computing, and the Institute for AI and Computational Research at URI is here to help.
Sincere thanks to Dr. Nick Pizzo and Andy Goering for taking the time to be interviewed for this article.
Written by Josh Port.