AI Is Changing How We
Forecast the Ocean
- Raleigh, NC
- July 9, 2026
- By: Michael Gray, Head of Engineering
Artificial intelligence is not overhyped. It’s overused. Not overused in the sense that we are applying AI too often, but because the term itself has become incredibly broad. Today, almost anything involving data, automation, or machine learning can be labeled “AI.” So let’s talk specifics about how AI is being used in ocean forecasting and on my team at Fathom Science.
I’ve spent the past several years working on this problem. Before joining Fathom Science, I was part of the Ocean Observing and Modeling Group at NC State University, where my research focused on developing OceanNet, one of the first AI-based ocean forecasting models. AI has been making its way into weather forecasting for years, but its use in oceanography is still relatively new.
I could spend a lot of time talking about neural networks and model architectures. But I think there is a more useful place to start: Why are we using AI at Fathom Science to forecast the ocean in the first place?
Simulating the Ocean Is an Enormous Computational Challenge
Computers have become dramatically faster, but ocean models are still constrained by the number of calculations they can perform. To better understand this constraint, let’s take a moment to consider the scale of computation for a typical global ocean model used by governments and industry around the world.
Global ocean models often have a resolution of roughly 9 kilometers. Covering the globe at that resolution requires around 9 million grid cells (4,300 longitudinal cells by 2,100 latitudinal cells). However, the ocean is not two-dimensional. Models also divide the ocean vertically, often into 50 or more layers. Those 9 million cells become nearly half a billion three-dimensional grid cells. It gets worse. Ocean models cannot calculate the ocean state once and stop. Currents move. Water temperatures change. Ocean features evolve. The model must continually advance forward in time, repeatedly recalculating the state of the ocean at each grid cell.
There are many technical details, which I’ll skip, but the frequency, usually called the time step, that the model does this recalculation, is on the order of 1-10 simulated minutes. If we assume five minutes, that is nearly 300 model time steps in a single simulated day. Half a billion grid cells calculated hundreds of times quickly becomes hundreds of billions of calculations for a single day forecast. Not to mention that at every grid cell the model is doing far more than a single calculation.
When you consider the computation needed, it is remarkable that we can produce global ocean forecasts days into the future at all.
For Fathom Science, this computational challenge is particularly important. Much of our work focuses on producing high-resolution ocean forecasts for specific locations and maritime applications. As model resolution increases, the number of calculations grows rapidly. Our team is focused on developing new ways to generate better forecasts without simply adding more computing power.
What If AI Could Learn From Decades of Ocean Modeling?
Ocean models have been running for decades, creating enormous archives of simulations describing how the ocean evolves. What if we used those previous simulations as training data? An AI model can analyze years of modeled ocean states and learn patterns in how the ocean changes over time such as how currents evolve, how water masses move, and how today’s ocean relates to what may happen next.
Something profound is happening here. Instead of explicitly solving the same complex physical equations every time we produce a forecast, an AI model can learn patterns from the results of those equations. Training an AI model can be computationally expensive. But once trained, forecasts can potentially be generated in seconds rather than hours or days.
This is one of the areas we are actively exploring at Fathom Science: how decades of ocean modeling and historical simulations can be used to train AI systems that complement the physics-based forecasting models we already operate.
Speed Changes What Is Possible
The value of speed is not simply getting a forecast faster. It allows us to run more forecasts. Traditional ocean models are deterministic. Start a model from exactly the same conditions, and it will produce the same forecast. This is not very useful if we are trying to predict how likely it is that the forecast will be accurate.
By contrast, If an AI ocean model can produce forecasts in seconds, we can potentially run hundreds or thousands of simulations from slightly different starting conditions, referred to as an ensemble forecast. Instead of asking, “What will the ocean look like?” We can begin asking, “What are the possible ocean states, and how likely is each outcome?”. This is particularly valuable farther into the future, when small differences in today’s ocean can lead to very different outcomes weeks or months later.
The Future Is Hybrid
Like other AI systems, AI ocean models can produce answers that look reasonable but are wrong. For a chatbot, that might mean answering a question incorrectly. For an ocean forecast, it could mean predicting conditions that are physically unrealistic. That is why our team at Fathom Science does not see AI replacing traditional ocean models. Instead, we are exploring where AI can make them better.
Ocean observations are sparse and come from satellites, buoys, gliders, ships, and other sources. AI can help fill in the gaps, providing a more complete picture of the ocean to initialize forecasts. It can also help extend forecasts weeks or months into the future, enable larger ensembles, and approximate computationally expensive parts of numerical models.
The future is not AI versus physics. It is combining the strengths of both. At Fathom Science, that is the future of ocean forecasting we are working toward. If you are interested in this work, or our products, feel free to reach out to me or others on the Fathom Science team.
- Michael Gray, Head of Engineering
