The paper grew out of The Fifth Forum on Artificial Intelligence Oceanography, held earlier this year in Jinan, in China’s Shandong Province. The forum convened more than 300 experts, scholars, and students drawn from more than 80 institutions, making it one of the most comprehensive gatherings to date at the intersection of marine science and machine learning. Across talks and panel discussions, a clear consensus crystallized among participants: the next stage of AI oceanography will increasingly depend on models that fuse the flexibility of learning algorithms with the governing principles of ocean physics. The authors of the paper distilled those discussions into a forward-looking assessment of where the field stands and where it must go.

At the heart of that assessment lies the concept of physics-constrained artificial intelligence. In conventional machine learning, algorithms sift through enormous volumes of historical observations and model simulations, extracting statistical patterns without any explicit understanding of the dynamical laws that produce them. That approach can yield impressive results when training data are abundant and conditions resemble the past. But the ocean is an unforgiving test bed. Observations beneath the surface are sparse and expensive to collect, and environmental conditions can shift in ways that leave a purely statistical model extrapolating blindly. Physics-constrained approaches counter this weakness by embedding physical knowledge, dynamical equations, and process-based relationships directly into AI learning and evaluation, so that every prediction the system makes must remain compatible with the known behavior of rotating, stratified, wind-driven fluids.

The practical payoff, the authors argue, is threefold: improved reliability, improved interpretability, and improved performance under precisely the conditions where conventional AI falters. When observations are scarce, physical constraints act as a form of regularization, narrowing the space of plausible solutions to those the real ocean could actually occupy. When environmental conditions diverge from the training distribution, embedded dynamics give the model a scaffold to lean on, reducing the risk of physically nonsensical outputs. And when scientists interrogate a prediction, a model organized around meaningful physical relationships is far easier to diagnose and trust than an opaque network whose internal representations may correspond to nothing in the real ocean.

Drawing on the forum’s presentations, the authors organized the field’s most promising applications into three complementary classes of problems. The first is reconstructing subsurface ocean states from surface information. Satellites can observe sea surface temperature, sea surface height, and surface salinity with global coverage, but the three-dimensional structure of the ocean beneath remains largely hidden. AI systems trained on historical observations and simulations can infer subsurface temperature and salinity fields from what is visible at the surface, and studies presented at the forum showed that incorporating physical knowledge into these reconstructions makes them more reliable and more useful for research and operations alike.

The second class is long-range climate prediction, exemplified by forecasting the El Niño Southern Oscillation. ENSO emerges from coupled interactions between the tropical Pacific Ocean and the atmosphere, and its swings reshape rainfall, temperature, and storm patterns across the globe. Machine learning models have recently matched or exceeded traditional dynamical forecast systems in some ENSO prediction tasks, but the forum’s participants emphasized that learning physically meaningful relationships among ocean and atmosphere processes is what elevates these forecasts from statistical tricks to genuine scientific tools. When an AI system captures the underlying coupled dynamics, its predictions become interpretable, its errors diagnosable, and its skill more likely to persist as the climate changes.

The third class is operational wave forecasting, a domain with immediate consequences for shipping, offshore engineering, coastal management, and marine hazard warning. Waves respond rapidly to winds and currents, and forecasting them demands models that are both fast and physically sound. AI-based wave prediction systems presented at the forum demonstrated that embedding wave dynamics and process relationships into learning frameworks can deliver the speed of machine learning without sacrificing the physical consistency that operational users require. In each of the three problem classes, the pattern was the same: physical knowledge was not a constraint on AI’s power but the source of its trustworthiness.

The authors are explicit about what comes next. In their view, the field must move beyond isolated demonstrations toward AI systems that systematically combine physical constraints, uncertainty estimation, and real ocean observations. Uncertainty estimation deserves particular emphasis, because a forecast that comes with an honest measure of its own confidence is vastly more valuable to decision-makers than one that does not. A shipping company rerouting around a storm, a coastal community preparing for a marine hazard, and a climate negotiator weighing long-term risks all need to know not just what a model predicts but how much that prediction can be trusted. Building calibrated uncertainty into physics-constrained AI is therefore a central challenge for the coming years.

The longer-term ambition articulated in the paper is equally clear. The authors state that the ultimate goal is to make AI a trustworthy complement to numerical ocean models, one capable of supporting faster and more reliable ocean forecasting, climate prediction, marine hazard warning, and scientific understanding. This framing matters. Rather than positioning machine learning as a replacement for the general circulation models and dynamical simulation systems that oceanographers have refined over decades, the authors envision a hybrid landscape in which AI accelerates computation, fills observational gaps, and extracts structure from data, while numerical models and physical theory supply the dynamical foundation that keeps learning algorithms anchored to reality.

The implications extend well beyond oceanography. The tension the paper addresses, between the pattern-finding power of data-driven models and the explanatory rigor of physics-based ones, is one of the defining questions of modern Earth system science. The ocean, with its sparse observations, slow memory, and outsized influence on climate, is arguably the harshest environment in which to resolve that tension. If physics-constrained AI proves itself in the deep blue, the lessons will carry directly into atmospheric science, hydrology, cryosphere research, and the broader effort to model a planet in flux. For now, the message from more than 300 researchers gathered in Jinan is unambiguous: the future of AI in ocean science belongs not to algorithms that merely memorize the sea, but to those that learn the motion of the ocean itself.

Subject of Research: Physics-constrained artificial intelligence for ocean forecasting and climate prediction

Article Title: AI needs to learn the motion of the ocean, researchers report

Article References: AI needs to learn the motion of the ocean, researchers report. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: artificial intelligence, oceanography, physics-constrained machine learning, El Nino Southern Oscillation, wave forecasting, ocean modeling, climate prediction, subsurface ocean reconstruction, uncertainty estimation, numerical ocean models, marine hazards, AI oceanography forum