← News
Pollution Wins Pollution Wins Planet

How AI Is Learning to Predict Ocean Waves — With a Little Help from Physics

How AI Is Learning to Predict Ocean Waves — With a Little Help from Physics
19.26 Cm Wave-height error
5.01% Error reduction
2017–2020 Training period
2022 Test year

A single buoy reading, combined with wind and depth data, can now predict the full shape of ocean waves 12 hours into the future with a median error of just 19.26 cm in significant wave height — and when guided by a simple physical rule, that error drops by another 5.01%. This isn’t from a supercomputer running a full ocean model. It’s from a compact AI system trained on just three years of Baltic Sea data, nudged toward realism not by brute-force simulation, but by a humble 12-hour relaxation time borrowed from wave physics.

That’s the quiet breakthrough in Sander Rikka’s 2026 study on data-driven wave forecasting. The number itself — 19.26 cm — sounds modest. But in the context of maritime safety, offshore engineering, and coastal resilience, it represents a shift: machine learning is no longer just mimicking nature. It’s beginning to understand it, even if only in fragments. And when those fragments are anchored in real physics, even weakly, they become more reliable than pure pattern recognition.

This isn’t about replacing global wave models like WAM or SWAN. It’s about augmenting them — or bypassing them — in places where computational resources are limited, where rapid updates are needed, or where local conditions defy coarse-scale assumptions. The Baltic Sea, with its irregular coastlines, shifting fetches, and wind-dominated waves, is the perfect testing ground. Here, waves don’t travel far. They’re born and die quickly, shaped by local winds over short distances. That makes them predictable in principle — but hard to simulate accurately with traditional models tuned for open oceans.

Rikka’s model doesn’t solve the full spectral action balance equation. Instead, it uses a long short-term memory (LSTM) network — a type of recurrent neural network good at sequences — to predict how a wave spectrum evolves hour by hour. It starts with one observed spectrum from a buoy, then uses local wind speed, fetch (the distance over which wind blows), and water depth to project forward. The inputs are simple. The output is rich: a full one-dimensional wave spectrum at each hour, from which you can extract not just wave height, but peak frequency, energy distribution, and storm structure.

The real innovation isn’t the architecture, though. It’s how the model learns. Rikka tests two strategies: data-driven supervision and physics-guided constraints. The first teaches the AI to match observed wave properties — variance, peak frequency, peak amplitude. The second embeds simplified physical rules, like how quickly waves should relax toward equilibrium with the wind.

And here’s where it gets interesting: the best model for predicting wave spectra isn’t the same as the best model for predicting wave height.

When the AI is trained only to minimize spectral error — the difference between predicted and observed energy at each frequency — the winning strategy is to supervise it on total wave variance and peak frequency. That combination yields the lowest spectral RMSE: 0.40 m²/Hz. But if your goal is to predict significant wave height (Hs), which integrates total energy, you’re better off also supervising the model on peak spectral amplitude. That variant, A3, achieves the lowest Hs error: 18.78 cm — though at the cost of slightly higher spectral error.

This trade-off matters. In applications like ship routing or offshore crane operations, Hs is the critical metric. But for understanding storm dynamics, resonance risks, or long-term coastal erosion, the full spectral shape is essential. You can’t optimize for both at once. The choice of training objective shapes what the model learns — and what it misses.

Then comes the physics. Rikka introduces two types of physical guidance. The first, P1, imposes upper bounds on wind-sea energy and peak period based on equilibrium wave growth theory (the Young–Verhagen relations). The idea is simple: waves can’t grow faster than physics allows. But the result is counterintuitive: these bounds increase both spectral and wave-height errors. Why? Because real waves often exceed theoretical limits during rapid wind changes, especially in fetch-limited seas like the Baltic. Enforcing equilibrium too strictly makes the model less accurate, not more.

The second physical constraint, P2, is different. Instead of capping wave growth, it nudges the model toward a relaxation process: the idea that wave energy gradually approaches equilibrium with the wind over time. This is encoded as a simple ordinary differential equation (ODE):

where is spectral variance, is the equilibrium variance for current wind and fetch, and is a relaxation time. Rikka calibrates empirically to 12 hours — meaning, on average, waves adjust to new wind conditions over about half a day.

When this relaxation term is added to the loss function, something remarkable happens: wave-height RMSE drops by 5.01% to 19.26 cm, outperforming even the best data-driven model. The spectral error increases slightly, but the trade-off is worth it for Hs-focused applications.

Even more striking is P3: extending the relaxation constraint beyond the 12-hour prediction window, all the way to 24 hours — even though there are no spectral observations to learn from in that second half. The model is essentially being told: "even if you’re not being graded on hours 13–24, behave as if physics still applies."

And it works. P3 achieves the lowest spectral RMSE of any physics-guided model: 0.39 m²/Hz, a 2.31% improvement over P2. This suggests that physical consistency, even when unobserved, stabilizes the model’s behavior and prevents drift. It’s a small nudge, but one that pays off.

Spectral and Wave-Height RMSE by Lead Time

Comparison of models across 12-hour forecast window

Spectral and Wave-Height RMSE by Lead Time
LabelValue
Persistence0.63
A00.42
A20.4
A30.41
P20.41
P30.39

The benefits aren’t uniform. In weak, steady winds, the relaxation constraint changes little. But in strong winds — whether increasing, decreasing, or steady — the gains are clear. During wind ramps, when waves are catching up or decaying, the physics-guided model adapts better than pure data-driven ones. This is exactly when traditional forecasts often struggle: during transitions.

Normalized Spectral Error by Wind Condition

Error relative to observed RMS, by wind class and station

Normalized Spectral Error by Wind Condition
LabelValue
Low Speed0.81
Middle Speed0.65
High Speed0.49
Increasing Wind0.58
Decreasing Wind0.56

The study also reveals station-specific patterns. The Gulf of Finland (GoF), a narrow, sheltered basin, has lower absolute wave energy but higher relative errors. The Bothnian Sea and Northern Baltic Proper, more exposed, see clearer benefits from physical guidance. This hints at a broader truth: one model doesn’t fit all seas. Local hydrodynamics matter — and so should local tuning.

So what does this mean beyond the Baltic?

First, it challenges the assumption that more physics means better forecasts. Hard constraints — like upper bounds — can backfire if they’re too rigid. But soft guidance — gentle nudges toward plausible behavior — can improve accuracy, even with imperfect assumptions.

Second, it shows that AI doesn’t need to simulate everything to be useful. A simple ODE, derived from decades-old wave theory, can outperform complex supervision schemes when embedded in the right way. This is the promise of physics-informed machine learning: not replacing physics, but using it to guide learning where data is sparse or noisy.

Third, it opens the door to lightweight, deployable forecasting systems. Traditional spectral wave models require high-resolution wind fields, complex numerics, and supercomputers. This LSTM model, by contrast, runs on modest hardware. With buoy data and public wind reanalysis (like ERA5), it could be deployed at any coastal station in the world — from the South China Sea to the Great Lakes.

Imagine a network of smart buoys, each running a local AI forecaster, updated hourly, sharing predictions across regions. Or a mobile app for fishermen in the Philippines, warning of dangerous wave buildup before the swell hits. Or a port authority in Rotterdam using real-time spectral forecasts to schedule cargo operations during lulls in storm energy.

We’re not there yet. Rikka’s model is a hindcast — it uses perfect wind data from the past. Real forecasting requires imperfect wind predictions, which add their own error. And the model hasn’t been tested on independent stations; transferability remains an open question. The retained frequency band also limits Hs estimates, meaning the full wave climate isn’t captured.

But the direction is clear. As climate change intensifies storms and sea levels rise, the demand for accurate, localized wave forecasts will only grow. Traditional models will remain essential for global coverage. But at the edges — in harbors, estuaries, and small seas — AI-guided systems could fill the gaps, offering faster, cheaper, and surprisingly accurate predictions.

The ocean is chaotic. But it’s not arbitrary. And when machine learning respects that, even in small ways, it stops being just a pattern matcher. It starts becoming a partner in understanding.

The Science

Sander Rikka, a researcher at Tallinn University of Technology, set out to test how different forms of physical knowledge affect short-range wave predictions. The study focuses on the Baltic Sea, using hourly wave spectra from three Datawell buoys: in the Bothnian Sea (BS), Gulf of Finland (GoF), and Northern Baltic Proper (NBP). These stations span different wave climates — from exposed open waters to sheltered basins — allowing the model to learn diverse behaviors.

The core model is a four-layer LSTM: three encoder layers process the initial spectrum, wind, and fetch; one decoder layer generates hourly predictions over a 12-hour window. The model predicts logarithmic spectral increments, which are accumulated to form the next hour’s spectrum. This autoregressive design means the model feeds its own output back in — a setup that mimics real forecasting but risks error accumulation.

Training data spans 2017–2020, validation in 2021, and testing in 2022. Winds come from ERA5 reanalysis — a global atmospheric dataset — ensuring consistent, high-quality input. Fetch is calculated from high-resolution coastlines, and depth from bathymetric grids. All inputs are local, avoiding the complexity of spatial wave propagation.

The key innovation lies in the loss function. Rikka compares seven variants:

  • A0–A3: Data-driven supervision on spectral variance, peak frequency, and peak amplitude.
  • P1: Empirical upper bounds on wind-sea energy and period.
  • P2–P3: A relaxation ODE that nudges wave energy toward wind equilibrium over 12 or 24 hours.

The relaxation time hours was chosen empirically to balance fidelity and stability. This isn’t a first-principles derivation; it’s a data-informed approximation of how fast Baltic waves adjust.

What They Found

The results reveal a hierarchy of accuracy depending on the objective:

  • Best spectral fidelity: A2 (variance + peak frequency supervision) achieves 0.40 m²/Hz RMSE.
  • Best wave-height accuracy: A3 (adding peak amplitude) hits 18.78 cm Hs error.
  • Best physics-guided performance: P2 reduces Hs error to 19.26 cm, a 5.01% improvement over the matched A2 reference.
  • Best long-term consistency: P3 cuts spectral RMSE to 0.39 m²/Hz by extending relaxation to 24 hours.

Persistence — simply repeating the initial spectrum — serves as a baseline. All models beat it, but the gains vary. P2 achieves 51.55% skill in wave height (i.e., 51.55% reduction in RMSE vs. persistence), while P3 reaches 36.35% skill in spectral accuracy.

Wind conditions modulate these gains. In strong winds (>9.7 m/s), P3 reduces spectral error by up to 2.8% relative to A2. In weak, steady winds, the benefit vanishes. This makes sense: in calm conditions, waves change slowly, and even naive models can track them. But during wind transitions — when waves are growing or decaying — physics guidance helps the model stay on track.

Why This Changes Things

Wave forecasting has long been dominated by physics-based models like WAM and SWAN, which solve the spectral action balance equation on large grids. These are powerful but computationally expensive, requiring expert tuning and high-resolution inputs.

Rikka’s work shows that a much simpler approach — a compact AI model guided by minimal physics — can achieve competitive accuracy at a fraction of the cost. More importantly, it demonstrates that physics doesn’t have to be all-or-nothing. You don’t need to simulate every wave interaction to get useful forecasts. Sometimes, a single ODE — a crude approximation of reality — is enough to keep the model honest.

This has implications far beyond the Baltic. In developing coastal nations, where computational resources are limited, such models could provide life-saving storm warnings. In offshore wind farms, they could optimize turbine maintenance during lulls. In coral reef management, they could predict wave-driven sediment transport.

It also reframes the role of AI in Earth system science. The goal isn’t to replace physics, but to complement it — using data to learn what’s hard to simulate, and physics to constrain what’s easy to get wrong.

What’s Next

Several questions remain. Can this model work with forecasted winds, not just reanalysis? Can it generalize to new locations without retraining? How does it perform during extreme events, like the Baltic’s 8-meter storms?

The study also leaves open the question of which physical constraints help — and which hurt. Upper bounds failed here, but might work in open oceans. Other forms of guidance — energy conservation, directional spreading, or nonlinear dissipation — could be tested.

Ultimately, the future may lie in hybrid systems: AI models that handle local adaptation, nested within global physics-based frameworks. Or decentralized networks of AI buoys, each learning its own patch of ocean, sharing insights in real time.

The sea doesn’t care how we predict it. It only cares that we do — accurately, reliably, and where it matters most.