Aller au contenu principal

NORDR · · 4 min read

Physics-Aware ML Without the AI Theater

The useful part of ML in surf forecasting is not hype. It is correcting physical estimates while staying inside what the ocean can plausibly do.

The useful part of ML behind NORDR is narrow: take physical marine signals, compare them with observed outcomes, learn local bias, and keep predictions inside realistic bounds. That layer is NJORD ML, an internal research system — NORDR's shipping forecast does not consume its output today.

Physics First

Wave propagation has rules.

Period affects travel speed. Direction affects exposure. Bathymetry affects attenuation and refraction. Tide affects breaking over shallow banks. Wind affects surface quality and local energy.

Those relationships should not be thrown away because an ML model is available.

NJORD ML is framed as physics-aware because the baseline matters. The model should refine a plausible physical estimate, not invent a disconnected answer.

ML Is Best at Local Bias

Every coast has bias.

A model can overestimate wave height for one beach and underestimate another. A buoy may represent one exposure better than another. A sandbank can make a spot more tide-sensitive than the generic coastline around it.

This is where ML helps. It can learn repeating errors between offshore conditions and nearshore outcomes.

The product value is not that the model is clever. It is that the model can become locally less wrong.

Guardrails Matter

Without guardrails, a model can output nonsense with confidence.

That is unacceptable for a marine decision app. Forecasts should stay within plausible ranges. Sparse data should lower confidence. Mixed signals should produce conservative messaging. Missing inputs should not be fabricated into certainty.

This is why NORDR treats confidence and limiting factors as part of the forecast.

The User Should See the Result, Not the Machinery

Most surfers do not need to know the model family.

They need to know whether the window is worth watching, what is limiting the session, and how confident the call is.

The public language should stay practical: physics-aware estimates, location-specific guidance, timing windows, confidence signals.

njord-mlmachine-learningsurf-forecastphysics