AI Training Cut Storm Surge Underestimation by Up to 97 Percent

A model trained to keep the rare surge visible

Coastal defenses are built for events that may happen once in a generation—or once in a century. Engineers therefore need to explore many possible futures, but physics-based models that simulate tides, wind, pressure and coastal geometry are computationally demanding across large climate ensembles.

A team from Politecnico di Milano, TU Delft, Deltares, Vrije Universiteit Amsterdam and the CMCC Foundation tested a different approach: train a neural network to emulate the Global Tide and Surge Model, then penalize it more heavily when it underestimates rare high surges.

In a New York City case study, the best storm-surge model achieved a test-set root-mean-square error of 4.5 centimeters against Global Tide and Surge Model output. A model trained with an asymmetric loss function reduced underestimation of simulated return-level extremes by 59% to 97% compared with conventional mean-squared-error training across return periods from one to 1,000 years.

The result is not a forecast of exactly when New York will flood. It is evidence that an AI surrogate can preserve more of the extreme tail of a physics-based simulation—the part most likely to shape resilience decisions.

Research references: [1]

Why average accuracy can hide the worst events

Most machine-learning systems are rewarded for doing well on the average case. For storm surges, ordinary hours vastly outnumber extreme ones. A model can therefore post a strong average score while smoothing down the peaks that matter for seawalls, evacuation zones and infrastructure standards.

The researchers addressed that problem with a loss function combining quantile and expectile terms. In plain language, the model received a stronger penalty for underestimating large positive surges. The authors call it weighted quantile and expectile, or WQE, loss.

They tested feed-forward, two-dimensional and three-dimensional convolutional networks, and a long short-term memory network. Inputs included wind, air pressure, temperature, mean sea level and the positions of the moon and sun. The target was a Global Tide and Surge Model node near the southern tip of Manhattan.

The strongest storm-surge configuration used a two-dimensional convolutional network, a 10 by 15 grid of atmospheric data and a 24-hour input window. A separate feed-forward model handled the astronomical tide. Combining the two produced a total-water-level test-set root-mean-square error of 5.7 centimeters against Global Tide and Surge Model output.

Research references: [1]

Training for the tail changed what the model could see

The trade-off was revealing. Conventional mean-squared-error training produced 14% lower mean-squared error than WQE training, while WQE produced 25% lower quantile loss. The method tuned for extremes gave up some average-case neatness to better represent high-water events.

Across historical return periods from one to 1,000 years, the WQE-trained convolutional model reduced underestimation of simulated return levels by 59% to 97% relative to the conventionally trained model. Its remaining underestimation was 3.4% for the one-year return level, 2.9% for the 10-year level, 3.7% for the 100-year level and 0.2% for the 1,000-year level.

Those labels describe statistical return levels derived from simulated time series. They are not promises about how often a future flood will occur. The longest return periods depend on greater extrapolation, and Hurricane Sandy strongly influences the historical extreme-value analysis.

Research references: [1]

The future test showed both promise and dependence

The team then tested whether a model trained on historical reanalysis could continue to emulate storm surges under changing-climate forcing. It compared the neural network with Global Tide and Surge Model projections through 2050.

Fine-tuning the neural network on historical simulations from the same climate model improved its performance on the future scenario. For the WQE model, fine-tuning reduced average return-level bias by 73.7% across the 10- to 1,000-year return periods. The surrogate reproduced the direction and approximate size of many changes projected by the physics-based model.

This is an evaluated model-to-model result, not an independent discovery about future New York floods. The AI was trained to emulate the Global Tide and Surge Model, so it inherits the reference model’s assumptions and the climate model’s forcing. It cannot independently validate the physical projection that taught it.

The paper’s methods and conclusion name the future forcing model as CMCC-CM2-VHR4, while one figure caption says VHR2. Because of that internal inconsistency, this analysis does not rely on the disputed suffix.

The dependence also reveals the scaling challenge. Performance under future forcing improved after fine-tuning with historical hydrodynamic simulations from the same climate model. If those simulations do not exist, the shortcut still needs an expensive reference run before it can accelerate exploration.

Research references: [1]

What this could unlock for coastal resilience

Climate adaptation decisions are rarely based on one forecast. Cities need ensembles covering emissions pathways, storms, sea-level conditions, model assumptions and protection options. A fast surrogate could help planners screen more combinations, test sensitivity to uncertain extremes and identify where full hydrodynamic modeling deserves the most attention.

The current study does not measure runtime savings, planning quality, construction costs or avoided flood damage. It evaluates numerical fidelity to one physics-based model at one New York location. The neural network also did not receive local bathymetry or coastline geometry as explicit inputs; it learned those local dynamics indirectly from simulations, making the trained model location-specific.

The practical target is a layered workflow. Physics-based models remain the source of physically consistent training data and high-stakes confirmation. AI surrogates expand the number of scenarios that can be screened. Human planners decide which uncertainties matter, which communities face the greatest exposure and which options merit deeper engineering analysis.

If later studies reproduce the result across more coastlines and climate models, AI could become a multiplier for adaptation planning: not a machine that predicts the future by itself, but one that helps cities examine more futures before concrete is poured. The authors’ open code and analysis scripts make that claim testable by other researchers.

Research references: [2]