Smarter AI Training Strategy Beats Leading Numerical Models at Predicting Sea Level Shifts

Coastal flooding rarely arrives without warning signs, but those signs need accurate interpretation to be useful. Sea level can shift temporarily - by centimeters to tens of centimeters - due to atmospheric pressure changes, storm winds, and ocean current anomalies, creating risks for coastal communities and disruptions to maritime operations. The time between warning and impact determines whether affected areas can prepare.

Short-term sea level prediction relies on tracking what scientists call sea level anomaly (SLA) - the difference between current sea surface height and long-term averages, derived from satellite altimetry measurements. Numerical ocean models have traditionally handled this task, but they carry persistent biases and computational costs that limit regional applications. AI-based approaches have shown promise for improving forecast accuracy, but the most capable global AI forecasting systems are designed for planetary-scale prediction and require enormous computing resources that most regional applications cannot justify.

A team from Sun Yat-Sen University, the Zhejiang Institute of Marine Planning and Design, and Pusan National University in South Korea took a different approach: rather than building a more complex model, they focused on training an existing model more intelligently. Their findings, published January 23 in the journal Ocean-Land-Atmosphere Research, demonstrate that training strategy alone can produce substantial forecast improvements.

Two Training Strategies That Change What the Model Learns

The team's model is based on Earthformer, a state-of-the-art AI architecture capable of processing spatial and temporal information in parallel rather than sequentially - a capability suited to ocean data where conditions at one location influence nearby areas. Applied to the North Pacific Ocean and trained on satellite altimetry data, the base Earthformer model performs reasonably well for very short-range forecasts but accumulates errors over longer time horizons.

The first training adjustment addressed this error accumulation at its source. Rather than training the model to predict absolute SLA values at future times, the researchers trained it to predict the daily change in SLA - the temporal tendency. This shift in prediction target captures the underlying dynamics of how the ocean surface evolves, rather than trying to predict a value that accumulates prior errors with each forecasting step.

The second strategy addressed what the team calls the "training-forecast gap." During training, the model learns to predict the next day's conditions given today's data. During actual use, it must forecast multiple days ahead by feeding its own previous predictions back into itself. This gap between training conditions and forecasting conditions is a common source of degraded performance in rollout prediction tasks.

The solution was to train the model explicitly for multi-step prediction from the start - a technique called multistep training - so that the model's learned parameters account for how errors propagate over successive predictions rather than treating each day's forecast as an isolated task.

"We developed two training strategies and used them to build a new forecasting model, which delivers much more accurate predictions than existing approaches," said Wenfang Lu, an associate professor at Sun Yat-Sen University's School of Marine Sciences and the paper's corresponding author. "Moreover, these strategies can also be widely applied to other AI-based prediction tasks."

Performance Against Current Methods

The resulting model, called Multistep-Earthformer, outperformed two comparison benchmarks. One was persistence forecasting - the naive assumption that tomorrow's sea level will look like today's, which serves as a baseline for evaluating model skill. More significantly, Multistep-Earthformer also outperformed the GLO12v4 model, the current state-of-the-art numerical ocean forecasting system from the European Copernicus Marine Service.

The performance advantage was most pronounced at medium-range forecasts - the time horizons between roughly three and ten days - where numerical models typically struggle and where improved prediction would have the greatest practical value for coastal management and maritime planning.

Generalizability and Next Steps

One of the study's broader claims is that the training strategies developed here are not specific to ocean forecasting. Any AI-based prediction task that involves rollout forecasting - generating sequences of future states from an initial condition - faces the same training-forecast gap that the multistep approach addresses. The temporal tendency approach similarly applies wherever the rate of change, rather than absolute value, contains the signal of interest.

The current work is limited to the North Pacific, and the researchers acknowledge that ocean dynamics vary significantly across regions. Different ocean basins have different current systems, temperature gradients, and boundary conditions that could require region-specific training adaptations rather than a single universal strategy.

"Our next step is to expand the study area to the global ocean and develop training strategies tailored to different ocean regions. Our ultimate goal is to further optimize these strategies and build a reliable AI-based global sea level forecasting system," said first author Jiangnan He, a graduate student at Sun Yat-Sen University.