Can AI Weather Models Accurately Predict Extreme Events Like El Niño?

A strong El Niño is testing how far AI weather models can be trusted when conditions fall outside anything they’ve learned from. Agriculture, energy and financial markets have all handed a chunk of their planning to AI forecasts, and some scientists doubt those models will hold up once the weather flips the script.

The US National Oceanic and Atmospheric Administration (NOAA) issued an El Niño Advisory on 8 October and expects a strong-to-very-strong event to last through winter.

Why A Historic El Niño Makes For A Hard Exam

NOAA has reported sea-surface temperature anomalies above +4.0°C in parts of the eastern equatorial Pacific. Its outlook puts the chance of a historic event, stronger than any El Niño since 1950, at 83% for October to December.

That becomes a hurdle because AI weather models learn from past climate records, and record-breaking conditions are outside anything the algorithms have studied.

Strong On Ordinary Days, Shakier At The Extremes

Even the scientists raising concerns acknowledge that AI models handle everyday forecasting well. The worry is that rare, extreme events are where strong routine accuracy doesn’t guarantee reliable predictions.

A peer-reviewed trial of the GraphCast system in Bangladesh illustrated the pattern. It performed well on standard rainfall forecasts a few days ahead but struggled with extreme downpours of 100mm to 300mm; false alerts were triggered, and the worst of the rain was missed.

The study emphasised that AI must incorporate better data on extreme events alongside hybrid physics-based methods to deliver reliable high-impact forecasts.

The European Centre for Medium-Range Weather Forecasts (ECMWF) reached a similar conclusion after evaluating rainfall forecasts in Ecuador. Machine-learning tools like AIFS and GraphCast performed well on broad statistical measures. However, every model struggled to predict extreme rainfall, whereas higher-resolution physical models did a much better job tracking the size and spread of major storms.

The Case For Merging Physical Equations With Machine Learning

Conventional weather modelling works by solving the fundamental equations of atmospheric movement. This keeps predictions grounded in physics, though the process requires heavy computing resources and still struggles with localised convection and regions with sparse data.

Neural networks learn from large amounts of past data as opposed to solving those equations. That delivers fast results and strong accuracy on broad measures, but rare climate anomalies don’t show up often enough in the training data, and it can be hard to see why a model reached a particular forecast.

Ultimately, the industry is moving toward a combined approach, retaining physical models as the foundation of forecasting while leveraging AI to help measure atmospheric risk.

The Rising Commercial Pressure On AI Climate Risk Platforms

As NOAA points out, El Niño swings affect agriculture, firefighting strategy and coastal infrastructure, which makes seasonal forecasts central to gauging drought and rainfall hazards.

Once AI forecasts feed energy and commodity trading decisions, a bad one carries a cost. That makes selling an opaque prediction a tough pitch for climate tech businesses unless they can set out their confidence thresholds, system limits, testing data and physical-model backups.

Buyers in insurance, energy, agriculture and commodities can ask how a product performed on rare events instead of how it averaged in a season, which puts a premium on probabilistic outputs, scenario ranges and transparent validation.

A forecast product that can show its workings has an easier case to make.

What Winter Will Reveal

Evidence on how AI models cope with the event will build over the coming months.

The agency also cautions that a stronger El Niño makes major climate impacts more likely without making them a certainty. As a result, all forecasters must work with probabilities instead of absolutes.

As the season plays out, forecasts can be checked against outcomes as they happen, and that record is likely to do more to settle the argument than any study of past events.