MIT Builds AI That Can Imagine Extreme Weather We Have Never Seen Before

 

MIT engineers have developed an artificial intelligence system designed to forecast extreme weather events that have never actually appeared in a region's historical record.

Instead of learning only from past disasters, the new method can generate statistically plausible scenarios that are more extreme than anything previously recorded.

That could change the way governments, cities, utilities, insurers, and disaster-response agencies prepare for climate emergencies.

The system was developed by Kai Chang, a mechanical engineering graduate student at MIT, and Professor Themis Sapsis, who holds the William I. Koch Professorship in Mechanical and Ocean Engineering.

Their method is called Extreme Event Aware learning, also referred to as η-learning.

The research was published in Nature Communications on August 20, 2026.

The Problem With Predicting Disasters Using the Past

Most disaster-risk models depend heavily on historical data.

If planners want to estimate what a major flood, typhoon, wildfire, or heatwave might look like, the model usually studies previous extreme events and tries to project similar conditions into the future.

But there is an obvious problem.

What happens when the next disaster is worse than anything we have ever recorded?

Traditional systems struggle because they often require examples of extreme events before they can reliably model them.

Chang explained that many existing approaches assume the most severe events in a dataset already provide enough information to estimate future risk.

But nature does not have to repeat history exactly.

The next catastrophic event may be larger, longer, stronger, or geographically different.

What Does a Disaster That Has Never Happened Look Like?

Professor Themis Sapsis uses Hurricane Katrina as an example.

A catastrophic storm like Katrina may occur every few decades.

But what would an even rarer storm look like?

What would happen during a once-in-a-century event?

How much worse could rainfall become?

How large an area could be affected?

How long might the storm last?

Those are exactly the kinds of questions the MIT system is designed to answer.

The objective is not to predict the exact date of a future disaster.

Instead, the AI asks:

What extreme scenarios are physically and statistically possible, even if we have never seen them before?

That distinction is important.

AI Generates Maps of Possible Future Disasters

The system can produce detailed maps showing hypothetical extreme weather events.

Each generated scenario can include estimates of:

• intensity
• duration
• geographic coverage
• affected area
• probability of occurrence

This means planners could potentially ask:

What might a once-in-a-century rainstorm look like in this city?

The AI could then generate multiple plausible maps showing how that event might develop.

Not one prediction.

Multiple possible futures.

This allows engineers and governments to stress-test infrastructure against scenarios more severe than their historical records.

How Does the MIT AI Work?

The system combines two different types of information.

The first is point statistics.

These describe how often certain levels of intensity appear in the data.

For rainfall, for example, the model may calculate how frequently the highest rainfall measurement across a region reaches a particular amount.

The second is spatial data.

This tells the system how rainfall, flooding, heat, or another hazard spreads across a geographic area.

By learning the relationship between extreme statistical values and geographic patterns, the algorithm can construct new disaster scenarios beyond anything present in the training examples.

That is the breakthrough.

The model does not need to have previously seen the exact disaster it is generating.

The Researchers Tested It on Rainfall Across the United States

Chang and Sapsis tested the system using precipitation data covering the continental United States.

They began with around 25 years of hourly rainfall observations, which they converted into daily weather maps.

From this larger dataset, the researchers calculated statistical information showing how frequently different rainfall intensities occurred.

But the AI's spatial training used something much more limited.

The model was trained using paired low-resolution and high-resolution weather maps taken from only the first six months of the 25-year dataset.

That period contained few, if any, examples of the most extreme rainfall events.

Despite this limitation, the AI learned how broad weather patterns corresponded to finer geographic detail.

The researchers then combined that knowledge with statistical information from the full dataset.

This allowed the system to generate extreme scenarios beyond what it had directly observed during training.

Imagine New York Getting 300 Millimeters of Rain

The researchers give a striking example.

If the highest recorded rainfall event in New York City is around 200 millimeters, the model can generate a statistically plausible scenario involving 300 millimeters of rainfall.

There may be no historical event matching that exact amount.

But the AI can still estimate what such a storm could look like spatially.

Where would the heaviest rain fall?

How large would the affected area be?

Would the event be concentrated or widespread?

How long might it last?

These questions matter enormously for infrastructure planning.

Stress-Test the City Before the Disaster Happens

This technology could function like a virtual disaster laboratory.

Instead of waiting for a catastrophic storm to discover that a seawall is too low, planners could test that seawall against generated extreme scenarios.

Instead of discovering during a heatwave that the electrical grid cannot handle record demand, utilities could simulate far more severe conditions beforehand.

The same principle could eventually be applied to:

• floods
• wildfires
• storm surges
• heatwaves
• extreme rainfall
• power-grid stress
• water systems
• transportation networks
• supply chains

The question becomes:

Can our infrastructure survive the disaster that has not happened yet?

Why This Matters More in the Climate Change Era

Historical records are useful because they tell us what happened before.

But climate change complicates that assumption.

A hundred years of historical weather does not guarantee that the next hundred years will behave the same way.

Rainfall patterns are changing.

Heat records are being broken.

Storm behavior is evolving.

Sea levels are rising.

Infrastructure designed using old assumptions may be exposed to conditions outside the range engineers originally expected.

That means governments cannot rely only on the question:

"What was the worst disaster we experienced before?"

They also need to ask:

"What is the worst disaster that could realistically happen next?"

Artificial intelligence may help answer that.

There Are Still Limitations

The MIT research does not mean AI can suddenly forecast every disaster.

To apply the technique to another type of hazard, researchers still need appropriate statistical and geographic data.

For example, generating extreme wildfire scenarios would require reliable information about wildfire intensity, frequency, spread, and spatial behavior.

Flooding would require its own relevant datasets.

The system's output is also not a guarantee that a specific disaster will happen.

It generates statistically plausible risk scenarios, not exact future events.

That distinction is critical.

AI cannot tell a city:

"This exact storm will happen on this exact date."

What it may be able to say is:

"A storm of this magnitude is plausible, and this is what it could look like."

For disaster preparedness, that can already be extremely valuable.

Why the Philippines Should Pay Attention

For the Philippines, this kind of technology could become especially important.

Our country faces repeated exposure to:

• typhoons
• extreme rainfall
• flash floods
• landslides
• storm surges
• drought
• extreme heat
• coastal flooding

Many communities design infrastructure based partly on previous disasters and historical hazard maps.

But what if the next typhoon produces rainfall beyond anything previously recorded in Metro Manila?

What if storm surge levels exceed the assumptions used when coastal infrastructure was designed?

What if a heatwave lasts much longer than anything currently included in power-demand forecasts?

What if rainfall in a mountain watershed produces flooding beyond historical river measurements?

These are not purely academic questions.

They affect lives.

They affect power.

Food.

Transportation.

Communications.

Hospitals.

Schools.

Water.

Entire cities.

AI Could Become a Tool for Disaster Preparedness

Imagine the Philippine government being able to ask:

What would a once-in-200-years rainfall event look like in Metro Manila?

Or:

What happens if a typhoon produces rainfall 30 percent beyond the historical maximum in Bicol?

Or:

Which substations fail first during an unprecedented heatwave in Luzon?

AI-generated scenarios could help engineers identify weak points before disaster strikes.

This does not replace PAGASA, climate scientists, civil engineers, hydrologists, disaster-response experts, or local knowledge.

It gives them another tool.

And in disaster management, better preparation can mean the difference between disruption and catastrophe.

AI Is Moving From Prediction to Possibility

Much of artificial intelligence has focused on predicting what is most likely to happen.

This MIT research explores another important question.

What could happen, even if it has never happened before?

That is a powerful shift.

Because some of the most dangerous disasters are dangerous precisely because we are not prepared for them.

Professor Sapsis points out that many modern systems have been optimized for efficiency.

Supply chains are lean.

Power systems operate close to demand.

Food distribution relies on tightly connected networks.

Infrastructure often has limited redundancy.

One extreme event can create failures that spread far beyond the location where the disaster began.

A storm can become an energy crisis.

An energy crisis can become a supply-chain crisis.

A supply-chain crisis can become a food and economic crisis.

That is why understanding low-probability but high-impact events is becoming increasingly important.

The Bigger Story

The most valuable AI may not always be the system that tells us what will happen tomorrow.

Sometimes it may be the system that shows us what we are not prepared for.

This MIT technology is still research.

But the idea behind it could become essential in a world facing more complex climate risks.

Instead of training AI only to recognize history, researchers are teaching it to explore the boundaries of what is possible.

And that could help societies prepare before the unthinkable becomes reality.