Can We Predict the Next Disaster? The Science of Wildfires, Droughts, Tornadoes and Tsunamis

From the fire-risk maps of Los Angeles to experimental tornado forecasts: what science can predict today, what AI is improving, and where nature still leaves us guessing.

Next Horizon cover showing a wildfire, massive storm clouds, drought-stricken ground, and a distant coastline, illustrating the science of natural disaster prediction.

Can science predict disasters before they strike? From wildfires and droughts to tornadoes and tsunamis, new technologies are changing how we anticipate natural hazards.

In January 2025, destructive fires swept through neighborhoods of Los Angeles. The winds were fierce, the landscape dangerously dry, and thousands of people had to flee. In a study published that April, researchers showed that a machine-learning system designed to forecast fire activity—not just dangerous fire weather—had identified elevated risk in the region five days ahead. It did not predict the first spark or the house that would burn. But it had seen the conditions for trouble coming.

That distinction matters. A forecast can tell firefighters when to prepare, even if it cannot tell a family which home will be lost. In a very different event in Missouri that same year, forecasters used an experimental storm-prediction system to give local emergency responders more than two hours of advance notice of a tornado threat. That was not a routine public two-hour tornado warning; it was an example of what extra information might eventually make possible.

These stories raise a bigger question. Could we someday know about wildfires, droughts, tornadoes and tsunamis before they happen? The answer is stranger—and more useful—than a simple yes or no. Radar can see circulation hidden inside a thunderstorm. Satellites can watch plants lose moisture or measure a tsunami moving across the ocean. Supercomputers simulate countless possible futures. And AI can sometimes find patterns faster than older methods.

But every disaster keeps a different kind of secret. Sometimes we can forecast the dangerous conditions, sometimes the event itself, and sometimes only what happens after it begins.

First, what do we actually mean by “predict”?

There are four very different jobs hidden inside that word. Scientists can estimate where a disaster is more likely over a season; forecast hazardous conditions days in advance; predict how a fire, storm or wave will behave once it starts; or issue an urgent warning when a threat is imminent. A system can be excellent at one job and poor at another.

Think of a medical checkup. A doctor can estimate someone’s risk of heart disease without knowing the exact day of a heart attack. Fire agencies can recognize a dangerously dry week without knowing whether lightning will strike a particular tree. And meteorologists can warn that a region is primed for violent thunderstorms long before they know which storm, if any, will produce a tornado.

Tsunamis make this distinction particularly clear. They are not weather events: most begin with undersea earthquakes, while others result from landslides or volcanic activity. We cannot reliably predict the precise time of an earthquake. We can, however, detect a dangerous earthquake, measure the waves it generates and estimate when the tsunami could reach distant shores.

Put simply, the important question is not always “Will it happen?” It may be “How much time will we have once the risk becomes clear?”

How far ahead can we see? These are typical scales of useful guidance, not guaranteed warning times.

Wildfires — weather and landscape danger: days to seasons; the exact ignition often cannot be forecast.

Droughts — regional risk: weeks to months; local outcomes remain uncertain.

Tornadoes — severe-storm risk: days; official tornado warnings are often measured in minutes, with experimental guidance sometimes earlier.

Tsunamis — the triggering earthquake cannot be reliably timed; once it begins, arrival forecasts range from minutes to hours depending on distance.

How forecasts turn observations into warnings

A modern forecast begins with the imperfect picture we have of the world right now. Weather stations measure wind and temperature; balloons sample the air overhead; ships and buoys watch conditions at sea. Doppler radar detects precipitation and movement inside storms. Satellites fill in much of the space between instruments, especially over oceans and remote landscapes.

Computers combine these scattered observations into a single, best-available starting picture, a process scientists call data assimilation. Think of it as reconstructing a moving scene from thousands of photographs taken at different angles. A physics-based model then calculates what could happen next as air rises, moisture condenses and pressure systems move.

The problem is that the atmosphere is chaotic. A small error in the starting picture can grow as the forecast reaches farther ahead. That is the real meaning of the butterfly effect: not that forecasts are pointless, but that some details—especially small, fast-moving ones—become increasingly hard to pin down.

One answer is an ensemble forecast. Instead of asking a model for one future, forecasters run it many times with slightly different starting conditions. Imagine planning a picnic and seeing 40 of 50 forecasts predict heavy rain. You would probably change your plans. The spread between those outcomes matters as much as their average: it is a window into uncertainty, not an admission of failure.

A major change is happening inside this process. In February 2025, the European Centre for Medium-Range Weather Forecasts (ECMWF) put an AI forecasting system into operational use alongside its physics-based models. In July 2025 it added a 51-member AI ensemble, and in May 2026 both systems were upgraded. AI does not make the atmosphere non-chaotic, but it can calculate forecasts quickly and identify patterns learned from enormous historical datasets.

And yet no amount of speed gives a model knowledge that the observations do not contain. The crucial advance is better guidance about the range of possible outcomes, not a perfect countdown to disaster.

Wildfires: the forest can be ready to burn long before it ignites

A wildfire needs a source of ignition, something that can burn and conditions that allow flames to spread. Heat, low humidity and strong winds make an obvious combination. Yet the same hot day may pass quietly in one forest and turn disastrous in another. The difference can be the moisture inside plants, the amount of dead vegetation, the slope of the ground, the direction of the wind—and one unpredictable ignition.

This is why a fire-risk map is not a calendar of future ignitions. In the United States, the Storm Prediction Center issues fire-weather outlooks for the coming days, while seasonal outlooks help agencies distribute people and equipment. Such forecasts can identify a dangerous combination of wind, heat and dry vegetation; they cannot say when or where somebody will drop a match.

Researchers are making the forecasts more intelligent by studying the landscape, not only the sky. A 2025 study in Geophysical Research Letters examined the daily spread of 196 wildfires in the northern Rocky Mountains. Soil moisture and plant water stress proved particularly important. Adding these measurements improved estimates of daily fire growth compared with models using weather and fuels alone. It is a regional result, not a universal formula, but it suggests why yesterday's drought can matter as much as today's wind.

The next step is to forecast fire activity, not just the weather that might feed a fire. In a 2025 Nature Communications study, researchers combined information about vegetation, weather and likely ignition drivers. Their model had identified elevated fire activity risk around Los Angeles five days before the January 2025 fires. That is a result about probabilities and regions—not a prediction of the precise ignition point. Importantly, the study evaluated a system already producing daily forecasts, rather than claiming scientists had invented a time machine after the event.

Forest wildfire spreading across a dry mountain slope, with a satellite and illustrative fire-spread prediction overlays showing how scientists monitor wildfire behavior.

Satellites, weather models, and artificial intelligence help scientists identify dangerous fire conditions, detect emerging fires, and estimate how flames may spread.

Once a fire begins, speed matters. Existing satellite systems can flag heat anomalies, while aircraft and ground crews provide detail that an orbiting sensor may miss. A newer project, FireSat, launched its first three operational satellites in July 2026 following a prototype mission. Its ambition is more frequent detection of smaller fires. But the first three spacecraft do not yet amount to a fully deployed, continuously watching global fire alarm.

Fire-spread models tackle a different problem: where flames might travel over the next hours. They combine the mapped fire edge, terrain, vegetation and expected winds. For emergency teams, that can help prioritize roads, crews and evacuation zones. Yet a gusty shift in the wind—or an unexpected new ignition—can make yesterday’s confident-looking map obsolete.

Wildfire science is therefore building a sequence of decisions rather than one miraculous prediction: assess seasonal danger, prepare for a severe fire-weather day, detect new flames, update the likely spread, and give people clear instructions while routes remain open.

Drought: the slow disaster that can arrive surprisingly fast

A drought is less spectacular than a tornado, but it can affect far more people over a much longer period. It also has more than one face. A farmer may have dry soil around the roots while a city reservoir is still full; a reservoir may decline months after a failed rainy season. Heat can make the atmosphere draw moisture from fields faster even without a dramatic drop in rainfall.

Scientists combine rainfall records, temperature, snowpack, rivers, groundwater, plant condition and soil moisture. Satellites allow them to survey large areas where ground sensors are scarce. NOAA’s Evaporative Demand Drought Index tracks how strongly the atmosphere is demanding water from the surface—roughly speaking, how “thirsty” the air has become. That extra pressure can be an early sign of agricultural stress.

The most unsettling case is flash drought. Instead of developing over a season or longer, dry conditions can intensify in a matter of weeks. A 2025 review of 122 scientific studies found that researchers have become better at detecting these events, but predicting their onset remains difficult. A field that looked healthy at the beginning of a month can be struggling by the end.

Agricultural landscape transitioning from healthy green crops to drought-damaged fields and cracked soil, with illustrative moisture contours showing changing water availability.

By combining satellite observations, soil moisture measurements, and AI forecasting models, researchers are improving their ability to anticipate drought conditions weeks in advance.

A 2025 Nature Communications study offered a promising step forward. Researchers combined deep learning with conventional forecast models and previous observations of root-zone soil moisture—the water plants can actually reach with their roots. The hybrid approach produced useful forecasts up to four weeks ahead and improved flash-drought prediction in several evaluated cases. It is not a guaranteed four-week warning for every farm: accuracy differed across conditions and regions, and the results were assessed on historical data.

In October 2026, a Nature study tackled the especially difficult two-to-six-week window from another angle. Its machine-learning system learned to correct the probabilities produced by existing forecasts, improving the estimates of several weather extremes. The model performed strongly in a 2025 real-time forecasting competition. The advance is less glamorous than the promise of an AI that can “see a month ahead,” but it is scientifically more meaningful: a better estimate of uncertainty can help water managers and farmers act sooner.

Drought predictions matter when they change choices: managing reservoirs, scheduling irrigation, planning crops or protecting power supplies. The connection between rainfall, water and agriculture is central to our recent exploration of the future of food. Forecasting cannot manufacture water, but it can help people use the water they have before a shortage becomes a crisis.

Tornadoes: knowing a storm is dangerous is not the same as knowing it will touch down

A tornado looks like a single spinning column of air, but its formation depends on a complicated chain of events. Many of the strongest tornadoes form beneath supercells: thunderstorms with long-lived rotating updrafts. Meteorologists know the broad ingredients—warm, moist and unstable air, together with winds that change speed or direction with height. The stubborn question is why some storms produce tornadoes and others, apparently similar, do not.

Days in advance, meteorologists can forecast a region with elevated severe-storm risk. That is the role of broad outlooks and watches. Once thunderstorms form, Doppler radar reveals areas of rotation; debris seen by specialized radar can offer strong evidence that a tornado is already occurring. A tornado warning is more urgent and more localized: it means the threat is imminent or happening, not merely that conditions are favorable.

This gap between a watch and a warning explains why knowing that Thursday may be dangerous does not tell us which street will face a tornado at 4:37 p.m. The finer the detail we demand, the more the forecast depends on rapidly changing processes inside a single storm.

NOAA says the average US tornado warning gives about 13 minutes of lead time, although actual lead times vary widely. That is a figure about official warnings—not the maximum number of hours a forecaster may anticipate severe weather. Warning too early without sufficient evidence risks repeated false alarms; warning too late can leave little time to reach shelter.

Powerful tornado developing beneath a rotating supercell thunderstorm over farmland, with an illustrative Doppler radar display highlighting storm rotation.

Doppler radar helps meteorologists identify rotating storms, while experimental high-resolution forecasting systems aim to provide more time for communities to prepare for tornadoes.

A research program called Warn-on-Forecast is trying to narrow that gap. At NOAA’s National Severe Storms Laboratory, rapidly updated, high-resolution computer simulations estimate how individual thunderstorms might develop over the next few hours. They can show an increasing risk of dangerous rotation; they cannot yet identify every future tornado reliably or replace the judgment behind an official warning.

An example from Missouri shows the potential. In March 2025, forecasters used Warn-on-Forecast guidance to alert local emergency responders more than two hours before an EF3 tornado. NOAA described the case in January 2026. It was a research demonstration and an early briefing to responders, not a sign that every US household now receives a two-hour public tornado warning.

That is where progress may matter most: moving from a broad risk on the morning weather map to more specific advice for the communities that will need it, without asking people to act on every uncertain signal.

Tsunamis: when we cannot predict the trigger but can track the wave

A tsunami is not an unusually large wind-driven wave. Most are generated when an undersea earthquake suddenly moves the seafloor and displaces water above it. Underwater landslides and volcanic activity can also start one. In deep water, the disturbance may be barely noticeable to a ship, while still traveling across an ocean at the speed of a passenger aircraft.

The obstacle comes first: earthquakes themselves cannot currently be predicted with reliable dates, locations and magnitudes. The US Geological Survey is explicit about that limit. Long-term probabilities can show where large earthquakes are more likely over decades, but they cannot tell a coastal town exactly when a particular fault will rupture.

Once the earth moves, however, prediction becomes a different problem. Seismic networks rapidly locate the quake and estimate its magnitude. NOAA’s warning centers often issue initial messages for US and Canadian coasts within roughly five minutes, based mainly on seismic information. Deep-ocean DART pressure sensors and coastal sea-level gauges then help confirm whether a tsunami has formed and how the water is moving.

Numerical models combine the best available earthquake-source estimate with seafloor depth and coastline shape to estimate travel times, offshore wave heights and potential flooding. For a coast across the Pacific, that may provide hours to prepare. For a town close to the earthquake, the first dangerous wave may arrive before an official message can reach everyone.

Scientific illustration of an undersea fault rupture, seafloor uplift, tsunami propagation, and a DART ocean monitoring system near a vulnerable coastline.

Tsunami warning systems combine earthquake measurements, deep-ocean pressure sensors, and computer simulations to estimate how dangerous waves may travel toward coastlines.

The July 2025 Kamchatka earthquake offered an extraordinary research opportunity. About 70 minutes after the magnitude 8.8 quake, the NASA–CNES SWOT satellite recorded an unusually wide view of the tsunami’s sea-surface patterns. A study published in March 2026 used those observations to infer more about where the seafloor had shifted. The work may improve future tsunami models; it does not make satellites a reliable instant alarm for every earthquake or nearby coast.

One practical distinction should never be lost in the technology: if you are at the coast and feel strong or prolonged shaking, see the sea suddenly pull back or rise unusually, or receive an official tsunami warning, move to higher ground immediately. A local tsunami can arrive before a phone alert. Do not wait to watch the water.

What AI changes—and what it still cannot know

AI has become the headline in almost every forecasting breakthrough, but it is better understood as another kind of model than as an oracle. Traditional weather systems follow equations describing the movement of air and water. Many AI systems learn from enormous collections of past observations and model-generated weather, discovering relationships that help them forecast new conditions quickly. Different tools excel at different jobs: some predict broad weather patterns, while others recognize fire hot spots or interpret radar signals.

There is now evidence beyond controlled laboratory examples. ECMWF runs AI forecasting systems operationally alongside physics-based models. A Nature paper published in August 2026 described an AI system for tropical cyclone tracks, strength and wind fields that, in the researchers’ 2023–2025 evaluation, gained roughly a day or more of useful lead time over leading operational baselines. Tropical cyclones are not the same as tornadoes, but they provide a powerful test of how much additional warning better forecasting can offer.

The most promising future may involve many specialized systems working together. Satellites catch a suspicious hot spot. Ground sensors confirm the conditions. A model estimates where danger might spread. An alert reaches the people who can act. At every step, a skilled human still has to decide what the information means and when to warn others.

AI also inherits the weaknesses of the world it learns from. The most extreme events are rare, and high-quality historical observations are concentrated in some countries more than others. Models can falter when conditions fall outside their training experience. And the fact that a prediction looks convincing on a screen does not mean it has been tested well enough to direct an evacuation.

A 2025 review in Nature Communications emphasized careful evaluation of AI methods for extreme events and their responsible use in decision-making. This is a useful distinction: performing well on archived storms is a scientific result; earning trust in a live emergency requires reliable performance, independent testing and clear communication about uncertainty.

The last mile of prediction is human

Suppose a forecast gives your town a 30 percent chance of damaging winds tomorrow. Should schools close? Should an outdoor festival be canceled? What about a hospital moving vulnerable patients? The percentage alone cannot answer. Someone has to weigh the potential harm against the cost of taking precautions—and decide which actions are still possible.

There are two ways to lose trust. Repeated false alarms make warnings easier to ignore; missed threats can be disastrous. Precision can be misleading, too. A computer map that colors one street red and the next green may look authoritative even if the underlying data cannot support decisions at that scale.

Access is another problem entirely. In November 2025, the World Meteorological Organization reported that 119 countries—about 60 percent of the global total—had reported a multi-hazard early warning system. That is progress, but the existence of a national system does not mean every resident receives a timely, comprehensible alert or has a safe place to go.

Even a nearly perfect forecast is useless to someone with no mobile coverage, no reliable electricity, no accessible evacuation route or no safe place to go. As we discussed in our look at the future of the internet, digital services depend on very physical infrastructure. During a disaster, that dependence becomes impossible to ignore.

For someone facing a real emergency, the best warning answers three questions in ordinary language: What might happen here? How certain are the forecasters? And what should I do now?

Will we ever predict every disaster?

Almost certainly not in the sense of knowing exact times and locations for every event. Too many pieces of the system are unresolved: the ignition of a fire, the behavior of a rapidly evolving supercell, the future rainfall that governs drought, or the moment a fault under the sea will rupture. Some uncertainties may shrink with new data. Others may remain fundamental.

That does not make the work futile. The realistic future is a connected warning chain: better monitoring of vulnerable forests, earlier signs of water stress, higher-resolution storm guidance and tsunami calculations that combine more measurements in real time. The tools will be judged by what people can do with their extra minutes, hours or weeks.

Climate change raises the stakes, but it needs careful language. A warmer climate can intensify some drought and fire-weather risks, and higher sea levels can increase coastal inundation during a tsunami. That does not mean every fire, tornado or tsunami can be blamed on climate change. A climate projection describes how risks may shift over decades; a disaster forecast concerns an event at a particular place and time.

Picture a town using a forecast to stage fire crews before a windstorm. Or farmers changing irrigation plans while enough water remains. Or a hospital checking its evacuation route well before the roads fill. None of those decisions requires perfect certainty. It requires an honest estimate of danger that arrives soon enough to be useful.

The future of forecasting is measured in time to act

Science is steadily extending that time. We can monitor drying soil from space, recognize storms with dangerous rotation, detect new fires and follow tsunami waves once they exist. AI may make these signals quicker and easier to interpret, but the warning still has to survive the journey from a research model to an actual person.

A forecast that never changes a decision is only information. A forecast that helps someone reach higher ground, shelter from a tornado or move crews before a fire spreads can change the outcome—even if it never predicts the disaster perfectly.

The most valuable breakthrough may be neither a sharper satellite image nor a more powerful supercomputer. It may be the hour, the day or the week they give us back.