Fire Season Is Now the Algorithm Season

The Camp Fire burned 153,336 acres in Butte County, California over seventeen days in November 2018. It killed 85 people, destroyed the town of Paradise (pop. 27,000), and caused an estimated $16 billion in damages. It was the deadliest and most destructive wildfire in California history.

The Camp Fire’s behavior on November 8, when it grew from ignition to a 15,000-acre inferno in six hours, was unprecedented in recorded California fire history. Wind speeds exceeded 50 miles per hour with gusts above 70 mph. Relative humidity dropped below 10%. The fuel—forests that had been in drought stress for years, with heavy deadwood accumulation from decades of aggressive fire suppression—was at historic low moisture content.

Weather forecasters knew the conditions were extreme before the fire started. They issued Red Flag warnings. The warnings were not sufficient. Eighty-five people did not have enough time to leave.

The problem AI wildfire prediction is trying to solve is not detection—satellites and cameras have been detecting fires within minutes of ignition for years. The problem is behavior prediction: given that a fire has started here, with these fuel conditions, this wind field, this terrain, how will it spread over the next six, twelve, and twenty-four hours? And how certain are you about that prediction?

The physics of wildfire spread is genuinely complex. Fire moves faster uphill than downhill (updrafts on hillsides accelerate convection). It jumps roads and rivers under high winds through spotting—embers carried far ahead of the fire front. It accelerates dramatically in drainages that channel wind. It can shift direction suddenly with wind shifts, cutting off evacuation routes that were safe an hour earlier. The standard models in use before roughly 2020—FARSITE, Phoenix, FlamMap—are physics-based models that handle average conditions reasonably well and perform poorly during the extreme events when prediction matters most.

The AI approaches address this in two ways. First, training neural networks on historical fire spread data—satellite-detected fire perimeter growth paired with the weather and terrain conditions at each time step—to learn the empirical patterns of how fires actually move in conditions that the physics models handle poorly. This includes the tails of the distribution: the rare extreme events where spotting distance is ten times the model default, where spread rates are five times the physics prediction.

Second, ensemble forecasting: running dozens or hundreds of model variants simultaneously, with different wind scenario assumptions, different fuel moisture estimates, different ignition locations, and aggregating the results into probabilistic spread forecasts. A forecast that says “70% probability the fire reaches this road intersection within 3 hours, 30% probability it reaches the town center within 6 hours” is operationally more useful than a single-line forecast that may or may not account for the scenario that unfolds.

The National Interagency Fire Center integrated more sophisticated ensemble forecasting into its operational tools in 2022. The California Department of Forestry and Fire Protection (CAL FIRE) has been piloting AI-assisted spread modeling through a partnership with the Pacific Northwest National Laboratory. The early operational experience suggests meaningful improvements in the 6-12 hour spread accuracy during moderate fire events; the extreme event prediction remains harder.

The infrastructure to put this prediction capability in the hands of people who need to make evacuation decisions is the other part of the problem. The Camp Fire evacuation failed partly because the predictions, even had they been available, had no pathway to translate into mandatory evacuation orders that reached everyone in the fire’s path in time.

Butte County had a population warning system. It was not comprehensive. It relied on phone calls and text alerts to registered numbers, missing renters, recent arrivals, people with non-local numbers, and people who didn’t respond to unlisted calls. The town of Paradise had one primary evacuation route for 27,000 people.

The combination of better fire spread prediction and better emergency alerting infrastructure—the nationwide Wireless Emergency Alert system can reach all cell phones in a geographic area simultaneously, without requiring prior registration—represents a genuine improvement in the ability to warn populations. The 2023 Maui Lahaina fire, which killed 102 people, demonstrated a different kind of failure: sirens not activated, emergency alerts not sent to phones, and officials who hesitated to issue mandatory evacuation orders. The technology for alerting was available. The decision to use it was not made in time.

AI evacuation optimization—given a spreading fire perimeter and a road network with real-time traffic data, compute the optimal evacuation routing for a given population—is an active research area. The algorithms are tractable. The implementation requires integration with emergency management systems that move slowly, and the operational trust required to let an algorithm influence evacuation routing decisions is high.

The structural conditions that are generating larger, more intense fires are not addressable through prediction and response systems, however good. The combination of a century of fire suppression (which has accumulated dense fuels in forests that historically burned frequently), prolonged drought stress from climate change, increasing temperatures that reduce fuel moisture, and expanding human development in the wildland-urban interface—these are the factors that make modern megafires possible.

Forest management—prescribed burning, mechanical fuel reduction, forest thinning—is the intervention that addresses the root fuel accumulation problem. The US Forest Service managed approximately 3 million acres with prescribed fire in 2022, the highest level in modern records. The historical fire-adapted ecosystems of the American West burned 4-12 million acres per year before European settlement. The gap between what is being treated and what needs treatment to restore something like historical fire behavior is enormous.

AI is being applied to forest management planning through optimization of prescribed burn programs: given available resources, fuel accumulation maps, weather forecasting, air quality regulations, and infrastructure proximity, what areas should be treated in what sequence to most efficiently reduce fire risk? This is a scheduling and spatial optimization problem that ML tools handle well in principle. The execution is constrained by air quality regulations that restrict when prescribed burns can occur, by liability concerns around escaped burns, and by a Forest Service workforce that is underfunded relative to the treatment backlog.

The wildfire AI investment has followed the money—significant private investment from insurance companies facing catastrophic losses, startups building detection and spread modeling tools for premium applications, and government contracts for operational forecasting support. The forest health investment—fuel treatment, controlled burning at ecological scale, addressing the conditions that make the fires—has followed the public budget, which is a different and less generous number.

Reforestation AI and forest carbon credit monitoring are also active areas, with companies like Pachama, NCX, and SilviaTerra applying satellite analytics and ML to assess forest carbon storage and track restoration outcomes. These tools serve the carbon markets and have some alignment with fire risk management, but they are optimized for carbon accounting rather than fire hazard reduction.

The Camp Fire’s 85 victims died partly because the fire behaved in ways that were outside the historical envelope. Climate change is making those out-of-envelope events more frequent—the conditions that produced the Camp Fire will occur more often as temperatures rise and droughts deepen.

Better prediction models extend the warning time before those events become fatal. That is genuinely valuable. A family in Paradise with an extra hour of warning is a family that might make it out.

But better prediction doesn’t change the forest that burns. It doesn’t address the housing developments in high-risk zones that communities have permitted for decades. It doesn’t rebuild the water infrastructure that Paradise’s survivors discovered they couldn’t reassemble quickly. It doesn’t change the insurance withdrawal from California that is leaving millions of homeowners without coverage and removing the financial risk signal that might deter further development in fire-prone areas.

Fire season is the algorithm season because the people building AI tools are doing what they can do with the capabilities they have. The problems the algorithms can’t solve require different tools and different political decisions. The eighty-five people who died in Butte County in November 2018 would have needed both.