A promising future for fire weather forecasting using AI

By Theresa Barosh | May 2025

Three smoky fires in Minnesota

Is it more important to know what is most likely to happen or the worst-case scenario?

CIRA researcher Ryan Lagerquist is working on making weather predictions that provide multiple potential outcomes and their likelihoods – or uncertainty – using machine learning. A typical weather forecast, in contrast, only provides one answer, such as that tomorrow’s high temperature will be 84 degrees.

Two photos of Lagerquist in a race
As a distance runner, Lagerquist takes advantage of the Colorado trails and races. 2025

“With traditional methods that don’t involve AI, it’s often very difficult to quantify the uncertainty in your predictions,” said Lagerquist. “There’s methods that have been really embraced by the meteorology community within the last few years that allow you to use AI to get really useful and informative uncertainty quantification.”

For Lagerquist’s fire weather research, for example, instead of just giving one prediction for every fire-weather index, he creates an ensemble of predictions. That provides a range of likely scenarios. Lagerquist hopes that this will allow the eventual algorithm users – like forecasters and firefighters – to look at a range of likely scenarios and figure out the best-case and worst-case scenarios around if and where a fire might ignite. Users can evaluate the information based on their specific threshold for risk. Some people, or entities, have really low risk thresholds because they need to prepare for the worst-case scenarios.

The Fire Weather Puzzle

The puzzle of predicting wildfires or forest fires can be divided into three parts: fire weather, fire ignition, and fire behavior after it starts. Lagerquist focuses on fire weather, the conditions that happen before ignition and can impact how likely a fire is to start. He works to define how conducive the atmosphere and available fuels are to wildfire ignition, growth and spread.

Lagerquist focuses specifically on predicting seven fire-weather indices: fine-fuel moisture code, duff moisture code, drought code, initial-spread index, build-up index, daily severity rating, and total fire-weather index. He provides predictions at lead times of 1-14 days, across the full United States, including Alaska and Hawaii.

Lagerquist and his team use a deep neural network to post-process weather forecasts from the Global Forecast System (GFS) model into skillful forecasts of fire-weather indices. The GFS model is a physics-based weather prediction tool, run multiple times daily by the National Weather Service and used widely by the weather forecast and research communities. Physics-based means that the model runs based on known rules about how the natural world and the atmosphere work. “I’m using statistics to make those predictions better.”

Lagerquist’s deep neural network outperformed the GFS on almost every one of its 98 tasks – providing seven fire-weather indices, each at 14 lead times – while also providing skillful uncertainty quantification via an ensemble. Each of the 98 outcomes has likelihoods to provide users with a broad understanding of what could happen.

“My AI algorithm is providing value because it’s making predictions more accurate,” said Lagerquist, “So that’s something that can allow people to plan for big fire weather days, where a lot of bad things might happen.”

North american maps with best case scenario, median scenario, and worst case scenarios
Two-day-ahead forecasts (made with information available on Dec 28, 2021) of daily severity rating (DSR), valid for Dec 30, 2021, the day of the Marshall Fire in Colorado.  In both the median and worst-case scenarios, the most extreme category of DSR (red) covers a large swath of the southern Rocky Mountains, including the location of the Marshall Fire.

Fire-Weather Indices Explained

Three of the fire weather indices describe how much moisture is in the fuels that could possibly ignite and burn. Those are called the fine-fuel moisture code, the duff moisture code, and the drought code.

There’s an index called the initial-spread index that basically tells you, if a fire gets ignited, how quickly it can spread in its early stages.

The build-up index describes the total amount of fuel that could be combusted.

The daily severity rating and the total fire-weather index are summaries of all the other indices, so they blend all the fire-weather indices into just a couple big ones that describe the overall danger on a given day at a given location.

Those seven indices had been computed with physics-based models like the GFS prior to Lagerquist’s research. He can now create those indices using AI with his team’s weather forecast model that is currently running and provides those indices.

What is needed: continued CIRA funding in fiscal year 2026 to support researchers like Lagerquist in conducting behind-the-scenes work that helps plan for worst-case scenarios, saving lives and protecting infrastructure.

Graph of root mean squared error over lead time in days
Root mean squared error (a measure of forecast accuracy, where lower is better) for fast-changing fire-weather indices — the fine-fuel moisture code (FFMC), initial spread index (ISI), and daily severity rating (DSR) — for both the physics-based GFS model and Lagerquist’s AI-based neural network (NN).  For the 42 tasks (3 fire-weather indices at 14 forecast lead times) in this graphic, the AI-based neural network is more accurate for 41 of the 42 tasks.
Graph of root mean squared error over lead time in days with up to 90 RMSE
Root mean squared error (a measure of forecast accuracy, where lower is better) for slow-changing fire-weather indices — the duff moisture code (DMC), drought code (DC), and buildup index (BUI) — for both the physics-based GFS model and Lagerquist’s AI-based neural network (NN).  For the 42 tasks (3 fire-weather indices at 14 forecast lead times) in this graphic, the AI-based neural network is more accurate for 31 of the 42 tasks.