Using machine learning and math, a Brigham Young University student improved a key tool firefighters rely on during wildfire season

As Utah enters the heart of wildfire season, wind might be the most unpredictable — and most dangerous — factor in the spread of flames. Wind gusts can send fires surging over hillsides, push smoke into neighborhoods and complicate firefighting from the ground and the air.

For Jane Housley, a Brigham Young University mathematics graduate student, wind was also the key to improving wildfire prediction models. Her research, completed this spring and published as her master’s thesis, could help make a widely used wildfire tool faster and more accurate when needed most.

Housley partnered with the U.S. Forest Service’s Missoula Fire Sciences Lab and focused on improving WindNinja, a simulation tool created by the agency and used by fire crews and analysts to predict how wind will move through terrain during a fire. It’s a helpful software, but not perfect.

“WindNinja struggles to model what’s called a cavity zone,” Housley said. “That’s the area directly behind a mountain or ridge where wind tends to swirl backward and create eddies.” These eddies are important because they can dramatically shift how and where a fire spreads.

WindNinja offers two simulation modes: a mass-conserving solver that’s fast but less accurate, and a computational fluid dynamics (CFD) solver that’s more precise but much slower, typically taking 14 seconds per simulation compared to one to two seconds for the faster solver.

Housley was determined to bridge the gap between speed and accuracy. She came at the problem from two directions.

First, she used a mathematical approach inspired by airflow around buildings. Decades-old architectural research showed how wind wraps around square structures in cities, and Housley repurposed that math to approximate how wind might similarly move across complex natural terrain.

She wrote an algorithm to approximate hills and peaks as “rectangular buildings,” then applied formulas to predict where cavity zones should appear. The result? A lightweight model that clearly outlines trouble spots where wind flow is likely to become turbulent.

Although the method doesn’t yet predict the exact wind speed inside those zones, it efficiently pinpoints where WindNinja’s simpler solver is likely to miss key dynamics.

The second part of her project took a more modern twist: machine learning.

Utilizing the computer science skills gained through her coursework at BYU, Housley built a custom U-net convolutional neural network (a type of artificial intelligence often used in image recognition) and trained it on nearly 6,000 wind simulation images provided by the Missoula Fire Sciences Lab. Each data pair included terrain, vegetation type, wind direction and outputs from both WindNinja solvers.

Aiming to couple the accuracy of the WindNinja CFD solver with the speed of the WindNinja mass-conserving solver, she trained the neural network to learn patterns of error in the mass-conserving solver, using the CFD solver accuracy as the goal. Without having to solve complicated physical equations like traditional methods such as the CFD solver do, the neural network was able to produce a pipeline 7x faster than industry-leading models while retaining high accuracy.

The results were impressive. Think of it like teaching a computer to recognize wind patterns the way facial recognition spots a familiar face — quick and accurate. The results included the following:

• The model cut one type of error by 75%.
• It sliced the average error in half.
• On a test that measures how close two wind maps look — kind of like comparing photos — it scored 0.77 out of 1, a big jump from 0.60.
• Best of all, it did it in just 0.07 seconds per simulation.

Housley said she still remembers the feeling of excitement when she saw how accurate and efficient her new model was. It was a sweet payoff after months of grueling work.

“Once I had the network built and plugged in the data and ran the simulation, the results were really good. I thought, ‘I must be doing something wrong,'” Housley said. “I combed through every single line of code and found that it was working correctly. I was really excited.”

Housley said collaborating closely with scientists at the Missoula Fire Sciences Lab was an inspiring part of her experience at BYU and she appreciates the experiential learning opportunities she had as both an undergraduate and graduate student.

“I didn’t appreciate the research opportunities I had been given at BYU until I talked to some friends at other universities during my internship. Most of them didn’t have the chance to do any research. It made me realize how unique BYU really is,” she said.

Combining her passion for math and her love for the gospel of Jesus Christ at BYU is something she said she’ll take with her as she prepares to move to Wisconsin to start a new job.

“The spiritual side of BYU has been amazing,” she said. “I had so many math professors make gospel connections during class. Math, to me, is a beautiful art form, and it’s a source of truth. Within the church, we’re always searching for sources of truth.”

Housley’s work, while deeply technical, carries clear and immediate value for wildfire modeling, especially as fire seasons grow longer, drier and more severe across the western U.S.

In the fast-moving world of wildfire response, that time savings could mean quicker briefings, more efficient evacuations and better on-the-ground decisions.