Neal McLoughlin, superintendent of predictive services with the B.C. Wildfire Service (BCWS), is trying to get one of B.C.'s most challenging fires, the Bald Range wildfire, to behave in his modelling software the way it did on the ground in the Okanagan Valley on Aug. 7.

On his computer, he's turning up the dials for extreme fire behaviour, replacing the forest fuels with more flammable types, programming a stronger wind from the west and swapping local topography of a downward sloping hill with a flat landscape.

Only with those interventions do his simulations get close to what actually happened on the ground that night in Summerland when more than 12,000 people were forced to flee as flames rushed toward the town in mere hours. One woman died in the chaos.

"Observed fire behaviour is doubling what our models would otherwise predict," said McLoughlin.

It's partly because the data that informs these computer models was collected at least 30 years ago, when extreme fire behaviour, driven by events like prolonged drought — a feature of climate change — was less common.

The challenge is especially pronounced in B.C.'s Southern Interior region, officially the driest place in Canada, according to the Canadian Drought Monitor in a wildfire season that saw less area burned than average, yet was one of the most devastating seasons on record for property destruction.

BCWS specialists look to their model predictions the way most of us consider a weather forecast. In their case, they're co-ordinating a response to a wildfire, thinking about how quickly it might spread and how their tactics and safety plan should change around that. It's part of the larger toolbox, on top of human resources, BCWS uses to evaluate and dedicate resources to wildfires.

While it's a conceptually robust prediction system, based on more than 100 years of data collection and analysis, even its website states it "was last updated in the early 1990s." Trusting the models In practice, McLoughlin says, his team have learned to become more skeptical of the model's predictions on those exceptionally hot and dry days, when even decades of high-grade data cannot overcome their own instincts that a wildfire could be significantly worse than the models say.

Ultimately it's people that make decisions about how to respond to a wildfire, not models, and on those fire-prone days McLoughlin has begun training his team to really apply their lived experience of past extreme fires. On some days, he says: "I don’t trust the models on this day. I expect it's going to do something very different." "Black swan events" is what Chelene Hanes, a research scientist with the Canadian Forest Service (CFS), calls those unpredictable fires, which fall outside what years of fire science has taught practitioners like her.

The basis of the prediction software is the Canadian Forest Fire Danger Rating System (CFFDRS), which forms the foundation of how agencies across the country evaluate the potential of wildland fires.

Hanes says updates to the Fire Behaviour Prediction System, part of the modelling that forecasters use, could be rolled out by 2029, but including new data from these more aggressively burning fires is unlikely because it's both very expensive and dangerous to collect. "A lot of it is reanalysis of data that was collected over numerous decades because it is really good data," she said. "But we do need to collect more information because we are trying to stretch our capabilities with the models with our knowledge of physics and math in the fire environment."

Deeper and hotter McLoughlin took us close to the ignition point of B.C.'s largest wildfire this summer, the Pear Lake Wildfire, near Clinton — a fire that had the hallmarks of behaviour that needs to be more accurately modelled.

The fire, which eventually scorched an area larger than Greater London, burned through more than 100 structures and, from where we stood, seemingly all the pines on the hill across from us, leaving a strange landscape of thousands of blackened stumps, standing like tiny spent matchsticks.

There's a lot of evidence of the impacts of drought, McLoughlin said as he walked through some of the burn scar. He pointed to 30-metre-tall trees that had burned from the top right down to the root plates, leaving huge holes in the ground, and other trees spotted black and yellow — so-called "leopard or cheetah trees," a phenomenon also observed in the Jasper wildfire — when a fire burns so hot bark explodes off the trunk in patches. Persistent drought means fires burn deeper and hotter, he said.

These are the kinds of conditions that can lead to a convective or column-driven fire, best known by their distinctive tall pyrocumulonimbus clouds that can look and act like huge thunderstorms as superheated gases and debris are pushed skyward, sometimes mixing in air from the upper atmosphere.

It used to be something you'd see once in your career, said McLoughlin. B.C.'s 2023 season had close to 100.

The fire ground we were standing in had multiple pyrocumulus events, one of them with lightning, which started a new fire that eventually merged together. "Those were not the conditions we collected data for," says McLoughlin.

Predicting the behaviour of a column-driven fire is simply "not part of our old systems," he said, as opposed to a more horizontal wind-driven fire where the models "do a great job." Lack of data on extreme conditions The last time that modelling software had a major update was 1992, but it's not as straightforward as putting some newer numbers in a spreadsheet said Hanes, from the CFS.

The capacity to do the kind of science where a timer could be put in the ground to measure the spread of a fire or infrared cameras could be used to look at the intensity of a fire is hard to imagine occurring in a life-threatening situation like the Bald Range or Pear Lake wildfire.

Even experimental burns — which is how scientists studied fires that informed the prediction systems used today — are too dangerous under the extreme conditions they're experiencing, according to Hanes.

"We don't have a lot of data on the extreme end," Hanes said, because the risk of an experimental blaze getting out of control during extreme conditions is just too high.

So, Hanes says they're often "piggybacking" their research on prescribed burns, though she admits many didn't happen this year because the season got too busy. They're also collaborating with some research groups like NASA whose planes can fly through fire storms to collect data.

McLoughlin says he's been asked multiple times about BCWS using artificial intelligence to help predict these new, seemingly unpredictable wildfires, but for him it still comes back to collecting more contemporary data.

"If we're not making an effort to collect that data on the ground to inform these models, they're not going to make good predictions," he said. "It's as simple as that."