Uncertainty quantification in coupled wildfire-atmosphere simulations at scale

Uncertainties in wildfire simulations pose a major challenge for making decisions about fire management, mitigation, and evacuations. However, ensemble calculations to quantify uncertainties are prohibitively expensive with high-fidelity models that are needed to capture today’s ever more intense and severe wildfires. This work shows that surrogate models trained on related data enable scaling multi-fidelity uncertainty quantification to high-fidelity wildfire simulations of unprecedented scale with billions of degrees of freedom. The key insight is that correlation is all that matters while bias is irrelevant for speeding up uncertainty quantification when surrogate models are combined with high-fidelity models in multi-fidelity approaches. This allows the surrogate models to be trained on abundantly available or cheaply generated related data samples that can be strongly biased as long as they are correlated to predictions of high-fidelity simulations. Numerical results with scenarios of the Tubbs 2017 wildfire demonstrate that surrogate models trained on related data make multi-fidelity uncertainty quantification in large-scale wildfire simulations practical by reducing the training time by several orders of magnitude from three months to under three hours and predicting the burned area at least twice as accurately compared to using high-fidelity simulations alone for a fixed computational budget. More generally, the results suggest that leveraging related data can greatly extend the scope of surrogate modeling, potentially benefiting other fields that require uncertainty quantification in computationally expensive high-fidelity simulations.