Scaling Uncertainty Quantification in Wildfire Simulations

publication Evidence: high

Author Information

Author(s): Paul Schwerdtner, Frederick Law, Qing Wang, Cenk Gazen, Yi-Fan Chen, Matthias Ihme, Benjamin Peherstorfer

Primary Institution: Courant Institute of Mathematical Sciences, New York University

Can surrogate models trained on related data improve the efficiency of uncertainty quantification in high-fidelity wildfire simulations?

Surrogate models trained on related data significantly reduce training time and improve prediction accuracy in large-scale wildfire simulations.

Surrogate models reduced training time from 3 months to under 3 hours.
Predictions of burned area were at least twice as accurate compared to high-fidelity simulations alone.
Correlation is more important than bias for effective uncertainty quantification.

This study shows that we can use simpler models to help predict wildfires better and faster, even if those models aren't perfect.

The study uses surrogate models trained on related data to perform multifidelity uncertainty quantification in wildfire simulations.

The surrogate models are trained on biased data, which may affect the accuracy of predictions.

The surrogate models may have a large bias, but they can still provide useful correlations for predictions.

Access the complete publication on the publisher's website