Uncertainty quantification in coupled wildfire-atmosphere simulations at scale

Area of Science:

• Computational science
• Environmental modeling
• Wildfire dynamics

Background:

• Wildfire simulations are crucial for fire management and evacuation planning.
• Quantifying uncertainties in high-fidelity wildfire models is computationally expensive.
• Current methods struggle with the scale and intensity of modern wildfires.

Purpose of the Study:

• To develop a scalable multifidelity approach for uncertainty quantification in wildfire simulations.
• To demonstrate the effectiveness of surrogate models trained on related data.
• To reduce the computational cost of wildfire uncertainty quantification.

Main Methods:

• Utilized surrogate models trained on biased but correlated data.
• Implemented multifidelity approaches combining surrogate and high-fidelity models.
• Applied the method to large-scale wildfire simulations with billions of degrees of freedom.

Main Results:

• Reduced training time for uncertainty quantification by several orders of magnitude (from 3 months to under 3 hours).
• Achieved at least twice the accuracy in burned area prediction compared to high-fidelity simulations alone within a fixed budget.
• Demonstrated the practicality of multifidelity uncertainty quantification for large-scale wildfire scenarios.

Conclusions:

• Surrogate models trained on related data are effective for computationally expensive simulations.
• Correlation, not bias, is key for accelerating uncertainty quantification in multifidelity approaches.
• This method offers a scalable solution for wildfire simulation uncertainty and has broader applications in scientific computing.