FireSim
A large-scale high-fidelity wildfire simulator for early warning and extreme fire development prediction
Wildfire simulation
The ability to accurately predict when, where, and how fires spread is a long sought after goal for firefighters and fire risk managers. The ability to predict fire spread enables effective fire risk management before fires occur, and efficient resource deployment and planning during the fire operations, which all can greatly reduce damages caused by fire.
We are leveraging Google’s Compute, TPUs, and applying machine learning to bring improvements into wildfire prediction. This includes the development of FireSim, a large-scale high-fidelity wildfire simulator (published in The International Journal of Wildland Fire) that we’ve used to generate large quantities of data covering a wide range of wildfire scenarios. FireSim has allowed us to generate datasets, such as the recently announced FireBench dataset, that provide the wildfire research community with insights needed to develop effective mitigation solutions.
This simulator effectively addresses the data sparsity issue and allows for better ML being developed to address various fire predictions, such as early warning for extreme fire development.
A controlled burn field experiment (FireFlux II) simulated using Google’s high fidelity fire simulator.
Fire spread and propagation models
In a related effort, we are working in close collaboration with the US Forest Service Fire Lab, using Google parallel computing and machine learning technologies, to develop the next generation of operational fires spread models. A Google-US Forest Service joint presentation in ICFFR 2022 and the USFS Fire Lab's project page provides a glimpse into the ongoing work.
Our next challenge is to look at use cases and train ML-based models to incorporate weather and satellite data into forecasts of fire propagation, and to be able to predict fire propagation.
Google Research YouTube interview: Simulating wildfires before they happen