FireBench: Using high-performance computing to advance machine learning and wildfire research

The severity and frequency of large wildfires has increased significantly over recent years due to factors ranging from climate and weather pattern changes to increased human activities in wildland-urban interfaces. While wildfires play an important role in some forest’s natural cycle, extreme fires pose serious threats to communities and ecosystems. Frequent wildfires can disrupt, damage, and destroy infrastructure, livelihoods, lives, and properties. For example, the recent surge in US wildfires has expanded geographically with the annual burned-area estimated to be approaching 7M acres, annual wildfire economic burden to be between $394B and $893B, and annual wildfire CO2 emissions exceeding 50% of combustion emissions. In fact, greenhouse gas emissions from wildfires can wipe out years of emissions savings. This is expected to worsen worldwide, not only in fire-prone areas within the US, Canada, Australia, and southern Europe, but also in regions that haven’t had a history of extensive wildfires.

Firefighters and the research community have been studying paths to better understand and manage wildfire impacts. With rapid machine learning (ML) and high performance computing advancements, Google has explored ways to apply this technology to improve predictions for fire-risk assessment and fire resilience to help communities and authorities manage wildfires. Some examples include using AI for wildfire boundary tracking, using ML to predict fire spread from remote-sensing data, and releasing an efficient and scalable high-fidelity TPU-powered simulation framework that can reduce data scarcity for ML-based fire-prediction model development. However, a key element to effectively leveraging ML technologies for fire management is finding high-quality data, which can be difficult.

To that end, in “FireBench: A High-fidelity Ensemble Simulation Framework for Exploring Wildfire Behavior and Data-driven Modeling” we introduce a high-resolution, simulation dataset designed to advance wildfire research. FireBench enables investigations of wildfire spread behavior and the coupling between atmospheric hydrodynamics and fire physics by extending beyond just fire states to also include a comprehensive list of flow field variables in three dimensions. It also supports the development of robust and interpretable ML models by capturing the underlying dependencies between relevant variables. To provide the research community with the insights needed to mitigate the impact of wildfires, we have released the FireBench dataset on the Google Cloud Platform.

As a demonstration of the dataset, we plot below the instantaneous flow-field results for nine ensembles out of the 117 in the dataset. The snapshots are taken at a representative time when the fire front reaches the middle of the slope.

FireBench is available on Google Cloud Storage, and is organized following a tree structure. The GitHub repo contains a Colab notebook and Apache Beam pipeline that demonstrate how to access and process the dataset. The notebook presents a simple UI to browse the Firebench dataset. The script in the notebook scans the simulations in the public Firebench dataset to populate its UI dropdowns. Additionally, it can plot a simple slice of the data given the selections in the UI. The Apache Beam pipeline demonstrates how to postprocess the Firebench dataset. Each simulation contains several terabytes of data, so we use Apache Beam to run post-processing code in parallel on multiple machines to reduce the execution time. This sample script computes a time series of mean, minimum and maximum of all variables for a given simulation.

FireBench opens up the opportunity for fire propagation model development, fire and atmospheric dynamics investigation, and ML tasks that are related to turbulent multiphase fluid flows. In the near future, FireBench will be employed as a component of a dataset for the construction of the Future Learning Approaches for Modeling and Engineering (FLAME) AI workshop at Stanford University, which fosters a forum for exchanging ideas, data, methods, and models related to ML techniques for fluid dynamics, turbulence, and combustion fields crucial to the development of energy, propulsion, climate, and safety systems.

Google Research is committed to advancing wildfire research through the development and application of innovative technologies like FireBench as well as building technologies pushing state of the art for tracking wildfire spread. By fostering collaboration and knowledge-sharing within the research community, we aim to accelerate progress towards a future where AI-powered tools enhance wildfire prediction, prevention, and management. We believe that open access to high-quality datasets like FireBench will empower researchers to develop more sophisticated models and strategies, ultimately mitigating the devastating impacts of wildfires on communities and ecosystems globally.

This dataset is generated by the researchers in Google Research, Athena Team: Cenk Gazen, Yi-Fan Chen, Matthias Ihme, and John Anderson. Special thanks to Carla Bromberg for making it publicly available. Also thanks to Karl Alexander Toepperwien for generating the animation.