A high-resolution large-eddy simulation framework for wildland fire predictions using TensorFlow

Mengqing Wang, Matthias Ihme, Rod Linn, Yifan Chen, Vivian Yang, Fei Sha, Craig B. Clements, Jenna S. McDanold, John AndersonInternational Journal of Wildland Fire 2023 · DOI

TL;DR

This paper introduces SWIRL-FIRE, a novel simulation framework built on TensorFlow and Google's Tensor Processing Units (TPUs), designed for high-resolution, large-scale wildland fire predictions. It successfully couples atmospheric and fire dynamics, demonstrating its potential for more accurate and affordable wildfire modeling.

What Problem Does This Paper Address?

Wildfires are increasing in severity and frequency, posing significant threats. Accurate prediction of wildfire behavior is crucial for effective management, planning, and scientific understanding. While existing models have advanced in physical fidelity, they often struggle with the computational demands of high-resolution simulations needed to capture complex fire-atmosphere interactions, especially at landscape scales. This gap limits our ability to predict critical phenomena like fire intermittency and turbulent dynamics, which are vital for understanding and forecasting extreme fire events.

Methods

The core of this work is the development of SWIRL-FIRE, a coupled fire/atmosphere simulation framework. It leverages:

  • TensorFlow: A popular machine learning framework, adapted here for scientific computing, enabling efficient execution on specialized hardware.
  • Tensor Processing Units (TPUs): Google's custom hardware accelerators designed for matrix computations, offering significant speedups for the types of calculations involved in fluid dynamics and combustion simulations.
  • Large-Eddy Simulation (LES): A technique for simulating turbulent flows by resolving large-scale eddies and modeling smaller ones, providing a good balance between accuracy and computational cost for atmospheric and fire dynamics.
  • Low-Mach Number Approximation: This simplification in the governing equations for the gas phase allows for larger time steps by removing the constraint of acoustic wave propagation, further enhancing computational efficiency.
  • Coupled Fire Model: The framework integrates sub-models from FIRETEC to represent combustion and multi-phase interactions with vegetation, capturing the crucial coupling between fire and atmosphere.

Datasets

The primary dataset used for evaluation was experimental data from the FireFlux II experiment. This prescribed grassland fire provided measurements of ambient fuel and wind conditions, as well as fire-spread data, serving as a benchmark for validating the simulation framework's accuracy and performance. No new datasets were introduced.

Key Insights and Ideas

  • TensorFlow and TPUs for Wildfire Science: This paper demonstrates the viability and benefits of using machine learning frameworks and specialized hardware (TPUs) for complex physics-based simulations like wildland fire modeling, opening doors for more efficient and scalable research.
  • Mesh Refinement vs. Global Quantities: While finer mesh resolutions significantly improve the prediction of intermittent fire behavior and turbulent dynamics, they have a surprisingly limited impact on global quantities like fire scar area and spread rate, suggesting these are governed by larger-scale processes.
  • Computational Cost vs. Resolution Trade-off: The study highlights that achieving high-resolution simulations, crucial for capturing fine-scale fire dynamics, is becoming more feasible with advancements in hardware and software, making landscape-scale predictions at resolutions of 1 meter achievable within reasonable timeframes.
  • Importance of Fire-Atmosphere Coupling: The simulations effectively capture the interplay between atmospheric turbulence and fire dynamics, leading to phenomena like counter-rotating vortices and plume tilt, which are critical for understanding fire behavior.

Key Results

  • Reasonable Agreement with Experiment: The SWIRL-FIRE framework, when applied to the FireFlux II experiment, showed good qualitative and quantitative agreement with experimental data for fire spread, temperature, and velocity profiles, especially after the fire became established.
  • Scalability: The framework exhibits close to linear scalability on TPU hardware, meaning that as more processing units are added, the simulation time decreases proportionally, enabling larger and more complex simulations.
  • Mesh Resolution Impact: Finer mesh resolutions (e.g., 0.5m horizontal) were found to be essential for accurately capturing intermittent fire behavior and small-scale turbulent motion, while coarser resolutions (e.g., 2m horizontal) were sufficient for predicting global quantities like fire spread rate.
  • Computational Efficiency: The use of TensorFlow and TPUs allows for simulating meso-scale wildfires at high spatial resolutions (e.g., 0.5m) at an affordable computational cost.

Limitations and Future Work

  • Early Stage Discrepancies: The paper notes discrepancies in predicting fire behavior during the early stages after ignition. These are attributed to uncertainties in wind and fuel conditions, as well as limitations in the combustion model's representation of turbulence-interaction.
  • Fuel and Wind Uncertainty: The authors acknowledge that more accurate characterization of experimentally observed wind profiles and fuel conditions could further improve simulation accuracy.
  • Wall Models: Improvements in wall models could enhance predictions of turbulence/fire interaction in the viscous boundary layer region.
  • Machine Learning Integration: While the framework is built on TensorFlow, the current study focuses on its physics-based simulation capabilities. The authors suggest that the utilization of machine learning for aspects like parameterization or prediction enhancement will be addressed in future work.
  • Burning Rate Convergence: The burning rate was not fully converged in the current study, suggesting that further mesh refinement might be required for this specific metric.

Summary updated

Abstract

Background Wildfires are becoming more severe, so we need improved tools to predict them over a wide range of conditions and scales. One approach towards this goal entails the use of coupled fire/atmosphere modelling tools. Although significant progress has been made in advancing their physical fidelity, existing tools have not taken full advantage of emerging programming paradigms and computing architectures to enable high-resolution wildfire simulations. Aims The aim of this study was to present a new framework that enables landscape-scale wildfire simulations with physical representation of combustion at an affordable cost. Methods We developed a coupled fire/atmosphere simulation framework using TensorFlow, which enables efficient and scalable computations on Tensor Processing Units. Key results Simulation results for a prescribed fire were compared with experimental data. Predicted fire behavior and statistical analysis for fire spread rate, scar area, and intermittency showed overall reasonable agreement. Scalability analysis was performed, showing close to linear scaling. Conclusions While mesh refinement was shown to have less impact on global quantities, such as fire scar area and spread rate, it benefits predictions of intermittent fire behavior, buoyancy-driven dynamics, and small-scale turbulent motion. Implications This new simulation framework is efficient in capturing both global quantities and unsteady dynamics of wildfires at high spatial resolutions.
BibTeX
@article{wang2023high,
  author = {Mengqing Wang and Matthias Ihme and Rod Linn and Yifan Chen and Vivian Yang and Fei Sha and Craig B. Clements and Jenna S. McDanold and John Anderson},
  title = {{A high-resolution large-eddy simulation framework for wildland fire predictions using TensorFlow}},
  journal = {International Journal of Wildland Fire},
  year = {2023},
  doi = {10.1071/wf22225},
}
Figure