Source-linked AI summary
A wildland fire model with data assimilation
Jan Mandel, Lynn S. Bennethum, Jonathan D. Beezley, Janice L. Coen, Craig C. Douglas, Minjeong Kim, Anthony Vodacek
TL;DR
The paper addresses how a fast, simplified wildland fire model can incorporate observations despite imperfect initial conditions and limited physical detail. It formulates coupled heat-and-fuel PDEs, calibrates their coefficients from wildfire behavior, and applies a regularized ensemble Kalman filter to assimilate temperature data. The resulting system reproduces observed fire behavior and can successfully assimilate data even when the initial ignition location is substantially wrong.
Problem
Wildland fire prediction requires balancing complex multi-physics with fast execution, while simulations must also accommodate imperfect models, sparse observations, and erroneous initial fire locations.
Method
The paper uses two coupled heat-and-fuel PDEs with an Arrhenius-based reaction rate, calibrates coefficients from wildfire observations, and applies a regularized ensemble Kalman filter for data assimilation.
Results
The model reproduces wildfire sensor time-temperature behavior, while the ensemble Kalman filter successfully assimilates data into simulations with substantially displaced ignition locations.
Takeaways & Limitations
A simple PDE model combined with data assimilation can produce realistic wildfire behavior and track observations despite significant initial-condition errors.
Takeaways & Limitations
The present model does not yet include atmospheric coupling, and the assimilation correction in fireline location should not exceed the reaction-zone width.
Abstract
from arXiv · showhide
A wildfire model is formulated based on balance equations for energy and fuel, where the fuel loss due to combustion corresponds to the fuel reaction rate. The resulting coupled partial differential equations have coefficients that can be approximated from prior measurements of wildfires. An ensemble Kalman filter technique with regularization is then used to assimilate temperatures measured at selected points into running wildfire simulations. The assimilation technique is able to modify the simulations to track the measurements correctly even if the simulations were started with an erroneous ignition location that is quite far away from the correct one.
1 Introduction
The paper develops a fast, simple PDE-based wildland fire model and combines it with data assimilation to incorporate observations and correct substantial ignition-location errors. It calibrates model parameters from wildfire behavior and demonstrates realistic results despite simplified physics.
- Data assimilation: The data-assimilation method modifies simulations to obtain realistic results even when initial fire locations contain significant errors.The paper presents this as a demonstration of concept using a simple model and a data-assimilation technique.
- Data assimilation: The study combines the PDE model with an ensemble Kalman filter to assimilate fire data into a running simulation.The stated objective is a real-time system steered by atmospheric, fire, fuel, terrain, and related data, although atmospheric coupling is not yet included.
- Modeling approach: The model balances physical fidelity and speed using two coupled PDEs for heat and fuel with a semi-empirical Arrhenius-based reaction rate.The approach targets a faster-running model while retaining reaction-diffusion-convection behavior.
- Modeling approach: PDE-based reaction-diffusion equations can reproduce nonlinear, unsteady behaviors such as pulsation and bifurcation that empirical models cannot.
- Calibration: Model coefficients are calibrated systematically from wildfire observations by separating qualitative solution properties from temperature, time, and space scales.Combustion-wave temperature, burning-region width, and propagation speed are used for calibration.
2 Formulation of the model
The model describes fire in a layer above the ground using coupled energy and fuel equations derived from conservation principles. Its terms represent heat transfer, wind advection, combustion, and atmospheric cooling, with coefficients calibrated from wildfire behavior.
- The model derives coupled energy and fuel equations from conservation of energy, fuel balance, and the fuel reaction rate.
- The temperature equation combines short-range radiative heat transfer, wind-driven advection, burning heat release, and convective heat loss.
- The modified Arrhenius reaction rate is offset so combustion is zero at ambient temperature while remaining smoothly dependent on temperature.
- Model coefficients are approximated from prior wildfire measurements rather than derived directly from uncertain material properties.
3 Relation to other models
Prior wildland-fire research spans physically based reaction-diffusion, fireline, fluid-coupled, and data-calibrated models. The present model relates to these approaches while addressing nonlinear combustion and fuel consumption.
- Reaction-diffusion models: Reaction-diffusion models establish traveling-wave solutions, while related wildland-fire systems study wave speed, stability, ignition, extinction, and spread.These studies include analytical, asymptotic, numerical, and heuristic treatments.
- Numerical methods: Existing finite-element work often linearizes the reaction function, omitting traveling waves and fuel consumption.Other discretization studies address nonlinear reactions or mixed finite elements for combustion systems.
- Fireline models: Thin reaction zones motivate fireline-evolution models that represent combustion as an evolving interface, often with curvature-dependent normal speed.Many such models prescribe empirically observed spread properties rather than deriving them from reaction kinetics.
- Data calibration: Relatively few studies use data to calibrate fire models, including radiation-based spread, mass-loss, and laboratory-measured spread-rate models.These examples demonstrate calibration against experiments or measurements rather than data assimilation into running simulations.
- Atmospheric coupling: Wildland-fire models have also been coupled to computational-fluid-dynamics or numerical-weather-prediction environments to represent atmospheric interactions.Examples range from detailed combustion and fluid models to semi-empirical fire-spread models coupled with weather prediction.
4 Derivation of the model
The model is derived from vertically homogenized energy and fuel balances for a thin ground-layer fire, with heat transport, combustion, convection, and wind represented explicitly. The equations are then simplified into a coefficient-identification form using a temperature-dependent fuel reaction rate.
- Model setting: The fire is modeled in a finite-thickness ground layer, with all quantities treated as two-dimensional and vertically homogenized.Coefficients are not derived directly from material properties because homogenization is highly simplified and uncertain.
- Energy balance: Heat transfer combines atmospheric radiation and convection with short-range radiative and turbulent diffusion.The model assumes convective heat transfer dominates, while radiation into the atmosphere is included in the heat-loss formulation.
- Fuel reaction: Combustion heat is modeled through a temperature-dependent reaction rate, while fuel loss is proportional to both reaction rate and available fuel.The reaction rate is CSr(T), and the heat generated per unit area is proportional to fuel lost.
- Advection and slope: The homogenized air velocity is supplied by atmospheric state data and incorporates slope effects and sub-ground wind scaling.A surface-gradient contribution approximates uphill spread, and the wind is scaled by a constant less than one within the fire layer.
- Coefficient identification: The coupled balances are rewritten with a minimal coefficient set for identification, including k, A, and C0 in the temperature equation.The resulting equation contains diffusion, advection, temperature-dependent reaction, and ambient-temperature cooling terms.
- Model simplification: The chosen heat equation omits fuel disappearance on its left-hand side because the authors seek the simplest identifiable model.This omission mainly affects the temperature profile in the reaction zone, where temperatures are highest; fuel begins at S = 1 and soon approaches a residual value.
- Reaction-rate regularization: An Arrhenius reaction law is modified with a cutoff temperature so that oxidation does not occur below T0.The modification addresses the nonzero ambient-temperature reaction implied by the unmodified Arrhenius expression.
- Numerical properties: The resulting fuel-consumption rate is smooth in temperature, which supports numerical solution compared with a cutoff-function formulation.The paper contrasts this smooth rate with the nonsmooth cutoff used in prior work.
5 Identification of coefficients
The model coefficients are identified by matching observed or desired combustion-wave behavior, while modifying the reaction rate prevents fuel loss at ambient temperature and enables sustained waves.
- Coefficient calibration: Reduced reaction dynamics provide rough coefficient values before nondimensionalization separates qualitative behavior from scale-setting parameters.The resulting approximate values initialize coefficient identification.
- Reaction-rate modification: With the unmodified Arrhenius rate, nonzero ambient-temperature reaction causes fuel loss everywhere and prevents a traveling combustion wave from developing.The fuel becomes insufficient to sustain combustion after a relatively short time, producing the cold boundary effect.
- Reaction-rate modification: Setting the temperature offset to ambient temperature forces zero ambient reaction and produces a propagating combustion wave.The modified reaction rate is therefore used in the model.
- Wave mechanism: Traveling-wave propagation arises from reaction and diffusion: heat advances into fuel ahead of the wave, while fuel depletion and cooling govern the trailing edge.Convection does not contribute to the traveling waves discussed in this section.
- Nondimensional identification: Nondimensional parameters λ and β determine qualitative solution behavior and can be varied independently during coefficient identification.Matching wave-width ratios and remaining fuel fraction guides the selection of dimensional coefficients.
- Nondimensional identification: The scaled solution is constructed to preserve the desired nondimensional properties and the wave’s maximal temperature, width, and speed.The dimensional coefficients are recovered from the scales and nondimensional parameters.
6 Data Assimilation
The data-assimilation method uses an ensemble of wildfire simulations, sequentially updated with measurements, while regularization limits nonphysical analysis states. Spatial perturbations help represent uncertainty in ignition location, which standard smooth additive perturbations fail to capture adequately.
- Assimilation challenge: The filter’s goal is to track the fireline using temperature and remaining-fuel measurements at sampled domain points.The sharp fireline makes standard initial-ensemble construction poorly represent ignition-location uncertainty.
- Regularization: Spatially shifting ensemble states supplements smooth random perturbations so the prior represents uncertainty in ignition-region location.The implementation uses distributed parallelism because ensemble members advance independently, while the Bayesian update couples them through linear algebra.
- Ensemble Kalman filter: EnKF represents the modeled-state distribution with an ensemble and alternates model advancement with data-injection analysis steps.The forecast ensemble is advanced to observation times, updated with data, and then advanced again.
- Ensemble Kalman filter: Each state vector contains temperature and fuel values at mesh nodes, while an observation function maps simulation states to synthetic measurements.The measurements are represented by a data vector with an associated error covariance matrix.
- Ensemble Kalman filter: The EnKF applies Kalman updates to individual forecast members, replacing the unknown forecast covariance with the ensemble covariance.Randomly perturbed observations are used in the memberwise update.
- Regularization: Regularization adds an independent observation based on spatial gradients to prevent large, nonphysical gradients in the analysis ensemble.The regularization matrix is practically taken as ρI and the update can be implemented by running the EnKF formulas a second time.
7 Numerical results
The numerical experiments calibrate the fire model from measured temperature behavior and test a two-dimensional ensemble under perturbed initial conditions. Repeated assimilation can attract the ensemble toward a reference solution despite a deliberately displaced ignition region, although outcomes vary across stochastic runs.
- 7.1 Calibration of coefficients in one dimension: Model coefficients are calibrated using traveling-wave characteristics and measured time-temperature curves.The calibration targets wave properties such as temperature, width, propagation speed, and the measured profile in Fig. 6.
- 7.2 Numerical results in two dimensions: The two-dimensional implementation uses central finite differences on a 250 by 250 mesh with explicit Euler time stepping.The initial state includes a 50m×50m ignition region, ambient temperature elsewhere, and a central fuel break.
- 7.2 Numerical results in two dimensions: The initial ensemble combines smooth random perturbations with spatial shifts of temperature profiles in both horizontal directions.Perturbation magnitudes and shifts are controlled by cT, cx, and cy, with bilinear interpolation used for off-grid points.
- 7.2 Numerical results in two dimensions: Each analysis cycle advances the solution by 100s before injecting artificially sampled temperature and fuel data.The reference solution is sampled every 10m, uses diagonal observation covariance, and applies regularization with ρ = 750.
- 7.2 Numerical results in two dimensions: After 10 assimilation cycles, the ensemble mean shows remarkable agreement with the reference solution despite an ignition region initially displaced by 100m.The ensemble is additionally perturbed after each cycle to maintain spread for future assimilations.
- 7.2 Numerical results in two dimensions: Different stochastic runs do not always converge to the reference solution, depending on whether a small ensemble spans a good data match.This variability reflects the dependence of attraction on the available ensemble states.
8 Conclusion
The paper shows that a simple coupled PDE fire model can reproduce measured wildfire temperature behavior and support data assimilation. A regularized EnKF with spatial state perturbations successfully assimilates fire data, while more realistic coupled modeling remains a future direction.
- Model and calibration: A model based on two coupled PDEs reproduces the time-temperature curve recorded as a wildfire passes a sensor.The curve is treated as a measurable feature of fire behavior.
- Data assimilation: Figures 8–12 compare reference data with prior and posterior ensemble behavior before and after repeated assimilation.The comparison includes the ensemble mean and variance after a single update and after 10 analysis cycles, plus a no-assimilation solution at 1000s.
- Model and calibration: Model parameters are identified from wildfire observations by separating qualitative solution properties from temperature, time, and space scales.The approach uses observed fire behavior to calibrate the model rather than relying only on theoretical coefficients.
- Data assimilation: A version of the EnKF assimilates wildfire data using nonphysical-state penalization, spatial random transformations, and smooth additive perturbations.These additions address the particular difficulty created by the thin combustion region.