University of Thessaly (UTH)
Modeling and uncertainty quantification of wildfire spread using field and laboratory data
Abstract
dc:descriptionThe escalating frequency and severity of global wildfires underscore the need for accurate predictive models to support effective wildfire management and protect both communities and ecosystems. Fire spread is a multiphysics process governed by the combined effects of thermochemical degradation, fluid dynamics, and turbulent heat and mass transfer. To address this complexity, this thesis introduces ADfiRe, a wildfire spread simulation framework built on advection-diffusion-reaction principles that combines empirically established formulations to calculate crucial quantities typically resolved through CFD, without the associated computational cost. Specifically, this work makes four primary contributions: (i) detailed computational modeling and simulation of wildland fire spread, (ii) physics-constrained parameter estimation, (iii) geometry-aligned reduced-order modeling for accelerated uncertainty propagation and prediction, and (iv) comprehensive experimental investigation of controlled burns. Current wildfire modeling approaches occupy two extremes, ranging from empirical frameworks that enable rapid, operational predictions at the risk of generating physically inconsistent results, when unsteady dynamics are ignored, to high-fidelity CFD simulators that capture fuel-fire-atmosphere interactions but remain computationally prohibitive for real-time deployment. ADfiRe tries to bridge this gap through a surface-level model in which fire spread is governed by transient convective, diffusive, and radiative heat fluxes. No attempt is made to explicitly solve the fluid mechanics problem, instead, the flow field required in the energy balance is derived from analytical arguments. The model draws on established fuel databases and accommodates multiple fuels with distinct burning characteristics, incorporates arbitrary complex topography from available elevation data, and captures wind-induced flame tilting, distinguishing between plume- and wind-dominated spread through a modified Byram number. Its predictive capability is further validated against multiple sets of well-documented fires. Building on the ADfiRe framework, the thesis turns to parameter estimation, introducing a physics-informed neural network (PINN) that identifies the unknown parameters of the spread model. The methodology couples neural networks with the physical constraints of fire modeling, through a composite loss function that jointly penalizes mismatch with observed data, residuals of the governing physics-equations, and violations of the initial and boundary conditions. This loss is minimized using the Adam and L-BFGS optimizers. Furthermore, the network is trained using synthetic spatio-temporal data generated by the ADfiRe simulator, alongside thermal imagery from the 2002 Troy Fire in California. The PINN accurately estimates the unknown model coefficients, even in the presence of data noise. The study next addresses the computational cost of forward uncertainty propagation by introducing a geometry-aligned multi-fidelity strategy. Whereas traditional linear reduced order methods struggle with the sharp, moving fronts, producing numerical oscillations, the proposed approach maps low- and high-fidelity snapshots onto a common reference domain (via shift/stretch and affine transformations). This alignment enables direct comparison of combustion-front structures and non-linear stabilization without requiring large offline datasets. Validated on ADfiRe benchmarks, the surrogate overcomes the slow decay of the Kolmogorov n-width, removes Gibbs-type oscillations, reconstructs full-field distributions of temperature and fuel load, and finally accelerates model predictions by three orders of magnitude. Finally, this research examines the dynamic, often eruptive behavior triggered by interacting dual parallel fire fronts. Controlled experiments conducted at the University of Coimbra’s Forest Fire Research Laboratory (LEIF) examined parallel fronts initiated at various separation distances and across multiple slope angles, evaluating the approaching rate of spread (ROS) and the convective structures driving flame tilt. Results show that steeper slopes accelerate the approaching ROS, as slope-aligned buoyant flow reinforces lateral wind-induced flow, while the dynamic nature of the flames is denoted by the instantaneous ROS oscillating with the separation distance. Slope angle emerges as the dominant parameter governing the interaction, with fronts exhibiting mutual flame tilting during convergence and partial detachment after merging, driven by the formation of counter-rotating vortices.
Degree
thesis:*- Grantor dc:publisher
- University of Thessaly (UTH)
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Vogiatzoglou, Konstantinos
- Βογιατζόγλου, Κωνσταντίνος
Subjects
dc:subject × 16- Μοντελοποίηση εξάπλωσης δασικών πυρκαγιών
- Πλαίσιο συναγωγής-διάχυσης-αντίδρασης
- Ποσοτικοποίηση αβεβαιότητας
- Νευρωνικά δίκτυα με ενσωματωμένη φυσική πληροφορία
- Μοντελοποίηση μειωμένης τάξης πολλαπλής πιστότητας
- Αλληλεπίδραση μετώπων πυρκαγιάς
- Wildfire spread modeling
- Advection-diffusion-reaction framework
- Uncertainty quantification
- Physics-informed neural networks
- Multi-fidelity reduced-order modeling
- Fireline interaction
- Περιβαλλοντική Μηχανική
- Επιστήμες Μηχανικού και Τεχνολογία
- Environmental Engineering
- Engineering and Technology
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/62913