acceptodds
Under review as a conference paper at ICLR 2027

WUIRL: Reconstructing Wildfire Structure Ignition in the Wildland–Urban Interface with Reinforcement Learning

Abstract

Wildfires are increasingly burning through the wildland–urban interface (WUI), destroying homes and other structures where development meets wildland vegetation. Modeling which buildings are most at risk remains difficult because the physical mechanisms governing fire spread and ignition in the WUI are incompletely understood, data on WUI and wildland fuel are coarse, and historical fire progression is too sparsely observed to learn directly. Supervised methods currently map building, vegetation, and weather data to structure outcomes. However, these static inputs cannot capture the mechanisms by which fire reaches a structure over time, which is necessary to predict ignition and determine an appropriate method to mitigate risk. Capturing the ignition process requires temporal ignition data, which is rarely recorded, leaving the process that determines survival and guides mitigation largely undocumented. Therefore, we introduce Wildland–Urban Interface Reinforcement Learning (WUIRL), a framework that reconstructs how fire reaches and ignites buildings from limited observations by pairing proximal policy optimization (PPO) with a physical ignition model. The learned policy selects pathways through which burning vegetation, neighboring structures, and external sources expose buildings to heat, and the physical model determines how each building responds. WUIRL thus links structural vulnerability and evolving exposure, reconstructing progression from recorded damage and coarse timing without complete exposure histories. We apply WUIRL to the 2025 Eaton fire, recovering 78.9% of buildings labeled as ignited from damage records while achieving the highest balanced accuracy among the evaluated policies. The reconstruction is most sensitive to the modeled radiant heat and firebrand-proxy entry mechanisms, whose removal most degrades performance. WUIRL provides a foundation for future studies of structure loss under dynamic exposure, offering the potential to better test specific interventions' role in limiting future damage as higher resolution observations become available.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Reinforcement learning, physics-guided machine learning, wildfire reconstruction, graph representation learning, climate and sustainability

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ZcA6OpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes WUIRL, a physics‑constrained reinforcement‑learning framework. Using only limited post‑wildfire damage records and coarse temporal data, it reconstructs building ignition‑spread pathways in the wildland‑urban interface. On the 2025 Eaton fire case study, it achieves a reasonable balance between predicting burned and surviving buildings. Still, it suffers from inaccurate temporal reconstruction, training performance degradation, and a lack of cross‑fire generalization validation.

Critical strengths

This work innovatively couples a fixed domain‑specific physical simulator with reinforcement learning. The RL agent only learns to select fire‑exposure pathways instead of replicating intricate combustion physics, lowering the learning burden for RL and offering a novel scientific‑machine‑learning paradigm for post‑hoc wildfire temporal reconstruction.

Critical weaknesses
  1. Poor POMDP handling. Though recognized as a partially‑observable setting, no POMDP‑specific solutions are adopted. Vanilla memory‑free PPO is directly applied without recurrent networks for history encoding. The policy relies on static GraphSAGE embeddings and pooled statistics, which discards temporal information and may harm pathway‑selection quality.
  2. Heuristic reward design with insufficient ablation. Key reward hyperparameters (temporal/final‑outcome weighting, doubled weight for boundary buildings) are ad‑hoc without calibration. Potential conflicts between temporal and final‑outcome objectives are observed, yet ablation studies are missing to show performance trade‑offs under varied reward weights.
  3. Incomplete SMDP treatment and ambiguous credit assignment. Multiple actions can happen within one physical time window under the SMDP formulation. Despite time‑dependent GAE usage, concurrent‑action credit‑assignment challenges are undiscussed. This is a plausible cause for the observed training degeneration but receives little analysis.
  4. Risk of RL compensating physical‑model misspecification. The physical ignition simulator contains many assumed, fitted and extrapolated parameters. The RL policy could learn physically unrealistic pathways to compensate simulator errors and still match ground‑truth burn labels, threatening the physical plausibility of reconstructed trajectories.
  5. Incomplete ablation suite. Physical‑mechanism ablations are only executed at evaluation time (no retraining), limiting their interpretability. Systematic ablations for core RL components (GraphSAGE encoder, temporal reward, factorized action space) are absent, making it hard to quantify each module’s actual contribution.
Questions for authors
  1. This task is formulated as a POMDP, yet memory‑free vanilla PPO is used without LSTM/GRU‑style recurrent modules for history encoding. The GraphSAGE encoder is also pre‑trained purely on static building attributes. Please analyze biases induced by partial observability, justify the omission of recurrent architectures, and provide arguments that static graph embeddings are sufficient for dynamic fire‑spread pathway selection.
  2. Hyperparameters including temporal‑term weights, the scaling factor for final‑outcome term, and doubled weights for boundary buildings are set heuristically without sweeping or ablation. Please justify these hyperparameter choices and supply reward‑weight ablation experiments to demonstrate trade‑offs between final‑damage metrics and temporal‑reconstruction performance.
  3. This is an SMDP environment permitting multiple concurrent actions per physical time window, creating challenging credit‑assignment conditions which may contribute to training degeneration. Please explain how credit is assigned across co‑occurring actions and diagnose to what degree SMDP‑related credit‑assignment issues drive training degradation.
  4. The ignition simulator contains numerous assumed and extrapolated parameters. There is risk that the RL policy learns physically implausible exposure pathways to compensate for simulator defects and fit burn labels. Please provide diagnostic analyses or experiments to disentangle performance gains originating from physically plausible pathway reconstruction versus policy compensation for model misspecification. Also discuss whether evaluation‑time physical ablations could be biased by such compensation effects.
  5. Current physical‑mechanism ablations are executed at evaluation time without retraining, and ablations over core RL components are lacking. Please add controlled retrained baselines where physical modules are disabled from the start of training. Also perform component‑wise ablations over the GraphSAGE encoder, temporal‑reward term, and factorized action space to validate their individual contributions. Further discuss whether ablation conclusions are bounded by the single‑fire dataset.
7 comments

We thank the reviewer for identifying our "novel scientific-machine-learning paradigm for post-hoc wildfire temporal reconstruction." WUIRL is, to our knowledge, the first method to reconstruct building-level fire spread scored against an hourly record, and it reports the range of histories the physics and the record allow, with a confidence for each. We have addressed each concern below with additional experiments. We hope these additions address the reviewer's concerns. We would appreciate reconsideration of the evaluation in light of them, and welcome continued discussion.

The review summary notes limited data and three limitations, which we address in turn.

The limited data reflects the field rather than WUIRL. No fire records which neighbor ignited which building, and only a few earlier fires have detailed timelines [1, 2]. The Eaton Fire is what makes this work possible. To our knowledge, it is currently the only dataset that supports building-level reconstruction, with both an inspection of every structure and hourly progression across a whole community [3]. WUIRL is built for this kind of record, and its reconstructions will improve as better records become available. We will state this more clearly in the limitations.

Timing accuracy is limited by the record. The data give only coarse hourly progression, not when each building ignited (Sec. 3, line 145; Sec. 4.8, lines 453–454), so no model's timing can be checked more finely than the hour, and the paper treats timing as a secondary signal for this reason (Sec. 4.5, lines 385–399). We sampled 30 histories from the trained policy and compared their agreement on each building's hour with the record. Where the runs agree, they typically fall within an hour of the record, so agreement across runs is an informative measure of timing confidence. We will add this analysis to Section 4.5.

The reviewer brings a decline in training to our attention, which has improved our work. We reweighted the delayed final-damage term in training and retrained to 500k actions. Training stays stable (reconstruction score −4.45 at the 500k-action checkpoint, ±0.056 across seeds), and the selected checkpoint performs on par with the paper's (−4.32 against −4.39), so the results and conclusions of this paper remain unchanged, and our optimization's robustness is improved (Q3). We will update Section 4.7 with this default.

To test generalization within the data available, we held out each of five Eaton regions from training in turn. On unseen ground WUIRL's score falls half as far as CEM's (reconstruction score −4.40 on training regions, −4.73 on held-out ones). We will add the five-region results to Section 4.6.

RL fits this problem because reconstruction is a cascade of decisions with no labels for how each building ignited, only records to score against (Sec. 1, lines 48–55). A stochastic policy inside fixed physics also yields the many histories the data support, rather than one. We will clarify this motivation in the introduction.

The five weaknesses correspond to the five questions (Q1–Q5), so we answer each briefly here and give the full experiment under the matching question.

WUIRL handles partial observability through the physical state rather than a recurrent network (Sec. 3.3, lines 225–228), since each building's accumulated dose carries its exposure history (Sec. 3.2, lines 200–204). Adding the two quantities the policy cannot see and retraining gives no improvement (reconstruction score −4.39 with the added inputs vs −4.32 without) (Q1). We will add this test to the appendix.

The weights are design assumptions set from observation reliability (Appendix D.1, lines 1242–1266) and checked on a synthetic fire before training. We swept every weight the reviewer names and retrained. Timing and final damage do not conflict once training is stable, removing the boundary doubling costs 0.34 in score, and outcome weight 1 is the only setting stable at 500k (Q2). We will add the sweep to the appendix.

We agree that same-minute decisions share credit (Appendix E.7, lines 1614–1644). We retrained with two changes that leave same-minute credit untouched, a shorter horizon and a smaller outcome weight. Both remove the collapse (reconstruction score at the 500k checkpoint −4.447 and −4.445, against −5.486 for the paper's setting), so the delayed final-damage term caused it, not concurrency (Q3). We will add this diagnosis to Section 4.7.

The policy can choose only pathways the physics admits, and the physics alone decides ignition (Sec. 3, lines 157–161), using published relationships (Appendix B). We retrained without each mechanism and reran the trained policy under perturbed parameters. No policy fits without radiation or firebrands (reconstruction score −5.52 and −5.47, against −4.39), and the policy's score lead over an untrained one holds under ±20% heat flux. The policy is therefore not compensating for the physics, with the exception of firebrand timing, the one parameter without a published value (Q4). We will add these tests to the appendix.

We retrained every mechanism, and the RL components the review names. Radiation and firebrands dominate (removing either costs over 1.0 in reconstruction score), followed by vegetation (0.33). The learned receiver stage contributes 0.16, the GraphSAGE embeddings 0.10 and the hourly reward term 0.08 (Q5). We will add the retrained ablations to Section 4.4.

Q1. The physical state already carries the relevant history, so a learned memory adds little. The policy's observation is defined in Appendix C.1, lines 1172–1180.

Experiment 1. Missing information. We gave the policy the two quantities it cannot see, pending firebrand progress and heat absorbed across the frontier, and retrained. There was no improvement (reconstruction score −4.39 with the added inputs vs −4.32 without), so the accumulated dose (Sec. 3.2, lines 200–204) already carries the history a memory would provide. We will add this test to the appendix.

Experiment 2. Firebrand timing. We shortened the firebrand gate from 900 s (Sec. 4.4, line 372) to 600 s at evaluation. The trained policy's score lead over an untrained one falls from 0.25 to 0.09, so firebrand timing is the one place where a memory could still help. We will add this test to the appendix.

Experiment 3. Static embeddings. We zeroed the GraphSAGE embeddings and retrained. The score drops only 0.10, because the policy's view still changes every step through live heat and dose (Sec. 3.3, lines 225–231), so static embeddings are sufficient. We will add this test to the appendix.

We will also note a recurrent policy as future work.

Q2. The reward was designed and checked before training and then fixed, and the new sweep tests how sensitive the results are to it. The terms are defined in Appendix D.1, lines 1242–1266. Timing gets a ±1-hour window for hourly-map error, final outcomes get a tighter tolerance and twice the weight because inspections are near-complete, and boundary buildings get twice the weight because they carry the most information about spread.

Experiment 1. Synthetic check. On a synthetic 12-house, 8-hour fire, we scored every single-house timing change plus whole-cascade controls. The scores ranked as intended (correct −0.154, one-hour shift −0.208, wrong intermediate progression −0.444, wrong boundary outcomes −5.873, random −7.037). We will add this check to the appendix.

Experiment 2. Weight sweep. We changed one training weight at a time, kept the locked score for evaluation, and retrained to 500k actions. The outcome weight multiplies R_final (2 in the paper), the boundary weight multiplies boundary buildings' outcome score (2), the hourly weight multiplies R_hourly (1), and the curriculum phases the hourly term in from 150k to 300k actions.

Training settingScoreTimingFinal damageBurn / survivor recall500k checkpoint
Outcome weight 1−4.317−0.933−1.69284.7 / 47.0−4.445
Outcome weight 2 (paper)−4.385−0.946−1.72067.9 / 62.5−5.486
Outcome weight 4−4.644−0.976−1.83462.4 / 62.0−4.792
Boundary weight 1 (paper 2)−4.727−0.988−1.87057.9 / 64.7−5.249
Hourly weight 0.5 (paper 1)−4.468−0.958−1.75568.3 / 60.3−5.113
Hourly curriculum (0 to 150k, full by 300k)−4.386−0.940−1.72376.6 / 53.6−4.848

Timing and final damage move together across settings, so the mixed result in Appendix E.4 (lines 1507–1541) came from collapsed final checkpoints. The real trade-off is between burned and surviving buildings. The boundary weight matters (removing it costs 0.34 in score), and outcome weight 1 is the only setting stable at 500k (score −4.445 at 500k, against −4.79 to −5.49 for the other settings), so we adopt it, with the evaluation score unchanged (Q3). We will add the synthetic check and the sweep to the appendix.

Reviewer_jzuDOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes WUIRL, a reinforcement learning framework that reconstructs building-scale wildfire ignition sequences in wildland-urban interfaces from sparse observations. Fire spread is formulated as a sequential decision problem: a learned policy selects exposure pathways, and a fixed physics-based model determines whether a building ignites. It leverages graph representations, a GraphSAGE encoder and masked PPO, supports multiple ignition mechanisms, and is evaluated on the 2025 Eaton Fire dataset.

Critical strengths

The task formulation is distinctive. It recovers hidden ignition sequences and combines sequential path search with physical ignition rules, offering a new perspective compared with conventional damage prediction models.

Critical weaknesses

As acknowledged in the paper, different ignition sequences may produce identical observed damage outcomes, making it difficult to validate the authenticity of the reconstructed pathways. Furthermore, the paper only compares WUIRL against Random and Greedy baselines, which is insufficient to convincingly demonstrate the advantages of reinforcement learning over simpler reconstruction methods.

Questions for authors
  1. Given that multiple ignition pathways may explain the same observed damage outcomes, how do the authors quantify the uncertainty of the reconstructed pathways?
  2. Could the authors provide stronger evidence or further analysis to support their reliability?
6 comments

We thank the reviewer for noting that the formulation offers "a new perspective compared with conventional damage prediction models." We have addressed each concern below with additional experiments. We hope these additions address the reviewer's concerns. We would appreciate reconsideration of the evaluation in light of them, and welcome continued discussion.

The review raises two concerns, that pathways cannot be validated from damage records and that two baselines are too few. In short, WUIRL reconstructs the plausible fire histories that fit the physics and the hourly record, with a confidence for each, rather than a single true pathway, and our answers to Questions 1 and 2 test that claim directly. For the second concern, we ran six additional baselines, and PPO transfers to unseen regions better than the strongest of them.

We agree that different ignition sequences may produce the same observed damage, and we do not claim the single true pathway. WUIRL returns all the histories that fit the physics and the record, and how likely each one is. Currently available fire data cannot confirm pathways for any model (Sec. 4.8, lines 453–459), since fires are recorded by final damage and, at best, hourly perimeters, and without a shared theory of spread no model's reliability can be confirmed [1]. WUIRL also narrows them down. Sequences that miss the hourly timeline score worse (Sec. 3.3, Eq. 3, line 248), and sequences the physics cannot produce never arise (Sec. 3.2, lines 185–212).

To address the reviewer's concern about limited baselines, we include six additional baselines across areas of machine learning, and outline the advantages, limitations and suitability of each below.

Rule-based policies fix the decision rule in advance. These are simpler reconstruction methods as suggested by the reviewer. Random and Greedy (Sec. 4.2, lines 286–291) are the paper's baselines. Nearest source and earliest source choose the source by distance or arrival order instead of heat. Select-all accepts every pathway, an upper bound on burn recall. The heat-flux threshold rule is Greedy with a heat-flux cutoff tuned on the score.

Derivative-free optimization tunes a fixed rule directly against the score. These are non-RL alternatives to reinforcement learning. The cross-entropy method (CEM) optimizes a fixed weighting of 12 pathway features against the score, but cannot adapt to the fire's state [2]. Nelder–Mead searches the same weights locally [3].

Decision rule, all inside WUIRL's environmentReconstruction score
Rule-based policies
Random−4.582
Greedy−4.501
Nearest source−4.412
Earliest source−4.490
Select-all−4.507
Heat-flux threshold rule (tuned)−4.216
Derivative-free optimization
Nelder–Mead (tuned)−5.138
CEM (tuned)−4.140
Reinforcement learning (ours)
PPO−4.317

Why RL. Reconstruction is a cascade of decisions with no labels for how each building ignited, only delayed feedback from the record (Sec. 1, lines 48–55), which is the setting RL is designed for. PPO outperforms every untuned rule. The two methods tuned directly on the score lead it by 0.1–0.2 on the full fire, close to its run-to-run variation of ±0.11, but they search for one fixed rule, while PPO learns a policy that responds to the fire's state. The difference shows on regions held out of training.

OptimizerTraining regionsHeld-out regionsDrop
PPO (ours)−4.395−4.7330.34
CEM (tuned)−4.008−4.6970.69

Transfer. On held-out regions PPO loses half as much as CEM, so CEM's 0.39 lead shrinks to 0.04, and region by region the two are even (two regions each, one tied). The learned parts of the policy matter. Retrained with the learned receiver choice replaced by a uniform one, the score falls 0.16, and with the graph embeddings zeroed it falls 0.10.

Uncertainty. As the policy is stochastic, PPO is the only method that produces a range of candidate histories, which the uncertainty analysis in Question 1 requires. It also recovers more burned buildings than either tuned rule (burn recall 85% vs 77–79%).

Framework. Our contribution is the methodology of physics-constrained wildfire reconstruction. We pose reconstruction as sequential, physically constrained pathway selection scored against the record, every method above runs inside it (Sec. 4.2, lines 288–289), and any objective can be used. We chose PPO because it is standard and robust, and because physical rollouts, rather than training, dominate compute (Sec. 4.8, lines 464–465). We will add the full table to the appendix.

(Q1) We quantify uncertainty by sampling.

Experiment. Sampling histories from the policy. The policy is stochastic, so we ran it 30 times (3 seeds × 10 rollouts). Each run is a different physically possible history that fits the record about equally well (reconstruction score −4.32 ± 0.11 across runs). For each building we recorded how often it burned, from which neighbors, by which route and when, and how often a pathway appears across runs is its estimated likelihood. A burned building is ignited from a median of 6 different neighbors across runs, so several pathways fit the record, and that multiplicity is itself a finding. Of buildings burned in at least 27 of 30 runs, 72% were recorded as damaged in the inspection, against 48% for those burned in 4 to 26 runs and 23% for those burned in 3 or fewer. The share of runs in which a building burns therefore works as a confidence score. The figure below summarizes the result, and we will add this analysis to the manuscript.

Share of buildings recorded as damaged, by how many of 30 sampled histories burn them (https://anonymous.4open.science/api/repo/WUIRL--C8EC/file/fig_agreement_simple.png)

(Q2) We show that the pathways rely only on physics the data support, and that they hold on regions the policy never saw.

Experiment 1. Retraining without key mechanisms. We retrained WUIRL once without building-to-building radiation and once without firebrands. Neither model reproduces Eaton. The reconstruction score (higher is better) falls to −5.52 and −5.47, compared with −4.39 when both mechanisms are present, so the Eaton record cannot be matched without them. We will add this retraining to Section 4.4.

Experiment 2. Held-out regions. We held out each of five Eaton regions in turn, trained without it and scored WUIRL there. WUIRL's score falls half as far there as CEM's (reconstruction score −4.40 on training regions, −4.73 on held-out ones), so the policy learns a rule that transfers rather than fitting the regions it was trained on. We will add the five-region results to Section 4.6.

Experiment 3. Heat-flux perturbation. We changed building heat flux by ±20% and reran the trained policy. Its score lead over an untrained policy holds (0.24 at −20% and 0.25 at +20%, against 0.25 at nominal flux), so the reconstruction does not depend on the exact flux values used in training. We will add this test to Section 4.4.

We will also clarify this in the discussion and limitations.

References

[1] M. A. Finney, J. D. Cohen, S. S. McAllister and W. M. Jolly. On the need for a theory of wildland fire spread. International Journal of Wildland Fire 22(1):25–36, 2013.

[2] R. Y. Rubinstein. The cross-entropy method for combinatorial and continuous optimization. Methodology and Computing in Applied Probability 1(2):127–190, 1999.

[3] J. A. Nelder and R. Mead. A simplex method for function minimization. The Computer Journal 7(4):308–313, 1965.

Reviewer_1KkFOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces WUIRL, which simulates a wildfire in a wildland–urban interface (WUI) as a step-by-step decision process that moves through a neighborhood. A reinforcement learning policy (PPO) picks how fire reaches each building, while a fixed physics model decides whether that building actually catches fire. The goal is to reconstruct the hidden order of house-to-house fire spread, including effects from surrounding vegetation, since real records only show final damage and rough hourly timing, not how each building ignited. Tested on the 2025 Eaton fire, it recovers 78.9% of burned buildings while preserving 52.8% of survivors, and attains the highest reconstruction score among the Random and Greedy baselines, though Greedy recovers more burned buildings (87.4%) and has the highest burn F1 score.

Critical strengths
  • Clear, framing of a gap, where most prior work maps static building features to damage. This paper correctly argues that catching fire is a cascading process, and it affects how the fire spreads, which justifies the sequential formulation.
  • The approach maintains transparency by decoupling path selection from ignition modeling, where the RL policy identifies potential fire-propagation routes. A separate, physics-grounded mechanism determines building ignition. This makes the assumptions easier to examine than in a fully black-box model, although it does not validate the reconstructed paths.
  • The authors inspect different assumed spread routes and find that building heat and flying burning material have the largest effect on results. These are applied without re-training.
  • The paper is clear about several limitations, including candidate-pathway availability and training instability.
Critical weaknesses
  • The main reconstruction claim cannot be directly validated. The data show final damage and coarse timing, but not which source ignited each building. Many different ignition sequences could produce similar final damage patterns. Thus, the method may produce trajectories consistent with observations without recovering the true historical pathways.
  • The baselines are limited to Random and Greedy. Stronger comparisons such as non-RL search or optimization methods, or other domain-specific heuristic approaches, would help show whether PPO and sequential RL are necessary.
  • The paper should better justify its main metric. WUIRL has the best reconstruction score, but Greedy has higher burn recall and F1. The authors should explain why this score better represents the paper’s claims than burn recovery or F1.
  • Training appears unstable, based on three training seeds, and performance varies across policies and evaluation seeds. Performance declines with further training. Selected checkpoints at 50k–150k actions obtain 78.9% burn recall, while final 500k-action checkpoints obtain 22.2% recall with very high variance. The paper does not explain this degradation or show a stable training procedure. The main results therefore depend strongly on early checkpoint selection.
  • The evaluation and the generalizability of the method are limited because it uses only one fire dataset from Eaton. This is not a test of whether it works on unseen fires.

Minor comments

  • The text uses "reconstruction reward" and "reconstruction score" interchangeably.
  • Flux, and its notation q’’ are not intuitive to a reader unfamiliar with fire propagation.
  • The heat-related notations, like FTP, \phi_{i,c} (t) are very domain-specific and not explained for a non-domain reader.
  • Some key rules, like the ones governing outside fire entry, acceptance, and update timing, appear to be deferred to the appendix. A short, plain-language explanation in the main paper would make the method easier to judge.
Questions for authors

Please see the weaknesses above.

6 comments

We thank the reviewer for identifying the sequential approach taken, where we model wildfire spread through a community and separate pathway choice from physics-based ignition so that each assumption is open to inspection. WUIRL is, to our knowledge, the first method to reconstruct building-level fire spread scored against an hourly record, and it recovers the plausible ways the fire could have spread, each with a confidence. We have addressed each concern below with additional experiments. We hope these additions address the reviewer's concerns. We would appreciate reconsideration of the evaluation in light of them, and welcome continued discussion.

We agree that the method may produce multiple trajectories, and we do not claim the single true pathway. Instead, WUIRL generates the many fire histories that fit both the physics and the record, and estimates how likely each pathway is. Currently available fire data cannot confirm which neighbor ignited each building, for any model (Sec. 4.8, lines 453–459). Fires are recorded by final damage and, at best, hourly perimeters, and without a shared theory of spread no model's reliability can be confirmed [1]. In ecology, where the driving processes are also unobserved, models are still rejected by testing them against observed time series [2]. We take the same approach, and the experiments below test each part of WUIRL's claim that the record can check.

Experiment 1. Retraining without key mechanisms. We retrained WUIRL twice, once without building-to-building radiation and once without firebrands. Neither model reproduces Eaton. The reconstruction score (higher is better) falls to −5.52 and −5.47, compared with −4.39 when both mechanisms are present. The Eaton record therefore cannot be matched without these mechanisms, which limits how far the policy can compensate for missing physics. We will add this retraining to Section 4.4.

Experiment 2. Agreement across sampled histories. We ran the trained policy 30 times and recorded, for each building, how often it burned, from which neighbors, by which route and when. Of buildings burned in at least 27 of 30 runs, 72% were recorded as damaged in the inspection, against 48% for buildings burned in 4 to 26 runs and 23% for those burned in 3 or fewer. The share of runs in which a building burns therefore works as a confidence score, and high-confidence buildings match the inspection record far more often than low-confidence ones. The figure below summarizes the result, and we will add this analysis to the manuscript.

Share of buildings recorded as damaged, by how many of 30 sampled histories burn them-https://anonymous.4open.science/api/repo/WUIRL--C8EC/file/fig_agreement_simple.png

We will also clarify this in the discussion and limitations.

To address the reviewer's concern about limited baselines, we include six additional baselines across areas of machine learning, and outline the advantages, limitations and suitability of each below.

Rule-based policies fix the decision rule in advance. These are domain-specific heuristic approaches as suggested by the reviewer. Random and Greedy (Sec. 4.2, lines 286–291) are the paper's baselines. Nearest source and earliest source choose the source by distance or arrival order instead of heat. Select-all accepts every pathway, an upper bound on burn recall. The heat-flux threshold rule is Greedy with a heat-flux cutoff tuned on the score.

Derivative-free optimization tunes a fixed rule directly against the score. These are non-RL search or optimization methods as suggested by the reviewer. The cross-entropy method (CEM) optimizes a fixed weighting of 12 pathway features against the score, but cannot adapt to the fire's state [3]. Nelder–Mead searches the same weights locally [4].

Decision rule, all inside WUIRL's environmentReconstruction score
Rule-based policies
Random−4.582
Greedy−4.501
Nearest source−4.412
Earliest source−4.490
Select-all−4.507
Heat-flux threshold rule (tuned)−4.216
Derivative-free optimization
Nelder–Mead (tuned)−5.138
CEM (tuned)−4.140
Reinforcement learning (ours)
PPO−4.317

Why RL. Reconstruction is a cascade of decisions with no labels for how each building ignited, only delayed feedback from the record (Sec. 1, lines 48–55), which is the setting RL is designed for. PPO outperforms every untuned rule. The two methods tuned directly on the score lead it by 0.1–0.2 on the full fire, close to its run-to-run variation of ±0.11, but they search for one fixed rule, while PPO learns a policy that responds to the fire's state. The difference shows on regions held out of training.

OptimizerTraining regionsHeld-out regionsDrop
PPO (ours)−4.395−4.7330.34
CEM (tuned)−4.008−4.6970.69

Transfer. On held-out regions PPO loses half as much as CEM, so CEM's 0.39 lead shrinks to 0.04, and region by region the two are even (two regions each, one tied). The learned parts of the policy matter. Retrained with the learned receiver choice replaced by a uniform one, the score falls 0.16, and with the graph embeddings zeroed it falls 0.10.

Uncertainty. As the policy is stochastic, PPO is the only method that produces multiple candidate histories, which the uncertainty analysis in Comment 1 requires. It also recovers more burned buildings than either tuned rule (burn recall 85% vs 77–79%).

Framework. Our contribution is the methodology of physics-constrained wildfire reconstruction. We pose reconstruction as sequential, physically constrained pathway selection scored against the record, every method above runs inside it (Sec. 4.2, lines 288–289), and any objective can be used. We chose PPO because it is standard and robust, and because physical rollouts, rather than training, dominate compute (Sec. 4.8, lines 464–465). We will add the full table to the appendix.

The reconstruction score measures what the paper claims, namely how and when buildings burned. Burn recall, in contrast, rewards burning everything. WUIRL has the higher balanced accuracy (Sec. 4.3, Table 1, line 343), and Greedy's higher burn recall comes from burning more. We ran a policy that selects every offered pathway, and it matches Greedy's 87% burn recall but keeps only 40% of survivors. The score is class-balanced and scores timing (Sec. 3.3, Eq. 3, line 248), so it rewards getting both burned and surviving buildings right, at the right hour. Under the new default, WUIRL also matches Greedy's F1 (0.753 vs 0.751) while keeping more survivors. We will clarify the choice of metric in the manuscript and report burn F1 alongside the score.

The reviewer brings a decline in training to our attention (Sec. 4.7, lines 411–418), which has improved our work. We identified this instability as coming from the delayed final-damage term in the training reward, and reweighting that term stabilizes training. Only this training weight changed, chosen as the one setting stable at 500k actions rather than by its score, and the evaluation score is unchanged. The selected checkpoint performs on par with the paper's (reconstruction score −4.32 against −4.39), and the 500k checkpoint now scores −4.45 with 88% burn recall and ±0.056 variation across seeds, so the results no longer depend on early stopping. The results and conclusions of this paper remain unchanged, and our optimization's robustness is improved. We will update the manuscript with this default and add these runs to the appendix.

The Eaton Fire is what makes this work possible. To our knowledge, it is currently the only dataset that supports building-level reconstruction, with both an inspection of every structure and hourly progression across a whole community [5]. Damage inspections elsewhere record only final damage, and detailed timelines exist for only a few earlier fires [6, 7]. WUIRL is built for this kind of record, and because it learns each fire from its own records, it can be applied directly as future fires are recorded the same way. Within Eaton, the paper held out one region (Sec. 4.6, lines 400–409), and we now hold out each of five in turn, training without that region and scoring WUIRL on it. WUIRL's score falls half as far there as CEM's (reconstruction score −4.40 on training regions, −4.73 on held-out ones), and balanced accuracy moves from 0.65 to 0.60. We will add the five-region results to Section 4.6 and state the data limitation more clearly.

We have updated the manuscript with these important changes for consistency. We standardized to "reconstruction score" throughout and defined domain-specific terms at first use, including heat flux (q″), the flux–time product (FTP) that sets each building's ignition requirement, and normalized dose (Φ), which tracks progress toward ignition. To improve accessibility and context, we have included a short introduction to WUI fire spread and a plain-language summary of fire entry, acceptance and update timing in Section 3.

References

[1] M. A. Finney, J. D. Cohen, S. S. McAllister and W. M. Jolly. On the need for a theory of wildland fire spread. International Journal of Wildland Fire 22(1):25–36, 2013.

[2] C. Song and J. M. Levine. Rigorous validation of ecological models against empirical time series. Nature Ecology & Evolution 9:2265–2278, 2025.

[3] R. Y. Rubinstein. The cross-entropy method for combinatorial and continuous optimization. Methodology and Computing in Applied Probability 1(2):127–190, 1999.

[4] J. A. Nelder and R. Mead. A simplex method for function minimization. The Computer Journal 7(4):308–313, 1965.

[5] S. Kerber, D. Alkonis, D. Gorham and G. Horn. Southern California fires timeline report. UL Research Institutes, Fire Safety Research Institute, 2025.

[6] A. Maranghides et al. A case study of a community affected by the Waldo Fire. NIST Technical Note 1910, 2015.

[7] A. Maranghides et al. A case study of the Camp Fire: fire progression timeline. NIST Technical Note 2135, 2021.

Comments on the paper

1 comment

We thank the reviewers for their thoughtful feedback, and appreciate their unanimous acknowledgement of our novel approach, WUIRL, which models fire spread as a cascade of building-to-building ignitions decided by physics.

Our submission is in the applications to physical sciences track, utilizing machine learning to study wildfire progression. We would like to clarify two potential confusions present across the reviews to distinguish between limitations of the wildfire research field and limitations unique to our contribution:

1. "The main reconstruction claim cannot be directly validated" (1KkF, jzuD).

WUIRL reconstructs how fire spreads building by building through a wildland–urban interface (WUI), where homes meet wildland vegetation. Its claim is not the single true sequence of ignitions. Instead, it recovers the many possible ignition sequences consistent with the physics and the record, and estimates how likely each one is. Confirming any single sequence would require knowing which neighbor ignited each building, and no fire dataset contains this data (Sec. 4.8, lines 453–459). This limitation of the available data therefore applies to every model used for wildfire modeling. Wildfire science also lacks an agreed theory of fire spread against which models could be checked [1], so existing WUI models, which simulate spread from fixed relationships or predict final damage (Sec. 1, lines 48–53; Sec. 2, lines 105–116), are evaluated only against final damage or fire perimeters [2, 3]. Better records of how fires move through communities would benefit the whole field. We conduct specific experiments in the individual responses to verify our claim.

2. "It uses only one fire dataset from Eaton" (1KkF, ZcA6).

To our knowledge, the Eaton Fire dataset is the only dataset to date containing this information, with both an inspection of every structure and hourly progression across a whole community [4]. Damage inspections such as CAL FIRE's record final damage but not building-level timing, and only a few earlier fires have detailed timelines [5, 6]. As WUIRL learns each fire from its own records, it can be applied to any future fire recorded this way.

What WUIRL contributes. WUIRL is, to our knowledge, the first method to score building-level reconstructions against an hourly timeline, as ecology tests models against observed time series when the drivers go unobserved [7]. It returns many plausible fire histories that quantify pathway uncertainty, and it tests which physics the record supports. Its reconstructions will improve as records improve.

Summary of changes. Reviewers noted necessary additional baselines, training stability and ablations. We summarize the major changes here and include point-by-point responses for each reviewer's feedback under their respective reviews.

Baselines (1KkF, jzuD). We added six baselines, all run inside WUIRL's environment. PPO beats every untuned rule, and a tuned optimizer's (CEM) lead on the full fire shrinks to 0.04 on held-out regions.

Generalization (1KkF, jzuD, ZcA6). We held out each of five regions in turn. On held-out regions PPO's score drops 0.34 and CEM's 0.69, and region by region the two are even.

Uncertainty (1KkF, jzuD, ZcA6). We sampled 30 histories from the policy. Buildings that burn in most runs are far more often recorded as damaged (72%, against 23% for those that rarely burn), so agreement across runs works as a confidence score (figure below).

Training stability (1KkF, ZcA6). Reweighting the delayed final-damage term in training keeps training stable to 500k actions with the paper's results unchanged.

Physics and ablations (jzuD, ZcA6). Retraining without radiation or firebrands fails to reproduce Eaton, and gains hold under ±20% heat flux.

https://anonymous.4open.science/api/repo/WUIRL--C8EC/file/fig_agreement_simple.png

References

[1] M. A. Finney, J. D. Cohen, S. S. McAllister and W. M. Jolly. On the need for a theory of wildland fire spread. International Journal of Wildland Fire 22(1):25–36, 2013.

[2] A. Chulahwat et al. Integrated graph measures reveal survival likelihood for buildings in wildfire events. Scientific Reports 12:15954, 2022.

[3] D. M. J. Purnomo et al. Reconstructing modes of destruction in wildland–urban interface fires using a semi-physical level-set model. Proceedings of the Combustion Institute 40(1), 2024.

[4] S. Kerber, D. Alkonis, D. Gorham and G. Horn. Southern California fires timeline report. UL Research Institutes, Fire Safety Research Institute, 2025.

[5] A. Maranghides et al. A case study of a community affected by the Waldo Fire. NIST Technical Note 1910, 2015.

[6] A. Maranghides et al. A case study of the Camp Fire: fire progression timeline. NIST Technical Note 2135, 2021.

[7] C. Song and J. M. Levine. Rigorous validation of ecological models against empirical time series. Nature Ecology & Evolution 9:2265–2278, 2025.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.