pith. sign in

arxiv: 2606.11676 · v1 · pith:WCBKPZN7new · submitted 2026-06-10 · 💻 cs.CE · cs.LG· physics.comp-ph

Neural-Parameterized Cellular Automata for Wildfire Spread

Maksym Zhenirovskyy , Ion Matei , Rohit Vuppala , Takuya Kurihana , Hon Yung Wonga This is my paper

Pith reviewed 2026-06-27 08:05 UTC · model grok-4.3

classification 💻 cs.CE cs.LGphysics.comp-ph
keywords wildfire spread modelingcellular automataneural networksdata assimilationprobabilistic modelingfire perimeter forecastinghybrid physical models
0
0 comments X

The pith

A neural network dynamically parameterizes a cellular automata model to forecast wildfire spread with IoU exceeding 0.6 over 72-hour horizons after a 10-day assimilation period.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces a hybrid model that combines a multi-scale convolutional neural network with a probabilistic cellular automata framework to predict how wildfires spread. Traditional rigid-parameter models often underpredict fire growth, so this approach lets the network generate spatially varying parameters for spread probability, wind, and slope effects. The model is fitted incrementally to observed fire perimeters for ten days and then used to project future growth. Evaluations on six large western US wildfires show it maintains intersection-over-union scores above 0.6 for three-day forecasts. This produces conditional projections that reflect the suppression actions already present in the data rather than assuming no intervention.

Core claim

The central claim is that a Multi-Scale Convolutional Neural Network can be used to dynamically generate the parameters of a three-state probabilistic cellular automaton, allowing the model to capture nonlinear interactions in wildfire spread while remaining physically interpretable, and that after incremental fitting to observed perimeters over ten days the model achieves IoU greater than 0.6 in 72-hour forecasts on real wildfire events.

What carries the argument

The Multi-Scale Convolutional Neural Network that outputs spatially varying parameters for fire-spread probability, wind alignment, and slope influence within the probabilistic cellular automata rules.

If this is right

  • The forecasts represent conditional projections of fire growth under the suppression regime already encoded in the assimilation observations.
  • The hybrid design captures complex nonlinear environmental interactions while preserving the physical interpretability of the underlying three-state cellular automaton.
  • Performance is demonstrated across six large-scale wildfires in the western United States.
  • The approach supports gradient-based calibration of the network parameters.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the parameterization approach generalizes beyond the tested cases, similar neural control of cellular automata could be applied to other spreading processes such as epidemics or information diffusion.
  • Incorporating additional environmental inputs during the assimilation window might extend reliable forecast horizons past 72 hours.
  • Direct comparison against traditional static-parameter wildfire models on the same events would quantify the gain from dynamic parameter generation.

Load-bearing premise

That parameters fitted to ten days of observed perimeters enable the model to project fire growth forward rather than simply reproducing the patterns already seen in the assimilation data.

What would settle it

Observing IoU scores below 0.6 in 72-hour forecasts on additional wildfires after the same 10-day assimilation procedure would indicate that the fitted parameters do not support reliable forward projection.

Figures

Figures reproduced from arXiv: 2606.11676 by Hon Yung Wonga, Ion Matei, Maksym Zhenirovskyy, Rohit Vuppala, Takuya Kurihana.

Figure 1
Figure 1. Figure 1: Canopy Fuel Mask vs Burned Area. The charts show that the standard Canopy Cover (CC) and Canopy Bulk Density (CBD) do not capture the full extent of the burn. model is positioned at the intersection of these two tradi￾tions: it retains the three-state probabilistic CA substrate so that mass conservation and ignition dynamics remain physically transparent, while replacing the historically static parameter s… view at source ↗
Figure 3
Figure 3. Figure 3: Detailed architectural diagram of the hybrid wildfire model combining a MS-CNN parameter generator with a Probabilistic CA physics engine. where the weight map is defined as 𝑤(𝑥, 𝑦) = 𝑤𝑓 for frontier pixels (i.e., for 𝐹(𝑥, 𝑦) > 0.5), and 𝑤(𝑥, 𝑦) = 1 elsewhere. The hyperparameter 𝑤𝑓 (set to 10 in our exper￾iments) amplifies the gradient signal at the fire perimeter, steering the model toward accurate bounda… view at source ↗
Figure 4
Figure 4. Figure 4: Per-event temporal trajectories of four spatial-agreement metrics over the 20-day analysis horizon for all six events: (a) Intersection over Union (𝐼𝑜𝑈), (b) Manhattan Distance, (c) Precision, and (d) Recall. The vertical dashed line at Day 10 marks the end of the data-assimilation window; values to the right of this line correspond to forward forecasts produced with frozen parameters. All metrics are comp… view at source ↗
Figure 5
Figure 5. Figure 5: Bear 2020 fire analysis: (a) Days 1–7. Layout and color conventions as defined at the start of Section 5.2 [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Chimney 2016 fire analysis: (a) Days 1–7. Layout and color conventions as defined at the start of Section 5.2 [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Pier 2017 fire analysis: (a) Days 1–7. Layout and color conventions as defined at the start of Section 5.2 [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Brattain 2020 fire analysis: (a) Days 1–7. Layout and color conventions as defined at the start of Section 5.2 [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 8
Figure 8. Figure 8: (cont.) Brattain 2020. Days 15–20 [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Ferguson 2018 fire analysis: (a) Days 1–7. Layout and color conventions as defined at the start of Section 5.2. implicitly learned from the other five events ( [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 9
Figure 9. Figure 9: (cont.) Ferguson 2018. Days 8–14 [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Buck 2017 fire analysis: (a) Days 1–7. Layout and color conventions as defined at the start of Section 5.2 [PITH_FULL_IMAGE:figures/full_fig_p015_10.png] view at source ↗
read the original abstract

Traditional wildfire models rely on rigid, low-dimensional parameters and static fuel maps, frequently underpredicting fire spread. To address this weakness, we introduce a hybrid deep-learning parameterized Probabilistic Cellular Automata (CA) framework implemented in JAX. Our approach employs a Multi-Scale Convolutional Neural Network to dynamically generate spatially varying parameters that govern fire-spread probability, wind alignment, and slope influence. This hybrid design captures complex, nonlinear environmental interactions while preserving the physical interpretability of the underlying three-state CA. The JAX implementation enables hardware acceleration and gradient-based parameter calibration. Evaluated on six large-scale wildfires in the western United States, the model maintains IoU > 0.6 over 72-hour forecast horizons after a 10-day data assimilation window during which the model is fitted incrementally to observed perimeters; the resulting forecast is a conditional projection of fire growth under the suppression regime already ncoded in those observations.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper introduces a hybrid Probabilistic Cellular Automata model for wildfire spread in which a multi-scale CNN dynamically generates spatially varying parameters (spread probability, wind alignment, slope) from environmental inputs. Implemented in JAX to support gradient-based incremental calibration, the model is evaluated on six large western-US wildfires, reporting IoU > 0.6 on 72-hour forecast horizons after a 10-day assimilation window in which the CNN is fitted to observed perimeters; the forecasts are explicitly described as conditional projections under the suppression regime already present in the assimilation data.

Significance. If the hybrid parameterization demonstrably improves upon rigid traditional CA models while preserving physical interpretability and enabling reproducible JAX-based calibration, the work could contribute to data-assimilative wildfire modeling. The explicit acknowledgment that outputs are conditional projections under encoded suppression is a strength in transparency, but it also caps the significance for claims of genuine forward predictive power beyond pattern continuation.

major comments (2)
  1. [Abstract] Abstract: the central performance claim (IoU > 0.6 over 72 h after 10-day incremental fitting) is framed as a 'conditional projection of fire growth under the suppression regime already encoded in those observations.' This directly implies that the reported metric largely measures consistency with patterns already present in the assimilation window rather than extrapolation via the learned CA rules, undermining the claim that the hybrid structure supports independent forward simulation.
  2. [Evaluation section (implied by abstract)] The incremental gradient-based calibration of the CNN weights on observed perimeters (described in the evaluation) risks the fitted parameters simply encoding the realized trajectory and implicit suppression effects rather than generalizable physical rules; without reported baselines, ablation of the CNN component, or tests that hold out suppression information, it is unclear whether the hybrid model adds predictive value beyond direct perimeter continuation.
minor comments (1)
  1. [Abstract] Abstract contains a typo: 'already ncoded' should be 'already encoded'.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major comment point by point below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central performance claim (IoU > 0.6 over 72 h after 10-day incremental fitting) is framed as a 'conditional projection of fire growth under the suppression regime already encoded in those observations.' This directly implies that the reported metric largely measures consistency with patterns already present in the assimilation window rather than extrapolation via the learned CA rules, undermining the claim that the hybrid structure supports independent forward simulation.

    Authors: The manuscript explicitly frames the forecasts as conditional projections under the observed suppression regime to ensure transparency and avoid overclaiming predictive power. We do not assert independent forward simulation detached from the assimilation data. The hybrid CNN-CA design instead enables dynamic parameterization from environmental inputs within this data-assimilative context, preserving CA interpretability while supporting gradient-based calibration. This aligns with the referee's note on transparency as a strength for data-assimilative wildfire modeling. revision: no

  2. Referee: [Evaluation section (implied by abstract)] The incremental gradient-based calibration of the CNN weights on observed perimeters (described in the evaluation) risks the fitted parameters simply encoding the realized trajectory and implicit suppression effects rather than generalizable physical rules; without reported baselines, ablation of the CNN component, or tests that hold out suppression information, it is unclear whether the hybrid model adds predictive value beyond direct perimeter continuation.

    Authors: The CNN component maps from environmental inputs (e.g., wind, slope, fuel) at multiple scales to generate CA parameters, rather than fitting directly to perimeters, which constrains the model via the underlying physical CA rules. The JAX implementation facilitates this incremental adaptation. We will add baseline comparisons to traditional rigid CA models and an ablation removing the multi-scale CNN in the revised manuscript to quantify added value. Tests that explicitly hold out suppression information are not feasible with standard wildfire perimeter datasets, which embed suppression implicitly; we will expand the discussion of this limitation. revision: partial

standing simulated objections not resolved
  • Tests holding out suppression information, due to the absence of wildfire datasets that explicitly separate suppression effects from perimeter observations.

Circularity Check

1 steps flagged

Fitting to 10-day assimilation perimeters makes 72h forecast a conditional continuation under encoded regime

specific steps
  1. fitted input called prediction [Abstract]
    "the model maintains IoU > 0.6 over 72-hour forecast horizons after a 10-day data assimilation window during which the model is fitted incrementally to observed perimeters; the resulting forecast is a conditional projection of fire growth under the suppression regime already encoded in those observations."

    The 72h output is generated after gradient-based fitting to the exact assimilation perimeters and is defined as conditional on the regime encoded in those same observations, so the reported performance metric is statistically forced to reflect continuation of the fitted trajectory rather than extrapolation from independent parameters.

full rationale

The paper's central evaluation result reduces to a fitted-input-called-prediction pattern. The model is incrementally fitted to observed perimeters over a 10-day window, after which the 72-hour output is reported as a forecast but is explicitly framed as a conditional projection under the suppression regime already present in the assimilation data. Because the fitting objective directly minimizes mismatch to the specific observed evolution, the IoU metric largely measures reproduction of patterns already encoded during assimilation rather than independent forward simulation via the CA rules. No self-citation chains, self-definitional equations, or other load-bearing reductions are present in the provided text; the hybrid CA+NN structure itself is not shown to be circular.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Based solely on the abstract, the central claim rests on the sufficiency of a three-state CA, the ability of a multi-scale CNN to capture nonlinear interactions, and the validity of data assimilation for producing usable parameters.

free parameters (1)
  • CNN weights for parameter generation
    The multi-scale convolutional neural network parameters are learned incrementally during the 10-day assimilation window to produce spatially varying CA parameters.
axioms (1)
  • domain assumption A three-state probabilistic cellular automata is sufficient to represent wildfire spread dynamics
    The framework is built on an underlying three-state CA whose parameters are modulated by the neural network.

pith-pipeline@v0.9.1-grok · 5702 in / 1331 out tokens · 19562 ms · 2026-06-27T08:05:02.033766+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Aerial Wildfire Suppression Planning with a Hybrid CNN-Cellular Automata Fire Model

    eess.SY 2026-06 unverdicted novelty 5.0

    A hybrid neural-cellular automaton wildfire model is paired with gradient-based optimization of aerial drops to generate suppression plans that reduce fire-affected area while quantifying uncertainty.

Reference graph

Works this paper leans on

16 extracted references · 8 canonical work pages · cited by 1 Pith paper

  1. [1]

    doi:10.1016/j.amc.2008.06.046

    Applied Mathematics and Computation 204, 191–201. doi:10.1016/j.amc.2008.06.046. Anastasiou, N., Kondylatos, S., Papoutsis, I.,

    work page doi:10.1016/j.amc.2008.06.046 2008

  2. [2]

    arXiv preprint arXiv:2505.17556 URL: https://arxiv.org/abs/2505.17556

    Wildfire spread forecasting with deep learning. arXiv preprint arXiv:2505.17556 URL: https://arxiv.org/abs/2505.17556. Bradbury, J., Frostig, R., Hawkins, P., Johnson, M.J., Leary, C., Maclau- rin, D., Necula, G., Paszke, A., VanderPlas, J., Wanderman-Milne, S., Zhang,Q.,2018. JAX:ComposabletransformationsofPython+NumPy programs. URL:http://github.com/jax...

    Pith/arXiv arXiv 2018

  3. [3]

    arXiv preprint arXiv:2512.06102 URL:https://arxiv.org/abs/2512.06102

    JaxWildfire: A GPU-accelerated wildfire simulator for reinforcement learning. arXiv preprint arXiv:2512.06102 URL:https://arxiv.org/abs/2512.06102. California Department of Forestry and Fire Protection,

    arXiv

  4. [4]

    California Department of Forestry and Fire Protection,

    Buck fire in- cident report.https://www.fire.ca.gov/incidents/2017/9/12/buck-fire. California Department of Forestry and Fire Protection,

    2017

  5. [5]

    Chen,X.,Tian,Y.,Zheng,C.,Liu,X.,2024

    2020 fire season incident archive.https://www.fire.ca.gov/incidents/2020. Chen,X.,Tian,Y.,Zheng,C.,Liu,X.,2024. AutoST-Net:Aspatiotemporal feature-driven approach for accurate forest fire spread prediction from remote sensing data. Forests 15,

    2020

  6. [6]

    Finney, M.A.,

    doi:10.3390/f15040705. Finney, M.A.,

    work page doi:10.3390/f15040705

  7. [7]

    Natu- ral Hazards and Earth System Sciences 19, 169–179

    Using cellular automata to sim- ulate wildfire propagation and to assist in fire management. Natu- ral Hazards and Earth System Sciences 19, 169–179. doi:10.5194/ nhess-19-169-2019. Jain, P., Coogan, S.C.P., Subramanian, S.G., Crowley, M., Taylor, S., Flannigan, M.D.,

    2019

  8. [8]

    EnvironmentalReviews28,478–505

    A review of machine learning applications in wildfirescienceandmanagement. EnvironmentalReviews28,478–505. doi:10.1139/er-2020-0019. Lahrichi,S.,Bova,J.,Johnson,J.,Malof,J.,2025. Improvedwildfirespread prediction with time-series data and the WSTS+ benchmark. arXiv preprint arXiv:2502.12003 URL:https://arxiv.org/abs/2502.12003. Mathur, S., Manjunath, S.B....

    work page doi:10.1139/er-2020-0019 2020

  9. [9]

    MuñozSabater,J.,2019

    Spa- tiotemporal wildfire prediction and reinforcement learning for helitack suppression.arXivpreprintarXiv:2601.14238URL:https://arxiv.org/ abs/2601.14238. MuñozSabater,J.,2019. ERA5-Landmonthlyaverageddatafrom1981to present. Copernicus Climate Change Service (C3S) Climate Data Store (CDS). doi:10.24381/cds.68d2bb30. Rothermel, R.C.,

    work page doi:10.24381/cds.68d2bb30 2019

  10. [10]

    NatureReviewsEarth&Environment 4, 552–567

    Differentiable modelling to unify machine learning and physicalmodelsforgeosciences. NatureReviewsEarth&Environment 4, 552–567. doi:10.1038/s43017-023-00450-9. Weinhouse, C., Augustin, J.,

    work page doi:10.1038/s43017-023-00450-9

  11. [11]

    arXiv preprint arXiv:2510.09708 URL:https://arxiv.org/abs/2510.09708

    Leveraging cellular automata for real-time wildfire spread modeling in California. arXiv preprint arXiv:2510.09708 URL:https://arxiv.org/abs/2510.09708. Xia, Z., Cheng, S.,

    arXiv

  12. [12]

    MendeleyData, V1

    Data for: PyTorchFire: A GPU-accelerated wildfiresimulatorwithdifferentiablecellularautomata. MendeleyData, V1. doi:10.17632/nx2wsksp9k.1. dataset. Xia, Z., Cheng, S.,

    work page doi:10.17632/nx2wsksp9k.1

  13. [13]

    Environmental Modelling & Software 188, 106401

    PyTorchFire: A GPU-accelerated wildfire simu- lator with differentiable cellular automata. Environmental Modelling & Software 188, 106401. doi:10.1016/j.envsoft.2025.106401. Xu,F.,Gong,W.,Wu,H.,Peng,L.,Zhao,X.,etal.,2026. Advancedglobal wildfire activity modeling with hierarchical graph ODE. arXiv preprint arXiv:2601.01501 URL:https://arxiv.org/abs/2601.0...

    work page doi:10.1016/j.envsoft.2025.106401 2025

  14. [14]

    arXiv preprint arXiv:2507.00761 URL: https://arxiv.org/abs/2507.00761

    Aprobabilisticapproachtowildfirespreadpredictionusingadenoising diffusion surrogate model. arXiv preprint arXiv:2507.00761 URL: https://arxiv.org/abs/2507.00761. Zheng, Z., Huang, W., Li, S., Zeng, Y.,

    arXiv

  15. [15]

    Eco- logical Modelling 348, 33–43

    Forest fire spread simulating model using cellular automaton with extreme learning machine. Eco- logical Modelling 348, 33–43. doi:10.1016/j.ecolmodel.2016.12.022. Zhou, N., Wang, H., Li, J., Li, Y., Zhang, X.P., Li, Y., Chen, X.,

    work page doi:10.1016/j.ecolmodel.2016.12.022 2016

  16. [16]

    arXiv preprint arXiv:2512.17152 URL:https:// arxiv.org/abs/2512.17152

    PhysFire-WM: A physics-informed world model for emulating fire spread dynamics. arXiv preprint arXiv:2512.17152 URL:https:// arxiv.org/abs/2512.17152. M. Zhenirovskyy et al.:Preprint submitted to ElsevierPage 16 of 16

    arXiv