Neural-Parameterized Cellular Automata for Wildfire Spread

This paper introduces a hybrid deep-learning framework that uses a Multi-Scale Convolutional Neural Network to dynamically parameterize a Probabilistic Cellular Automata model in JAX, significantly improving wildfire spread prediction accuracy on large-scale US fires by capturing complex environmental interactions while maintaining physical interpretability.

The Gist

Imagine trying to predict how a wildfire will spread across a landscape. Traditionally, scientists have used rigid, rule-based maps that say, "Fire can only burn here because there are trees," and "Fire cannot burn there because it's just grass or dirt." The problem, as this paper points out, is that nature doesn't follow these strict rules. Real fires often jump over "non-burnable" areas like grasslands or even water bodies due to flying embers, intense heat, or wind, leaving huge gaps between what the old maps say should burn and what actually burns.

The authors, a team from Fujitsu Research, have built a new kind of wildfire simulator that fixes this by mixing old-school physics with modern AI. Here is how their system works, explained simply:

1. The Old Way vs. The New Way

Think of the old models like a stiff, pre-written script. They have a fixed set of rules (like "fire spreads 10% faster on a hill") that apply everywhere, regardless of the specific weather or terrain at that exact moment. If the map says an area has no trees, the script stops the fire dead in its tracks, even if the fire actually jumped there.

The new model is like a smart, improvising director. It still uses a basic set of physical rules (the "script"), but it has a "smart assistant" (a Neural Network) that watches the landscape and rewrites the rules in real-time. Instead of saying "fire spreads 10% faster," the assistant says, "In this specific patch of grass, with this specific wind, the fire should spread 40% faster."

2. The "Brain" of the System (The Neural Network)

The core of their invention is a Multi-Scale Convolutional Neural Network (MS-CNN). You can think of this as a pair of glasses with three different lenses:

• Lens 1: Looks at the big picture (7x7 grid) to see general terrain and weather.
• Lens 2: Looks at the medium picture (5x5 grid).
• Lens 3: Looks at the fine details (3x3 grid).

By looking at the landscape through these different "lenses" simultaneously, the AI learns to generate a unique set of instructions for every single square inch of the map. It creates a dynamic "fuel factor" that tells the fire engine, "Even though this looks like non-burnable grass on the map, the heat and wind here make it act like fuel." This allows the model to predict fires spreading into areas that traditional maps claim are safe.

3. The "Engine" (Cellular Automata)

The actual fire spreading happens in a grid of cells (like a giant checkerboard), which the authors call a Cellular Automata (CA).

• The States: Each square on the board is either Unburned, Burning, or Burned.
• The Physics: The fire moves from a burning square to its neighbors based on probability. If the wind is blowing toward a neighbor, the chance of it catching fire goes up. If the neighbor is on a steep hill, the chance goes up.
• The Innovation: In the past, these probabilities were static numbers. In this new system, the "Brain" (the AI) constantly updates these probabilities based on the local environment.

4. Learning from Mistakes (Training)

The system doesn't just guess; it learns. The researchers fed it data from six massive wildfires in the western US (mostly in California, plus one in Oregon).

• The Process: They let the model watch the fire for the first 10 days. During this time, the AI adjusted its internal "knobs" to match the real fire's path as closely as possible.
• The Prediction: After 10 days, they froze the AI's settings and asked it to predict the next 10 days.
• The Result: The model successfully predicted the fire's path with high accuracy (over 60% overlap with the real fire) for most events, even in areas where the fire burned through "non-burnable" zones.

5. Why It Matters (The "Fuel Factor")

The most significant breakthrough is how the model handles the "Canopy Fuel Mask." Traditional models look at satellite data and say, "No trees here, so no fire."

• The Reality: In the "Brattain Fire" of 2020, 65% of the fire burned in areas the map said had no trees.
• The Solution: The new model learned a "Fuel Factor" that isn't just about trees. It learned that wind, heat, and ground cover can make anything burn. It effectively learned to ignore the "No Burning" signs on the map when the physics of the situation demanded it.

6. Where It Stumbles

The paper is honest about where the system fails:

• New Ignitions: If a fire suddenly starts in a completely new spot (a "secondary ignition") far away from the main fire, the model misses it. The model only knows how to spread fire from existing fire, not how to create new fires out of thin air.
• Different Firefighting Styles: The model was trained on fires where firefighters were aggressively trying to stop the blaze. When tested on a fire in a wilderness area where firefighters used a "let it burn" or passive strategy, the model predicted the fire would spread faster than it actually did. It learned the "aggressive suppression" pattern from the training data and couldn't adapt to the "passive" approach.

Summary

This paper presents a hybrid tool that combines the reliability of physics-based rules with the adaptability of deep learning. It acts like a smart director that rewrites the rules of fire spread every second based on the local terrain and weather, allowing it to predict wildfires more accurately than ever before, especially in the tricky areas where traditional maps fail. It is built using JAX, a software framework that makes these complex calculations run very fast on modern computer hardware.

Technical Summary

Technical Summary: Neural-Parameterized Cellular Automata for Wildfire Spread

Problem Statement
Traditional wildfire forecasting models, such as the Rothermel equations and operational simulators like FARSITE, rely on rigid, low-dimensional parameters and static fuel maps. These approaches frequently underpredict fire spread, particularly in complex topographies or when fires propagate through areas classified as lacking combustible vegetation (e.g., grasslands or shrublands outside mapped canopy cover). Furthermore, existing semi-empirical models are not differentiable, making calibration to observed fire perimeters dependent on manual tuning or computationally expensive outer-loop data assimilation. Conversely, purely data-driven deep learning approaches often abandon the explicit state-transition structure of cellular automata (CA), sacrificing physical interpretability for representational flexibility. A critical systemic limitation shared by both paradigms is the reliance on static vegetation metrics (e.g., Canopy Cover and Bulk Density) to define burnable boundaries, which observational evidence shows leads to significant underprediction of fire extent.

Methodology
The authors propose a hybrid deep-learning parameterized Probabilistic Cellular Automata (CA) framework implemented in JAX to address these limitations. The model integrates a physics engine with a neural parameter generator, creating an end-to-end differentiable pipeline.

• Architecture: The core is a three-state Probabilistic CA (unburned, burning, burned) governed by local transition rules. Unlike previous differentiable CA works that use global scalar parameters, this model employs a Multi-Scale Convolutional Neural Network (MS-CNN) to dynamically generate dense, spatially varying parameter fields.
• Neural Parameter Generation: The MS-CNN processes a 13-channel input tensor comprising static LANDFIRE data (elevation, slope, aspect, canopy metrics, fuel models), dynamic wind data (ERA5), and a learnable fuel embedding. The network outputs six spatially varying fields: base spread probability ($p_{base}$), wind speed scaling ($\alpha_{w1}$), wind direction alignment ($\alpha_{w2}$), slope influence ($\alpha_s$), ignition rate ($\gamma$), and a unified $fuel\_factor$.
• Learnable Fuel Embedding: A key innovation replaces hard-coded canopy burnability masks with a learnable continuous fuel-embedding matrix. This allows the model to infer propagation potential in regions traditionally classified as non-burnable, effectively capturing sub-grid processes like ember spotting and thermal preheating that standard fuel maps miss.
• Physics Engine: The CA engine computes directional propagation potential ($\phi$) based on wind and slope modifiers, aggregates cumulative heat exposure ($\lambda$), and updates cell states using Poisson-aggregated ignition probabilities. The system operates in 24-hour macro-steps, each containing 50 micro-steps for finer temporal resolution.
• Training Objective: The model is trained using a composite loss function comprising five terms: Frontier-Weighted Binary Cross-Entropy (emphasizing perimeter accuracy), Mean Squared Error, Area Constraint (ensuring total burned area conservation), Soft Intersection-over-Union (IoU) for geometric shape fidelity, and Pooled MSE for coarse-grained spatial matching.

Key Contributions

1. Spatially Varying Parameters: The replacement of global scalar parameters with dense, spatially varying fields generated by an MS-CNN, allowing the model to capture heterogeneous environmental interactions.
2. Learnable Fuel Embedding: The introduction of a learnable $fuel\_factor$ that supersedes static canopy-derived burnability masks, enabling accurate prediction of fire spread in areas lacking mapped canopy fuel.
3. Joint Multi-Event Learning: Unlike prior works requiring independent per-event calibration, this model learns a single set of network weights and embedding parameters jointly across six diverse wildfire events, demonstrating generalizable environmental representations.
4. Differentiable Framework: The implementation in JAX enables hardware acceleration and gradient-based parameter calibration, facilitating efficient data assimilation.

Results
The model was evaluated on six large-scale wildfires in the western United States (five in California, one in Oregon). The evaluation protocol involved a 10-day data assimilation window where the model was fitted to observed perimeters, followed by a 72-hour (and up to 20-day) forward forecast with frozen parameters.

• Performance: The model maintained an Intersection over Union (IoU) greater than 0.6 over 72-hour forecast horizons for five of the six events after the calibration period.
• Comparative Advantage: The model outperformed the differentiable CA baseline by Xia and Cheng (2025), which reported peak IoU values of 0.408 and 0.525 on specific events, whereas the proposed model sustained IoU > 0.6 across the full 20-day horizon for those same events.
• Specific Event Analysis:
  • Bear 2020 & Chimney 2016: Achieved high IoU (0.74 and 0.66, respectively), with Chimney demonstrating the model's ability to predict spread in areas outside mapped fuel boundaries (41.3% of the fire burned outside vegetation).
  • Brattain 2020: Successfully predicted spread in a sagebrush-juniper ecosystem in Oregon with 65.3% of the fire footprint occurring outside mapped canopy vegetation.
  • Ferguson 2018: Showed lower initial performance (IoU ~0.15) due to a secondary, independent ignition not modeled by the contagious CA, but recovered to IoU > 0.6 once the fronts merged.
  • Buck 2017: Exhibited reduced accuracy (IoU 0.61) attributed to a mismatch in suppression regimes; the fire was managed with passive containment tactics in a wilderness area, diverging from the active suppression patterns implicitly learned from the other training events.

Significance and Claims
The paper claims that this hybrid architecture successfully bridges the gap between physical interpretability and data-driven flexibility. By retaining the three-state probabilistic CA substrate, the model preserves mass conservation and ignition dynamics while leveraging neural networks to infer complex, nonlinear environmental interactions. The authors assert that the system addresses the systemic failure of static fuel masks, enabling the prediction of fire growth in regions previously deemed non-burnable by standard metrics.

The authors explicitly note limitations: the model does not explicitly simulate ember spotting beyond the Moore neighborhood, spontaneous secondary ignitions, or active suppression tactics. Consequently, the learned $fuel\_factor$ acts as an effective quantity that conflates true fuel propagation with the imprint of these unmodeled processes. The evaluation is geographically concentrated in California and adjacent Oregon, and while the methodology is geographically agnostic, cross-regional transfer to boreal or tropical biomes remains untested. The paper positions this work as a step toward operational disaster management systems but refrains from claiming universal applicability without further cross-regional validation.