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

arXiv: https://arxiv.org/abs/2606.13633 · HTML: https://mlsanity.com/paper/2606.13633/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 6 of its 62 references that are in the map.
- First posted on arXiv 11 Jun 2026.

## Builds on (most cited first, 6 of 6)

- [The CMA Evolution Strategy: A Tutorial](https://mlsanity.com/paper/1604.00772/) (2016)
- [A review of machine learning applications in wildfire science and management](https://mlsanity.com/paper/2003.00646/) (2020)
- [PyTorchFire: A GPU-Accelerated Wildfire Simulator with Differentiable Cellular Automata](https://mlsanity.com/paper/2502.18738/) (2025)
- [Analysis and design of model predictive control frameworks for dynamic operation -- An overview](https://mlsanity.com/paper/2307.03004/) (2023)
- [JaxWildfire: A GPU-Accelerated Wildfire Simulator for Reinforcement Learning](https://mlsanity.com/paper/2512.06102/) (2025)
- [Neural-Parameterized Cellular Automata for Wildfire Spread](https://mlsanity.com/paper/2606.11676/) (2026)

## Abstract (arXiv)

Aerial wildfire suppression requires decisions about when, where, and how to deploy limited aircraft. We present an intervention-design framework built on a frozen hybrid convolutional neural network and cellular automaton (CNN-CA) simulator trained jointly on six historical wildfires. First, we jointly optimize binary drop execution and continuous location and orientation, with aircraft-specific footprints, wind drift, and availability, turnaround, and grounded-day constraints. Second, we model water as an immediate transfer of burning probability to the unburned state and retardant as a persistent reduction in the fuel contribution to spread. Third, we remove drops by a rollout-verified backward elimination while limiting degradation in selected fire-performance metrics. Fourth, we evaluate fixed schedules under daily state sampling and a separate spatially correlated probability-field sensitivity test. Fifth, we compare the optimizer with random, tactical, greedy, and derivative-free planners under shared fleet, drop-budget, and simulator-evaluation allowances, and measure computational scaling. A 2020 Bear Fire case study considers two objectives: total fire-affected area and protected-region exposure. The two-stage total-area schedule reduces deterministic terminal extent by 89.5% relative to the simulator baseline with 1,111 drops. The nominal schedule is sensitive to small pose and effectiveness perturbations. The tactical heuristic is more economical at small evaluation allowances; the gradient planner achieves better objectives with more computation.