A Neural Network That Predicts Global Wildfires 7 Days in Advance

A new artificial intelligence model can predict where wildfires will burn across the globe up to a week in advance—accurately enough to improve air quality forecasts and public health warnings. In 2023, the system reduced prediction errors by 43% compared to current methods, even at the seven-day mark. This isn’t about detecting flames as they ignite. It’s about forecasting fire activity—where, when, and how intensely biomass will burn—before satellites can see it, using only recent fire history and environmental conditions. And it works.

That matters because today’s air quality models are flying blind. When smoke from Canadian wildfires darkened skies over New York and Chicago in 2023, the forecasts were reactive, not predictive. The reason? Operational systems like NOAA’s GEFS-Aerosols rely on satellite observations of fire radiative power (FRP)—a measure of how much energy a fire emits—as input. But those data arrive with a 36-hour delay. Worse, once used, they’re frozen for the duration of the forecast, assuming fires won’t grow, shift, or die. In reality, fires evolve. And so does their smoke.

"Current aerosol forecasts assume fire emissions are static over five to seven days," the researchers write. "This is clearly unrealistic."

The new model, developed by Li Zhang, Jun Wang, and colleagues, breaks that assumption. It uses a graph neural network (GNN) trained on three years of global fire, weather, and land data to predict daily FRP at 0.1° resolution—one fire pixel every 11 kilometers—up to seven days ahead. It doesn’t simulate combustion or fire spread physics. Instead, it learns the statistical fingerprints of fire behavior from data: how drought, wind, vegetation type, and recent burning history combine to make fires more or less likely in a given location.

And it learns them well. When tested on 2023–2024 data, the model reproduced the planet’s seasonal fire rhythms—Australia’s dry-season blazes, the Amazon’s August peak, Siberia’s summer flares—with striking fidelity. More importantly, it outperformed the standard "persistence" baseline—where tomorrow’s fire is assumed to look like today’s—by wide margins.

At one day ahead, the model cut mean squared error (MSE) by 32% in 2023 and 24% in 2024. By day seven, those improvements grew to 43% and 40%, respectively. That’s counterintuitive—usually, forecast skill degrades over time. Here, the relative advantage increases because persistence becomes increasingly wrong, while the GNN retains signal.

"The fact that we gain more skill relative to persistence at longer lead times suggests the model is capturing real drivers of fire evolution, not just short-term inertia," the authors note.

The Science

The model, called ST-FRPGNN (Spatiotemporal Fire Radiative Power Graph Neural Network), treats Earth’s surface as a dynamic graph. Each node represents a 0.1° grid cell—about 11 km × 11 km at the equator. Edges connect neighboring cells, allowing fire activity to propagate spatially, much like real flames spread across landscapes.

The input features include:

• Recent fire history (FRP from the past 7 days)
• Meteorological reanalysis data (temperature, humidity, wind speed, precipitation)
• Land cover type (forest, grassland, cropland, etc.)
• Vegetation greenness (from NDVI)
• Topography (elevation, slope)

The target? The next day’s FRP, observed from the GBBEPx satellite product—the same dataset used to initialize NOAA’s operational forecasts.

The architecture is built on a spatiotemporal GNN backbone, combining graph convolutional layers (to capture spatial dependencies) with temporal modules like GRUs (Gated Recurrent Units) to model fire evolution over time. The network is trained end-to-end to predict FRP one day ahead, then rolled forward autoregressively to generate seven-day forecasts.

Training used data from 2020 to 2022; evaluation was done on 2023–2024, a period that included extreme fire years in Canada and Australia. The model was not fine-tuned for specific regions or events—it learned global patterns from the start.

Crucially, the system operates at the same resolution and latency constraints as operational systems. It uses only data that would be available in real time, making it a drop-in candidate for integration into existing forecasting pipelines.

What They Found

The results are striking—not just in aggregate metrics, but in geographic and temporal fidelity.

At 0.1° resolution, the model reduced MSE by 32% at one-day lead and 43% at seven days in 2023 (

Mean Squared Error Reduction vs. Lead Time

Label	Value
1-day	32 %
2-day	35 %
3-day	37 %
4-day	39 %
5-day	41 %
6-day	42 %
7-day	43 %

Source: Zhang et al. (2026), evaluated on 2023 data

). In 2024, the reductions were 24% and 40%, respectively. These numbers reflect a consistent, significant improvement over persistence across all lead times.

$$\text{MSE Reduction}=1-\frac{{\text{MSE}}_{\text{GNN}}}{{\text{MSE}}_{\text{Persistence}}}$$

But MSE alone doesn’t tell the full story. Fires are spatially sparse—most grid cells are unburned on any given day. So the team also evaluated detection skill using the Critical Success Index (CSI), which balances hits, false alarms, and misses.

At a coarser but operationally relevant 1° resolution (~110 km), CSI ranged from 0.32 to 0.60 across lead times and regions (

Critical Success Index by Lead Time (1° resolution)

Label	Value
1-day	0.6 CSI
3-day	0.48 CSI
5-day	0.4 CSI
7-day	0.32 CSI

Source: Zhang et al. (2026), 1° resolution evaluation

). For context, a CSI of 0.3 is considered "useful" in meteorology; 0.6 is excellent. The model maintained strong detection skill even at day seven, with only modest decline.

"Detection is surprisingly stable," the researchers write. "It suggests that the large-scale environmental controls on fire occurrence—drought, wind, fuel availability—are predictable on sub-weekly timescales."

But intensity is harder. While the model reliably identifies where large fires will occur, it systematically underestimates their radiative power. At peak intensity, predicted FRP is about 60–70% of observed values. This bias grows with fire size.

"We detect the big fires, but we don’t capture their full energy," the authors admit. "This is a key limitation for aerosol forecasting, where emission strength scales nonlinearly with FRP."

Still, the spatial patterns are compelling.

shows predicted vs. observed FRP over Siberia in July 2023, a month of extreme fire activity. The model captures the broad swaths of burning across Yakutia, including the timing and location of major clusters. It even anticipates the northward shift of fire fronts as heatwaves moved poleward.

Similarly,

displays a forecast for the 2023 Canadian fire season. The model correctly predicts the outbreak of intense fires in central Saskatchewan and northern Alberta two days before they peaked—days before satellites could provide updated inputs to current air quality models.

Why This Changes Things

Today, when a megafire erupts in British Columbia or the Amazon, air quality alerts are issued only after smoke is already aloft. By then, it’s too late for schools to cancel recess, for hospitals to prepare for asthma surges, or for outdoor workers to get respirators.

This model changes that timeline. By predicting fire activity days in advance, it enables proactive smoke forecasting. That means:

• Earlier public health warnings for vulnerable populations
• Better aviation routing to avoid smoke-dense airspace
• More accurate solar energy forecasts (smoke reduces irradiance)
• Improved climate model initialization, since biomass burning is a major source of black carbon and aerosols

And because the model runs quickly on GPUs—each seven-day global forecast takes under two minutes—it could be integrated into operational systems like GEFS-Aerosols with minimal computational cost.

"This isn’t a replacement for physical fire models," the authors clarify. "It’s a bridge between now and the next satellite overpass. It fills the gap where we have no eyes in the sky."

The implications extend beyond air quality. Predictable fire activity means predictable carbon emissions. In 2023, wildfires released over 2 billion tons of CO₂ globally—equivalent to the annual emissions of 500 million cars. If we can forecast when and where that carbon will be released, we improve carbon accounting, climate modeling, and even carbon market integrity.

Consider the 2023 Canadian fires, which emitted an estimated 1.5 billion tons of CO₂. Had this model been operational, it might have flagged the risk weeks in advance, based on dry fuels and heatwave forecasts. That could have informed forest management, fire suppression resource allocation, and even carbon offset project risk assessments.

But the real win is in equity. Satellite data gaps hurt the Global South most. In Africa and Southeast Asia, where agricultural burning is widespread but monitoring is sparse, a 36-hour delay can mean missing entire fire cycles. A predictive model like this—one that doesn’t wait for a satellite pass—could democratize fire intelligence.

"In regions with limited real-time observation, predictability becomes a proxy for visibility," the researchers argue.

What's Next

The model isn’t perfect. Two challenges stand out.

First, intensity underestimation. Because aerosol emissions scale with FRP squared or higher, even a 30% low bias in FRP can lead to 50–70% underestimation of smoke. The authors suggest post-hoc calibration or hybrid approaches—using the GNN to predict fire location and likelihood, then coupling it with a physical model to simulate intensity.

Second, small fires. The model struggles with pinpoint accuracy for fires below 10 MW of FRP. That’s expected—small fires are more stochastic, often human-ignited, and harder to resolve at 11 km scale. But they matter. In Southeast Asia, small agricultural burns dominate emissions.

Future work could integrate higher-resolution data, human activity proxies (like nighttime lights), or active fire alerts from VIIRS to refine small-fire placement.

The team also plans to test the model in a coupled mode—feeding its FRP predictions directly into an aerosol transport model like GEFS-Aerosols—to measure downstream improvements in PM2.5 forecasts.

"The ultimate test," they write, "is not how well we predict FRP, but how much better we predict air quality."

There are broader lessons here. This work is part of a quiet revolution in Earth system science: the shift from purely physical models to hybrid systems that combine physics with machine learning. Weather models now use AI to downscale precipitation. Climate models use neural networks to emulate cloud processes. Now, fire prediction is joining the wave.

But unlike some AI models that operate as black boxes, this GNN offers interpretability. The graph structure mirrors real-world connectivity. Attention mechanisms can reveal which factors—drought, wind, prior burning—drive predictions in a given region. That transparency builds trust, especially in operational settings.

Still, caution is warranted. The model was trained on recent years, which may not represent future fire regimes under climate change. As fire seasons lengthen and megafires become more common, the statistical relationships the model learns could break down.

"We’re predicting based on the past," the authors warn. "But the future of fire may not look like the past."

For now, though, the model works. It shows that fire activity—often seen as chaotic and unpredictable—is, in fact, constrained by large-scale environmental drivers that are predictable. And that predictability, even if imperfect, is actionable.

In a world where wildfire smoke now routinely crosses continents, where a fire in Siberia can degrade air quality in Europe, the ability to see a few days ahead isn’t just useful. It’s essential.