Wildland Fire Modeling

Wildfires in Korea are becoming larger and harder to predict. My research aims at a physics-based understanding and prediction of wildland fire behavior, combining HIGRAD/FIRETEC, FARSITE and WRF-SFIRE to cover scales from meter-scale fire–atmosphere interaction to national-scale fire spread. Alongside the models, I build fuel maps for accurate fire modeling: after a national-scale fuel map from forest-type data, I am now refining fuel and fuel moisture with satellite and high-resolution meteorological data. I also run large ensembles of high-resolution FIRETEC simulations to develop AI emulators that improve fire-spread accuracy in large-scale simulations.

Figure. Tree-resolving 3-D fuel field (left) and its inputs (canopy, tree list, fuel moisture and loading) built from the national fuel map for FIRETEC.

Slope, cross-wind and vegetation

How do slope steepness and a wind blowing across the slope combine to steer a fire? HIGRAD/FIRETEC simulations at Los Alamos National Laboratory covered six slopes (0–50%), four cross-slope wind speeds (4–10 m s⁻¹) and grass, chaparral and forest fuel beds. The propagation angle grows with steeper slopes and shrinks with stronger cross-slope winds, and secondary upslope and downwind branches appear that Rothermel-type models do not predict (Seo et al. 2026, Agric. For. Meteorol.).

Figure. Rate of spread for primary, main upslope and main downwind propagation versus wind speed and slope, for grass (top), chaparral (middle) and Blodgett (bottom).

Fuel maps and Korean wildfires

Fire models need fuel. I built a national-scale fuel map for South Korea from the forest-type map (임상도) and land-cover data, classified into the Anderson 13 fuel models (NFUEL_CAT) used by WRF-Fire and FARSITE, with matching 3-D canopy fuels for FIRETEC and QUIC-Fire. With it, I compare WRF-SFIRE, FARSITE and FIRETEC on Korean cases such as the 2025 Gyeongbuk fire, and I am refining fuel and fuel moisture with satellite and high-resolution meteorological data.

Figure. Multi-scale setup: (a) WRF nested domains, (b) Sentinel-2 view of the burn scar, (c) WRF fuel categories and (d) FARSITE fuel map with the FIRETEC 3-D domain.

FIRETEC ensembles and an AI emulator

To learn what controls early spread, I ran 12,480 FIRETEC simulations varying fuel, moisture, slope, wind direction and wind speed on a 2-m grid. The wind is substantially modified by fuel and terrain before the fire arrives, and wind direction reorganizes the shape of the fire front rather than simply adding upslope and cross-slope effects. These ensembles are the training data for an AI emulator that improves fire-spread accuracy in large-scale simulations.

Figure. Top: burn probability at 120 s for wind directions 90–120° across wind speeds (each panel averages up to 72 runs). Bottom: mean final burned fraction versus slope and wind speed for four fuel types, averaged over moisture and wind direction.

Wildfires in Mongolia

In September 2026 I taught a WRF-Fire tutorial at the National University of Mongolia and gave an invited talk on physics-based fire prediction. Using a Mongolian fuel map built from land-cover data and ERA5 meteorology, we simulated the April 2021 Shiliin Bogd grassland fire and compared it with FIRMS satellite fire detections; the same workflow was then used to simulate the October 2026 Khalkh Gol fire.

Figure. Left: fuel map, FIRMS fire detections and candidate WRF-Fire domains for the April 2021 Shiliin Bogd fire. Right: WRF-Fire simulation of the October 2026 Khalkh Gol fire, showing the burned area and plume envelope.