Yutong Zhang, Xiao Liu
H. Milton Stewart School of Industrial and Systems Engineering,
Georgia Tech · 2026
Use High Resolution Rapid Refresh (HRRR) data to train a generative machine learning model to achieve the following goals:
Guided Generation: generate high-resolution AOD (aerosol optical depth) fields in space and time, guided by "imperfect" remote-sensing data streams.
Forward Forecasting: use the trained generative model as a fast forward simulation tool, with user specified initial and weather conditions.
Inverse Modeling: use the trained generative model as an inverse modeling tool to detect the source of smoke or weather conditions.
Details are available from NOAA: https://gsl.noaa.gov/hrrr/
Led by scientists at NOAA’s Global Systems Laboratory (GSL), the High-Resolution Rapid Refresh (HRRR) has been the National Weather Service’s flagship short-term, high-resolution weather model for over a decade. HRRR-NCEP-smoke is an operational wildfire smoke forecasting system developed by NOAA's Global Systems Laboratory (GSL) and operated by the National Centers for Environmental Prediction (NCEP). It extends NOAA's High-Resolution Rapid Refresh (HRRR) weather prediction model to simulate and forecast the transport, dispersion, and concentration of wildfire smoke across the United States.
Horizontal resolution: 3km; Forecast frequency: Hourly; Forecast horizon: 18 hours; Satellite inputs: VIIRS and MODIS; Output: 3D smoke concentration and surface smoke
[*Figure: Los Angeles fires · 2025-01-09 · 18 UTC to 03 UTC, one frame per hour]
It provides five major capabilities:
Wildfire detection and emissions estimation: Uses satellite-derived Fire Radiative Power (FRP) observations to estimate wildfire locations and smoke emissions.
Smoke plume modeling: Simulates plume rise and the three-dimensional evolution of smoke.
Smoke transport and dispersion: Predicts how atmospheric winds, turbulence, and weather conditions transport smoke.
Surface air quality forecasting: Predicts near-surface smoke concentrations, particularly smoke-related PM2.5.
Visibility forecasting: Estimates the impact of smoke on atmospheric visibility.
GOES-18 Aerosol Optical Depth (AOD) is a satellite-derived measurement that quantifies how much airborne aerosol particles—such as wildfire smoke, dust, and pollution—reduce the transmission of sunlight through the atmosphere. It estimates how much smoke is in the air column, at 2 km resolution.
[*Figure: Los Angeles fires · 2025-01-09 · 18 UTC to 03 UTC, one frame per hour]
Limitations:
No observations at night.
Missing data subject to noise
A conditional flow-matching U-Net
Two past smoke frames and 55 condition channels (HRRR weather, GOES fire detections, metadata) enter a 58.8 M-parameter U-Net trained with flow matching on HRRR-Smoke only. It outputs the next hour, which is fed back for autoregressive rollout.
The pre-trained model is used to learn the dynamics of smoke propagation process.
Use real-time satellite data (GOES 18 AOD) data as the guidance, and generate high-resolution and high-quality AOD process
California
2024-07-26 · 18–22 UTC
Oregon
2024-07-26 · 18–22 UTC
EXPERIMENT II(a)
use the pre-trained model as a forecasting tool: ten hours ahead forecast
California
Observed to 22 UTC · forecast 23–08 UTC
Oregon
Observed to 22 UTC · forecast 23–08 UTC
EXPERIMENT II(b)
use the pre-trained model as a forward simulation tool (add fires and generate smoke)
Source on the left
0 / 1 / 3 / 16 fire cells · 10 h
Source on the right
0 / 1 / 3 / 16 fire cells · 10 h
EXPERIMENT II(c)
use the pre-trained model as a forward simulation tool (add wind field and generate smoke)
Source on the left
Wind left / right / up / down · 10 h
Source on the right
Wind left / right / up / down · 10 h
EXPERIMENT II(d)
use the pre-trained model as a forward simulation tool (add precipitation and generate smoke)
Scene 1
0 / 1 / 5 mm per hour
Scene 2
0 / 1 / 5 mm per hour
Scene 3
0 / 1 / 5 mm per hour
HOW THE INVERSION WORKS
The trained model is never changed. The fire map or the rain rate becomes the unknown c. For a given smoke sequence, a condition is good if the model’s denoising direction agrees with it:
L(c) = 𝔼τ,ε ‖ v̂(zτ, τ, c) − (x1 − ε) ‖²
Then move c downhill: c ← c − η∇L(c), a dozen steps. On GOES cases the error is computed on real observed pixels only.
This uses the local flow-matching error, not the full generative trajectory, so the results are preliminary estimates, not calibrated posteriors.
EXPERIMENT III(a)
use the pre-trained model as a inverse modeling tool: estimate fire locations from (generated) smoke process
Synthetic case
green: the source that generated the smoke · circles: inferred candidates
Real GOES case
California 2024-07-26 · green: GOES fire detections · circles: inferred candidates
EXPERIMENT III(b)
use the pre-trained model as a inverse modeling tool: estimate precipitation from (generated) smoke process
Green: the rain used to generate each sequence. Dots: the rain the frozen model infers from the smoke alone, two repeated fits.
The order is always recovered (0 < 1 < 5 mm/h). Magnitudes are underestimated, about 0.3–0.8 for 1 mm/h and 3.2–3.6 for 5 mm/h.