Spatial-temporal source term estimation using deep neural network prior and its application to Chernobyl wildfires.

Area of Science:

• Environmental Science and Engineering
• Atmospheric Science
• Nuclear Safety

Background:

• Accurate source term estimation of atmospheric hazardous material emissions is vital for analyzing unintended releases, especially concerning radioactive contamination from wildfires.
• Existing atmospheric inverse problems are often ill-posed, with insufficient measurement data to fully characterize the 5D emission tensor (spatial location, time, height, particle size).

Purpose of the Study:

• To propose a novel method for regularizing atmospheric inverse problems in hazardous material emission analysis.
• To leverage deep learning, specifically a deep image prior, for improved source term estimation in complex scenarios like radioactive wildfires.

Main Methods:

• Introduction of a deep image prior method, utilizing a randomly initialized deep neural network structure without prior dataset training.
• Integration of variational optimization with the deep image prior to regularize the inversion process.
• Enforcement of a prior covariance structure in the source term to reduce unknowns and enhance spatial smoothness.

Main Results:

• The proposed method successfully introduces spatial smoothness in emission estimates.
• The approach effectively reduces the number of unknowns by incorporating prior structural information.
• Demonstrated effectiveness on estimating Cesium-137 (¹³⁷Cs) emissions during the 2020 Chernobyl wildfires.

Conclusions:

• Deep image prior combined with variational optimization offers a powerful, data-efficient regularization technique for atmospheric source term estimation.
• This method shows significant promise for improving the analysis of airborne radioactive material transmission, particularly in post-wildfire contamination events.