Nonlinear Atmospheric Inversion with Interpretable Bias Correction via Gaussian Process Prior

Accurate quantification of atmospheric pollutant emissions is essential for evaluating the consequences of environmental incidents. Inverse modelling of such releases commonly employs a linear framework based on a source–receptor sensitivity (SRS) matrix; however, this matrix can be substantially biased or may even fail to represent the true scale of the release. We introduce a method in which the SRS matrix is corrected jointly with the inversion, resulting in a nonlinear inverse problem. The SRS discrepancies are interpreted as small shifts of observation points, leading to a deformation of the sensitivity field. The shifts are regularized through a Gaussian process prior, which imposes smoothness and sparsity while allowing inference at unobserved locations. The resulting posterior predictions of the shift field offer a practical tool for hyperparameter selection: the inferred shifts can be visualized geographically and evaluated by domain experts. This leads to a Bayesian framework that integrates inversion, SRS correction, and a tuning strategy based on L-curve-type diagnostics combined with maps of the predicted shifts. It will be demonstrated on a selected real continental-scale scenario of an atmospheric release.

 

This research has been supported by the Czech Science Foundation (grant no. GA24-10400S). FLEXPART model simulations are cross-atmospheric research infrastructure services provided by ATMO-ACCESS (EU grant agreement No 101008004). Nikolaos Evangeliou was funded by the same EU grant. The computations were performed on resources provided by Sigma2 - the National Infrastructure for High Performance Computing and Data Storage in Norway.