<?xml version="1.0"?>
<rdf:RDF xmlns:rdf="http://www.w3.org/1999/02/22-rdf-syntax-ns#" xmlns:dc="http://purl.org/dc/elements/1.1/"><rdf:Description rdf:about="https://repozitorij.uni-lj.si/IzpisGradiva.php?id=189554"><dc:title>Preparing the initial conditions for numerical weather prediction with machine learning</dc:title><dc:creator>Melinc,	Boštjan	(Avtor)
	</dc:creator><dc:creator>Zaplotnik,	Žiga	(Mentor)
	</dc:creator><dc:creator>Skok,	Gregor	(Komentor)
	</dc:creator><dc:subject>data assimilation</dc:subject><dc:subject>3D-Var</dc:subject><dc:subject>background-error covariances</dc:subject><dc:subject>physical balances</dc:subject><dc:subject>control variable transform</dc:subject><dc:subject>analysis increments</dc:subject><dc:subject>machine learning</dc:subject><dc:subject>autoencoder</dc:subject><dc:subject>latent space</dc:subject><dc:description>The initial conditions for numerical weather prediction (NWP) are obtained by optimally combining the information from the previous forecast (the background) with newly available observations through data assimilation (DA). The output of a DA algorithm (the analysis) is considered to be the best estimate of the state of the atmosphere, and its correction with respect to the background is typically referred to as the analysis increment. Traditionally, DA of atmospheric observations relies on variational and Kalman filter methods, which are theoretically well established but require some simplifications in operational weather forecasting. 

In this thesis, we develop a standalone machine-learning-based equivalent to three-dimensional variational (3D-Var) DA, where the 3D-Var cost function is minimised in a reduced-order latent space of an autoencoder (AE)-type neural network.  When the atmospheric state is transformed from full physical space to latent space, it undergoes strong compression, which results in decorrelation among the latent elements. This leads to a quasi-diagonal background-error covariance ($\mathbf{B}_z$) matrix, effectively replacing the traditional control variable transform (CVT) and substantially simplifying the computational procedure. 

The proposed framework is evaluated in three configurations of increasing complexity, ranging from a univariate single-level setup to a multilevel system incorporating several atmospheric, oceanic and land variables. Both climatological and operational ensemble-based estimates of $\mathbf{B}_z$ are applied. Single-observation experiments produce analysis increments that are consistent with multivariate physical balances in the tropical and extratropical troposphere, the stratosphere, the atmospheric boundary layer, and coupled atmosphere--sea-ice interactions. These results indicate a possible pathway towards a unified representation of background-error covariances within the atmosphere and across Earth-system model components for DA in operational NWP.

The resulting analysis increments are inherently flow dependent and reflect local features such as orography and land--sea contrasts. Forecasts initialised from these analyses exhibit realistic weather evolution, including the excitation of an eastward-propagating Kelvin wave in the tropics. The proposed latent-space 3D-Var framework could be extended to a four-dimensional formulation, while the developed background-error covariance estimation method could also be applied in hybrid systems as a replacement for the traditional CVT.</dc:description><dc:date>2026</dc:date><dc:date>2026-10-09 08:15:09</dc:date><dc:type>Doktorsko delo/naloga</dc:type><dc:identifier>189554</dc:identifier><dc:language>sl</dc:language></rdf:Description></rdf:RDF>
