Application of the latent twins approach for clear sky retrieval from IASI observations

In recent years, data-driven approaches emerged as alternatives to traditional physics-based retrievals, taking advantage of machine learning techniques such as learnable pseudoinverse, random forests, or deep learning architectures. Classical data-driven models generalize poorly to out-of-sample regimes, as they optimize over finite datasets without incorporating underlying physical laws. This often requires large models and extensive data to achieve reliability. Physics-Informed Neural Networks address this by embedding physical constraints into the learning process, enabling improved extrapolation. However, they requires substantial computational cost due to the need to solve governing equations at each training step. In this work, we introduce a novel deep learning architecture, based on latent twin approach, that balances model complexity, dataset size, and training cost, while providing a quantitative measure of data quality. This architecture is applied to IASI spectra, with the goal to assess the robustness of this method for retrieving atmospheric profiles, including temperature, water vapor, ozone, surface emissivity, and surface temperature, in real-world clear-sky conditions. The algorithm is first applied on synthetic radiances derived from the NWP SAF database using the fast radiative transfer code sigma-IASI/F2N. After validating the architecture on synthetic data, the algorithm is applied to IASI Level 1C observations, along with their corresponding Level 2 products which serve as reference to evaluate the reconstruction accuracy of the autoencoder-based retrieval. The retrieval performances are discussed along with possible strategies to provide an error analysis for the reconstructed thermodynamical profiles.