Latent Twins Architecture Cuts Clear-Sky IASI Retrieval Errors on Synthetic and Real Spectra

The preprint introduces a latent-twins encoder-decoder that embeds physical consistency constraints between paired atmospheric states without solving radiative transfer equations at every training step. Training began on 50 000 synthetic IASI spectra generated from the NWP SAF database with sigma-IASI/F2N, then transferred to real MetOp observations. Reconstruction accuracy was quantified against operational Level 2 retrievals, showing reduced root-mean-square errors in the lower troposphere.

Standard data-driven models overfit finite clear-sky samples and degrade outside training distributions; Physics-Informed Neural Networks improve extrapolation but incur high cost. The latent-twins design balances these trade-offs by enforcing twin-state agreement in latent space, yielding a built-in data-quality metric. This directly supports operational needs for cloud-free radiances used in numerical weather prediction and long-term climate monitoring.

Related work on sigma-IASI fast models and earlier PINN applications to infrared sounding confirms that hybrid architectures consistently outperform pure ML or pure physics baselines when sample sizes are moderate. The current study still lacks independent validation against in-situ radiosondes or aircraft data, limiting claims about absolute accuracy.

Next steps include ensemble error propagation and testing on additional infrared sounders such as CrIS to assess cross-instrument transferability.