Latent Twins: A Framework for Scene Recognition and Fast Radiative Transfer Inversion in FORUM All-Sky Observations

TL;DR

Latent Twins present a data-driven, physics-aware surrogate framework for all-sky radiative-transfer inversion in FORUM observations, using paired autoencoders for atmospheric states and spectra with bidirectional latent mappings to form surrogate forward and inverse operators. The model is augmented with lightweight model-consistency corrections for cloud variables, enabling physically plausible cloud reconstructions. On synthetic FORUM-like data, the approach achieves near-real-time retrievals of temperature, humidity, ozone, and surface emissivity, with strong scene classification (clear vs cloudy) and informative, though more challenging, cloud-property estimates. The method promises fast, scalable, and interpretable priors or surrogates for full-physics data assimilation and climate studies, with clear paths for extension to uncertainty quantification and real instrument data.

Abstract

The FORUM (Far-infrared Outgoing Radiation Understanding and Monitoring) mission will provide, for the first time, systematic far-infrared spectral measurements of Earth's outgoing radiation, enabling improved understanding of atmospheric processes and the radiation budget. Retrieving atmospheric states from these observations constitutes a high-dimensional, ill-posed inverse problem, particularly under cloudy-sky conditions where multiple-scattering effects are present. In this work, we develop a data-driven, physics-aware inversion framework for FORUM all-sky retrievals based on latent twins: coupled autoencoders for atmospheric states and spectra, combined with bidirectional latent-space mappings. A lightweight model-consistency correction ensures physically plausible cloud variable reconstructions. The resulting framework demonstrates potential for retrievals of atmospheric, cloud and surface variables, providing information that can serve as a prior, initial guess, or surrogate for computationally expensive full-physics inversion methods. It also enables robust scene classification and near-instantaneous inference, making it suitable for operational near-real-time applications. We demonstrate its performance on synthetic FORUM-like data and discuss implications for future data assimilation and climate studies.

Figures (18)

• Figure 1: Schematic of the Latent Twin (paired autoencoder) architecture for inverse problems. Two autoencoders are trained on the parameter space $\mathcal{X}$ and the observation space $\mathcal{Y}$, producing low-dimensional latent representations $z_x$ and $z_y$, respectively. The inverse operator is learned directly in latent space through a trainable mapping $s^{\gets} : \mathbf{z}_y \mapsto \mathbf{z}_x$, which is composed with the observation encoder and state decoder to form the full inverse surrogate/ latent twin $f^{\gets} = d_x \circ s^{\gets} \circ e_y$. By learning the inverse transformation in a shared latent coordinate system, the framework enables efficient and stable recovery of high-dimensional atmospheric states from observations, while preserving consistency between the data and model representations.
• Figure 2: Average January emissivity profiles with standard deviation (shaded) for ocean (WAT) points (left panel) and land points dominated by deciduous (DEC) land cover (right panel).
• Figure 3: Workflow of the algorithm for the generation of the dataset. Atmospheric input data are retrieved from source databases (yellow, left) and processed through models and preprocessing codes (blue and purple) before being ingested by the forward model (green). The resulting high-resolution spectra are then convolved and corrupted with noise (red) to generate the final simulated spectral database (yellow, right).
• Figure 4: Variables associated with atmospheric scenarios, presented with their ranges across pressure layers. The lines indicate the minimum and maximum values in the database, while the shaded area shows the variability between them.
• Figure 5: Retrieval results for atmospheric state variables: temperature, water vapor, ozone, and spectral emissivity (clockwise from top-left). The blue solid line shows the MBE, while the blue and red dashed lines indicate $\textbf{MBE} \pm \textbf{MAE}$ and $\pm\,\textbf{MAD}$, respectively. For surface temperature, the mean bias error is $0.035~K$, the mean absolute error is $2.649~K$, and the mean absolute deviation of the scenes is $17.413~K$.