Deep Reinforcement Learning – Driven Optimization of Ice Cloud Bulk Scattering Properties in RTTOV-SCATT Using Arctic Weather Satellite Observations

Abstract. The Arctic Weather Satellite (AWS), launched by the European Space Agency (ESA) in August 2024, has enabled the first-ever global ice cloud remote sensing using 325 GHz terahertz channels. Owing to its short wavelength, the 325 GHz frequency band is highly sensitive to scattering by the three frozen hydrometeor types present in ice clouds—cloud ice, snow, and graupel. Radiative transfer (RT) models are the core tools for satellite remote sensing retrieval and data assimilation, and their accuracy directly determines the capability to retrieve ice cloud parameters from satellite observations. Due to the complexity of cloud microphysical processes and insufficient observational constraints, the parameterization of the Bulk Scattering Properties (BSPs) of frozen hydrometeors in the current operational RT model RTTOV still harbors considerable uncertainty. The resulting large discrepancies between simulated brightness temperatures (TBs) from high-frequency channels and AWS-observed TBs severely limit the retrieval accuracy of ice cloud parameters.

This paper proposes a physically-informed deep reinforcement learning (DRL) framework for optimizing the computation of frozen hydrometeor BSPs. The framework employs the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm to couple the RTTOV-SCATT v14.0 module with a differentiable surrogate neural network model, using AWS-observed ice cloud TBs as constraints to optimize the bulk extinction coefficient, scattering coefficient, and asymmetry factor of frozen hydrometeors. Training and validation experiments are conducted using approximately 176,000 AWS-observed TB data records collected since July 2025 (filtered to approximately 80,000 valid ice cloud samples). Results show that after optimization, the root-mean-square error (RMSE) between AWS-observed and simulated ice cloud TBs is reduced to 4.98–14.33 K for the 183 GHz channels and 9.54–13.32 K for the 325 GHz channels, representing an improvement of 56.3–69.7 % over the standard RTTOV-SCATT module. The distribution of observation-minus-background (O–B) deviations shifts from markedly skewed to approximately Gaussian. The optimized model significantly improves the simulation accuracy of spaceborne terahertz-band ice cloud TBs while retaining the computational efficiency of RTTOV (approximately 0.2 ms per sample).