There are significant gaps in both experimental and theoretical understanding of mixed-phase clouds, their impacts on the hydrological cycle as well as their effects on atmospheric radiation. Accurately identifying liquid water layers in mixed-phase clouds is crucial for estimating cloud radiative effects. A proof-of-concept study utilizing a machine-learning-based liquid-layer detection method called VOODOO is presented. This method was applied alongside a single-column radiative transfer model to compare downwelling shortwave fluxes of mixed-phase clouds detected by the standard Cloudnet processing chain and VOODOO to ground-based pyranometer observations. Our findings reveal that VOODOO creates more realistic liquid water content distributions and significantly influences profiles of heating rates. Moreover, our study demonstrates a substantial enhancement in the estimation of shortwave cloud radiative effects of VOODOO compared to conventional method Cloudnet. Specifically, we observe a remarkable reduction in the mean absolute error of simulated shortwave radiation at the surface of 70\%, particularly in homogeneous cloud conditions. The mean percentage error of SW cloud radiative effects between Cloudnet and pyranometer observations is 44\%, while VOODOO+Cloudnet reduces this error to 8\%. Overall, our results underscore the potential of VOODOO to provide new insights into deep mixed-phase clouds, which were previously inaccessible using traditional lidar-based remote sensing techniques.

Accurately identifying liquid water layers in mixed-phase clouds is crucial for estimating cloud radiative effects. Lidar-based retrievals are limited in optically thick or multilayer clouds, leading to positive biases in simulated shortwave radiative fluxes. At the same time, general circulation models also tend to overestimate the downwelling shortwave radiation at the surface especially in the Southern Ocean regions. To reduce this SW radiation bias in models, we first need better observational-based retrievals for liquid detection which can later be used for model validation. To address this, a machine-learning-based liquid-layer detection method called VOODOO was employed in a proof-of-concept study using a single column radiative transfer model to compare shortwave cloud radiative effects of liquid-containing clouds detected by Cloudnet and VOODOO to ground-based and satellite observations. Results showed a reduction in shortwave radiation bias, indicating that liquid-layer detection with machine-learning retrievals can improve radiative transfer simulations.