Towards Better Simulations of Liquid Water Path during Stratocumulus-to-Cumulus Transition: Insights from Gaussian Process Emulator, XGBoost and MAGIC Observations

Abstract. Simulating Liquid Water Path (LWP) during stratocumulus-to-cumulus transition (SCT) remains challenging for storm-resolving models, with biases varying across cloud regimes. We use the storm-resolving DP-EAMxx model and a perturbed-parameter ensemble of three warm-rain microphysical parameters to investigate LWP biases during an SCT event observed in the MAGIC field campaign. Gaussian process emulators trained on observation-derived metrics of mean LWP bias and LWP decorrelation timescale bias are used to identify low-bias parameter combinations within the explored parameter space. While similar parameter constraints are obtained for the stratocumulus (Sc) and transition (Tr) phases, the low-bias parameter combinations for the cumulus (Cu) phase differ substantially, indicating requirement of a stronger reduction in autoconversion and accretion rates for a given prescribed droplet number concentration. Using an overlapping low-bias parameter set from the Sc and Tr phases, the mean LWP bias improves from −31 and −22 g m⁻² to −1 and 3 g m⁻² in the Sc and Tr phases, respectively, but degrades in the Cu phase. An independent XGBoost model with SHAP attribution, trained using DP-EAMxx-simulated process-level diagnostics and meteorological state, supports the emulator sensitivities: LWP bias in Sc is strongly associated with warm-rain microphysics, whereas dynamical, radiative, and thermodynamic influences become more prominent in Tr and Cu. These results show where a limited set of parameters is effective in improving model performance and where additional sources of uncertainty likely need to be considered across regimes. More broadly, we demonstrate a proof-of-concept observation-constrained framework for the diagnosis of storm-resolving model bias.