Abstract Although many large‐scale sea ice models now represent salt dynamics, few studies have evaluated them, and uncertainty remains regarding key parameterization choices. Here, we conduct global 1° ocean–ice simulations with NEMO‐SI 3 v5, forced by climatological atmospheric data, tuned to the salinity–thickness relationship, and evaluated against more than 1,400 ice cores. Simulations investigate (a) the treatment of vertical ice salinity variations (diagnostic vs. prognostic), (b) gravity drainage physics (brine fraction‐based vs. Rayleigh number‐based), and (c) key parameter values. On average, ice salinity is simulated within 0.1 g/kg (Arctic) and 0.7 g/kg (Antarctic) of observations, capturing large‐scale features including the Arctic–Antarctic contrast, seasonality, vertical profile, and the salinity–thickness relationship. Using a high value for the new ice liquid fraction parameter is critical for simulating high salinity in the thin ice range. Vertically‐resolved salt dynamics generally outperform the diagnostic approach, consistently across hemispheres, seasons, depth and thickness. Brine fraction‐based and Rayleigh number‐based parameterizations of gravity drainage produce different salinity fields, both through direct effects and interactions with melt ponds. However, limited observational constraints on measurement biases prevent firm conclusions on which approach performs best. Uncertainties in ice salinity are generally larger in the Antarctic than in the Arctic. While salt dynamics strongly affect sea ice mass and salt budget terms, they affect sea ice volume moderately, and have little effect on ocean salinity due to compensating effects.
Abstract This study investigates the impact of convective microphysics parameterization on the simulation of the Madden–Julian Oscillation (MJO) in the Energy Exascale Earth System Model (E3SM). Specifically, the two‐moment, five‐class convective microphysics parameterization (CMP) scheme from Song et al. (2025, https://doi.org/10.1029/2024ms004656 ) was implemented into Zhang and McFarlane (1995, https://doi.org/10.1080/07055900.1995.9649539 ) convection scheme in E3SM. Results from 15 years of simulations demonstrate that the CMP improves the MJO simulation, particularly in terms of amplitude, period, propagation, life cycle, and seasonality. Wavenumber‐frequency spectra analysis reveals that, with the CMP, the simulated energy peak shifts from 80–180 days to 30–80 days, better aligning with observations. Composite MJO life cycle analysis shows that the CMP enhances precipitation amplitude, eastward propagation, and coherence with zonal winds, addressing weaknesses in the standard model, such as a weak MJO amplitude and delayed propagation. Additionally, in summer, the CMP allows E3SM to better capture the seasonal variation of MJO convective systems, particularly the northward shift in the Asian monsoon region. The CMP enhances MJO convection, making it stronger, and promotes pre‐conditioning east of the MJO convective center, favoring eastward propagation. Model‐process‐oriented diagnostics reveal that the CMP influences the MJO through complex interactions among convection, stratiform clouds, and circulation.
Abstract Ensemble perturbations in 4DEnVar‐type methods are pulled in two directions at once: estimating a reliable background‐error covariance (BEC) calls for perturbations large enough to sample real error structure, while a valid tangent‐linear (TL) approximation calls for perturbations small enough to stay near‐linear. This study resolves this conflict by introducing a shrinkage factor () into a multigrid NLS‐4DVar framework, decoupling the perturbation scale used for BEC estimation from that used for TL approximation. An independent finite‐difference test verifies this decoupling, showing that TL accuracy is governed by a trade‐off between nonlinear truncation error and numerical error. Based on this framework, an improved Regional Multiscale Nonlinear Least Squares (NLS) 4DVar Data Assimilation System was developed based on the CWRF‐CoLM coupled model. A 1‐month cycling assimilation experiment (July 2021) shows that the new scheme (DA2) outperforms the original (DA1): analysis RMSE and bias are reduced for U / V winds, T , and QVAPOR, 3‐day forecast RMSE improves broadly, and the 24‐hr precipitation threat score at moderate‐to‐heavy rainfall thresholds is notably higher, which is consistent with a preferential benefit for strongly nonlinear moisture and convective processes. Although only atmospheric observations are assimilated, land‐surface forecasts (2‐m temperature, 2‐m humidity, soil moisture) also improve indirectly, driven by a more accurate initial moisture field and better‐represented land‐atmosphere coupling. Beyond its practical performance, this decoupling strategy completes a previously unverified aspect of the NLS‐4DVar formulation and offers a generalizable approach for mitigating nonlinear errors in ensemble‐based variational assimilation.
Abstract The current study presents machine learning (ML)‐based schemes that independently represent model error and observation error during data assimilation. The Model Error ML scheme (MLME) uses a generative adversarial network (GAN) to produce flow‐dependent model corrections. The Observation Error ML scheme (MLOE) also uses the GAN to adaptively estimate the observation error covariance matrix. Furthermore, a generalized formulation (CorrFOE) is derived to incorporate the often‐neglected cross‐correlations between forecast and observation errors. The proposed schemes are rigorously evaluated via twin experiments with the Lorenz III model. Results show that the ML‐based schemes reduce the root‐mean‐square error (RMSE) of analyses relative to the control and, for model error, relative to the additive‐noise baseline. Additionally, the MLME scheme is particularly effective at recovering small‐scale structures. Crucially, the synergistic combination of both ML schemes with the explicit inclusion of cross‐correlation between forecast and observation errors produces the most accurate analyses. The added computational costs for MLME, MLOE and CorrFOE are moderate, confirming their operational potential.
Abstract Clouds strongly impact net surface radiation, creating differential heating patterns, and trigger land breezes that in turn organize cloud development. These dial processes determine the location of persistent ecohydrological regimes at local scales. Aerosol–cloud interactions (ACI) directly impact the diurnal cycle of cloud optical depth, and therefore the surface energy budget; yet they are generally not explicitly represented in studies of land‐atmosphere interactions. To address this gap, a 3D Large‐Eddy Simulation model was modified to describe ACI in the gray zone. The goal is to simulate realistic cumulus cloud dynamics consistent with the spatial dynamics of land surface moisture and temperature, and associated boundary layer processes. The model is tested and evaluated using simulations at hectometer resolution with prescribed diurnal surface heating corresponding to idealized wet (low Bowen ratio [LBR]) and dry (high Bowen ratio [HBR]) surface boundary conditions. Expected cloud sensitivity is well captured: compared to LBR, HBR clouds have a significantly higher cloud base (about 1 km), up to 50% reduction in total cloud fraction, and cloudiness ends earlier by about 2 hr due to enhanced turbulence and lower moisture availability. Analyses of the evolution of 3D cloud drop size distributions show that the model captures the cloud life cycle, including aerosol activation, condensational growth, enhanced collision‐coalescence, and aerosol regeneration. The results show that resolved eddy‐scale (sub‐km) within‐cloud circulations modulate cloud microphysical structure by locally altering supersaturation at minute timescales. Successive aerosol regeneration yields coarser cloud condensation nuclei, producing a coarse‐aerosol mode, and impacting mid‐day cloud fraction and cloudiness persistence.
