This study evaluates convective cell properties and their relationships with convective and stratiform rainfall within a season-long convection-permitting simulation over central Argentina using measurements from the RELAMPAGO-CACTI field campaign. While the simulation reproduces the total observed rainfall, it underestimates stratiform rainfall by 46% and overestimates convective rainfall by 43%. As Convective Available Potential Energy (CAPE) increases, the overestimation of convective rainfall decreases, but the underestimation of stratiform rainfall increases such that the high bias in the contribution of convective rainfall to total rainfall remains approximately constant at 26% across all CAPE conditions. Overestimated convective rainfall arises from the simulation generating 2.6 times more convective cells than observed despite similar observed and simulated cell growth processes, with relatively wide cells contributing most to excessive convective rainfall. Relatively shallow cells, typically reaching heights of 4–7 km, contribute most to the cell number bias. This bias increases as CAPE decreases, potentially because cells and their updrafts become narrower and more under-resolved as CAPE decreases. The gross overproduction of shallow cells leads to overly efficient precipitation and inadequate detrainment of ice aloft, thereby diminishing the formation of robust stratiform rainfall regions. Decreasing the model’s horizontal grid spacing from 3 to 1 or 0.333 km for representative low and high CAPE cases results in minimal change to the cell number and depth biases, while the stratiform and convective rainfall biases also fail to improve. This suggests that improving prediction of deep convective system growth depends on factors beyond solely increasing model resolution.

Mesoscale convective systems (MCSs) are an important component of our hydrologic cycle as they produce prolific rainfall in the tropics and mid-latitudes. Recent advancements in high-resolution modeling show promise in representing MCSs in regional climate simulations. However, how well do these models represent the complex interactions between convective dynamics and microphysics in MCSs remain unknown. In this study, we take advantage of observations collected during the Midlatitude Continental Convective Cloud (MC3E) experiment to evaluate multi-scale aspects of MCSs in convection-permitting WRF model. We conducted three sets of month-long simulations with Morrison and P3 (1-ice and 2-ice categories) microphysics, respectively, at 1.8 km grid-spacing over the Southern Great Plains. MCSs in observations and simulations were tracked using a newly developed FLEXTRKR algorithm. About 15-20 MCSs were identified in the simulations, consistent with observations. All three simulations underestimate observed monthly total precipitation which are primarily from MCSs, suggesting the biases might be caused by large-scale forcings rather than microphysics. All simulated MCSs overestimate convective area and precipitation amount but underestimate stratiform rain area and precipitation. Simulated MCS convective updraft intensities are comparable with radar retrievals for moderate depths of convective cores, but are too strong for deep cores. The two P3 simulations have smaller mean ice mass aloft but more frequent heavy convective rain rate at the surface than the simulation with Morrison, agreeing better with observations (Figure 1). Simulated stratiform area ice mass in the upper troposphere are generally larger than radar retrievals, but the P3 2-ice category has relatively smaller bias. We will also use polarimetric radar 3-D rain water retrieval to further evaluate the vertical evolution of rainfall to explain differences in simulated surface precipitation.

Computational models of the Earth System are critical tools for modern scientific inquiry. Efforts toward evaluating and improving errors in representations of physical and chemical processes in these large computational systems are commonly stymied by highly nonlinear and complex error behavior. Recent work has shown that these errors can be effectively predicted using modern Artificial Intelligence (A.I.) techniques. In this work, we go beyond these previous studies to apply an interpretable A.I. technique to not only predict model errors but also move toward understanding the underlying reasons for successful error prediction. We use XGBoost classification trees and SHapley Additive exPlanations (SHAP) analysis to explore the errors in the prediction of lightning occurrence in the NASA GEOS model, a widely used Earth System Model. This interpretable error prediction system can effectively predict the model error and indicates that the errors are strongly related to convective processes and the characteristics of the land surface.