Surface Solar Irradiance in Continental Shallow Cumulus Fields: Observations and Large-Eddy Simulation

Abstract

This study examines shallow cumulus cloud fields and their surface shortwave radiative effects using large-eddy simulation (LES) along with observations across multiple days at the Atmospheric Radiation Measurement Southern Great Plains atmospheric observatory. Pronounced differences are found between probability density functions (PDFs) of downwelling surface solar irradiance derived from observations and LES one-dimensional (1D) online radiation calculations. The shape of the observed PDF is bimodal, which is only reproduced by offline three-dimensional (3D) radiative transfer calculations, demonstrating PDF bimodality as a 3D radiative signature of continental shallow cumuli. Local differences between 3D and 1D radiative transfer calculations of downwelling surface solar irradiance are, on average, larger than 150 W m−2 on one afternoon. The differences are substantially reduced when spatially averaged over the LES domain and temporally averaged over the diurnal cycle, but systematic 3D biases ranging from 2 to 8 W m−2 persist across different days. Covariations between the domain-averaged surface irradiance, framed as a surface cloud radiative effect, and the simulated cloud fraction are found to follow a consistent diurnal relationship, often exhibiting hysteresis. In contrast, observations show highly variable behavior. By subsampling the LES domain, it is shown that this is due to the limited sampling density of inherently 3D observations. These findings help to define observational requirements for detecting such relationships, provide valuable insight for evaluating weather and climate models against surface observations as they push to ever higher resolutions, and have important implications for future assessments of solar renewable energy potential.

On the Relationship Between Shallow Cumulus Cloud Field Properties and Surface Solar Irradiance

“…Predictions from both the RF and the ANN are able to capture variations in shape, size and location of both modes but neither method shows consistent improvement over the other. For context, commonly used 1‐D radiation schemes do not capture the bimodal shape at all, let alone the variations in the details of each mode (see Gristey et al, 2020 Figure 4 or Schmidt et al, 2007 Figure 10b). Figures 3 and 4 therefore demonstrate, both quantitatively and qualitatively, that machine learning methods using just a few shallow cumulus cloud field properties as input can produce more realistic SSI PDFs than 1‐D radiative transfer.…”

“…Simultaneous 3‐D LES output (specifically cloud liquid water content and drop number concentration) is then ingested into offline 3‐D radiative transfer to compute SSI that is subsequently normalized by cos ( SZA ) to remove variations in the magnitude of incoming solar irradiance. We use the Education and Research 3‐D Radiative Transfer Toolbox (EaR 3 T), which is based on the Monte Carlo Atmospheric Radiative Transfer Simulator (MCARaTS Iwabuchi, 2006), with an identical configuration as described by Gristey et al (2020). Since some of the LASSO cases deviate substantially from shallow cumulus, we subset the LES 10‐min output for 3‐D radiative transfer using the following criteria: cloud fraction 5%–40%, no significant cloud ice, greater than 10 individual clouds in our domain, mean cloud size less than 2 km, and solar zenith angle less than 40° to capture the afternoon evolution of shallow cumulus where variations in the cloud field properties are most likely to influence the surface irradiance.…”

“…In this study, we use LES of 30 nonprecipitating shallow cumulus cases, selected from the LES Atmospheric Radiation Measurement (ARM) Symbiotic Simulation and Observation activity (LASSO; Gustafson et al, 2020), spanning the summers of 2015 (5 days), 2016 (10 days), and 2017 (15 days). Simulations are performed using the System for Atmospheric Modeling (SAM; Khairoutdinov & Randall, 2003) with two‐moment bulk microphysics (Morrison & Gettelman, 2008), a domain size of 24 × 24 × 15 km and a horizontal grid spacing of 100 m; further details are given in Gristey et al (2020) and Glenn et al (2020) who used an identical setup. Output is considered every tenth simulated minute throughout the duration of the runs from 07:00–22:00 local time (LT; Central Daylight Time or UTC–5).…”

“…For widely used one‐dimensional (1‐D) radiation schemes, which consider vertical columns independently and apply a plane‐parallel approximation within each column, the observed PDF bimodality is not always reproduced even with well‐resolved cloud fields from large eddy simulation (LES). However, for identical cloud fields, the bimodality can be reliably reproduced with Monte Carlo 3‐D radiative transfer (Gristey et al, 2020). While it is clear that the shape of the bimodal PDF is related to the properties of the cloud field, a direct mapping between the two has never been investigated.…”

Process‐Based Climate Model Development Harnessing Machine Learning: III. The Representation of Cumulus Geometry and Their 3D Radiative Effects

“…The sign of resulting 3D effects depends on solar zenith angle. Gristey et al (2020) found that 3D effects of sub-tropical land cumulus fields act to heat the surface when averaged over a diurnal cycle; neglecting these effects in climate models might introduce significant errors in the predicted evolution of the system.…”