| 25 Mar 2026
Capturing and explaining the effects of three-dimensional radiative transfer on cloud evolution with the dynamic TenStream solver
Abstract. Radiative transfer is an inherently three-dimensional (3D) process that, for computational reasons, is still approximated as one-dimensional (1D) in most atmospheric models. To address this limitation, Maier et al. (2024) introduced the dynamic TenStream solver, which reduces the cost of 3D radiative transfer calculations through incomplete solves. Here, we investigate how coupling dynamic TenStream to the large-eddy simulation model PALM affects cloud development compared to simulations using conventional 1D and full 3D radiation. Results show that during daytime, clouds driven by either of the 3D solvers organize into cloud streets oriented perpendicular to the solar incidence angle, whereas with 1D radiation they remain more or less randomly distributed. Moreover, daytime clouds grow larger, become thicker, and contain more liquid water with 3D radiative transfer. It is shown that these differences arise because, unlike in the 1D case, clouds coupled to 3D radiation are not positioned directly above their own shadows. Instead, they are located over areas of enhanced net surface irradiance, where values even exceed those in the clear-sky columns of the 1D simulation, strengthening rather than weakening the associated updrafts. Additionally, 3D radiation is shown to reduce the domain-averaged net thermal emission at the surface, which affects the surface energy budget and is primarily balanced by an increase in the domain-averaged latent heat flux, resulting in a greater release of water vapor into the atmosphere. Both effects are captured by dynamic TenStream, demonstrating its ability to represent 3D radiative effects on cloud development at a substantially lower computational cost.
This manuscript couples the dynamic TenStream 3D radiative transfer solver to the PALM large-eddy simulation model and investigates how 3D versus 1D radiative transfer treatments affect shallow cumulus cloud field evolution over a full diurnal cycle. It also serves as an online validation of the dynamic TenStream solver against the original, more expensive TenStream implementation. The manuscript provides a number of novel and significant insights including:
Separately from the physical insights provided by the manuscript, this work also includes validation of the dynamic TenStream solver against the original version:
Below are some suggestions to improve this already strong manuscript.
Manuscript organization
The manuscript tackles two objectives: (1) to validate the performance of the dynamic TenStream solver against the original, higher-fidelity TenStream solver in an online setting and (2) to discuss how 3D versus 1D treatments of radiation impact cloud field statistics of shallow cumulus. While there are some subtle lessons about the cloud field to be learned from the differences between the dynamic and original TenStream solvers (i.e. lingering surface enhancements in the dynamic version), for the most part these two objectives appear separate. It would benefit the reader for the authors to treat them as such, particularly within sections 3.2 and 3.3. Subsections might help here.
Emphasis of novel results
Several of the manuscript's most interesting results could be given greater prominence.
This work demonstrates for the first time that radiatively driven cloud streets can develop with 3D RT under realistic diurnal conditions, clearly distinguishing it from prior work by Jakub and Mayer (2017) in which solar zenith angle was kept fixed. The manuscript would benefit from better highlighting this result.
This reader is fascinated by the result that for higher cloud cover days, cloud coverage for simulations with a 1D treatment of radiation is higher than in simulations with a 3D treatment of radiation (supported by the results of Tijhuis et al. 2024). Why might this be? Further discussion would be greatly appreciated.
The manuscript shows that longwave 3D RT impacts the surface energy budget via reduced thermal emission, propagating into increased surface latent heat fluxes. This higher moisture flux into the atmosphere from the surface is hypothesized by the authors to partially explain the higher LWP in the 3D RT simulations. To this reviewer’s knowledge, this mechanism is novel in the literature and deserves greater focus. Do the authors believe it is equally significant to the mechanism whereby clouds in the 3D simulations are positioned over areas of enhanced net surface irradiance, strengthening rather than suppressing the associated updrafts? Can the two mechanisms be isolated to assess their relative importance? If computational resources allow, simulations with 3D RT for shortwave only and with 3D RT for longwave only would be of interest. Work from Veerman et al. 2022 and Tijhuis et al. 2024 suggests that the shortwave mechanism may be dominant as they see substantial LWP increases with 3D RT without longwave effects.
The differences in 3D versus 1D shortwave surface flux fields, as described at the start of section 3.3, are well documented in the literature. A briefer introduction of these concepts might allow the authors to place stronger emphasis on (a) the connection between surface flux heterogeneities and cloud field evolution (b) what the model artifacts of the dynamic TenStream solver and resulting cloud field differences can teach us about this mechanism.
Statistical rigor of dynamic TenStream validation
Dynamic TenStream and original TenStream solver differences should not be evaluated solely by comparing their differences with those between 1D and original TenStream; instead the authors should leverage their (modest) ensemble, asking if the observed differences are greater or less than the variability observed for original TenStream. Using additional ensemble members, if computationally feasible, would allow for a more statistically rigorous analysis of the signal to noise ratio of these differences. Is the top row of Figure 6 a useful and fair comparison for the performance of the dynamic TenStream model? Does it provide additional information beyond Figure 4?
General suggestions and questions
The authors should take care to distinguish between surface field lag artifacts of the dynamic TenStream solver and the physical displacement of cloud shadow from cloud root that occurs at moderate solar zenith angles in the original TenStream results (lines 415–420). The former is a solver artifact arising from incomplete solves; the latter is a real 3D radiative effect. The current text could more clearly distinguish between the two.
What is the surface heat capacity C₀ used in the PALM land surface model for this setup? If it is small, would the authors not expect a near-instantaneous response of the surface energy budget to changes in the radiative field (lines 425–435)?