Weaker land–atmosphere coupling in global storm-resolving simulation

Fig. 1A–C shows the correlation coefficient between soil moisture index (SMI) and precipitation, taken as a measure for the feedback strength (Materials and Methods). As we only have 1 y of simulation at storm-resolving resolution (5 km), we use the SD of the 60 correlation coefficients for each individual year of the coarse-resolution model as a measure of internal variability and to assess the significance of the difference between the two models. We exclude the results in cases where precipitation is smaller than 0.1 mm d${}^{-1}$ in both simulations or where the SMI-precipitation correlation coefficient of the storm-resolving model (SRM) falls within one SD of the year-to-year variability of the correlation coefficient of the coarse-resolution model. Given this definition, 46.8% of the rainy area exhibits a difference in feedback between the two models. For our subsequent analysis, we only consider those points and will always give percentage values with respect to this subset of points except if noted otherwise.

Fig. 1A clearly indicates that the soil moisture–precipitation feedback is weaker in the SRM than in the coarse-resolution model: 90.1% of the points display a weaker correlation. Moreover, 78.2% of the points in the SRM exhibit a correlation coefficient close to zero, with values comprised between $\pm$0.2, whereas a similar fraction of points (70.1%) exhibit a correlation coefficient larger than 0.2 in the coarse-resolution model. As a further check (SI Appendix, Fig. S3), we computed how often the correlation coefficient in the coarse-resolution model is smaller than that in the SRM. For the vast majority of rainy points (61.5%), this happens at best 10% of the time, meaning 6 y out of 60 y, confirming that the feedback is consistently weaker in the storm-resolving simulation. The areas where the coarse-resolution model shows frequent smaller coefficient mostly belong to the areas where the SRM falls within one SD of the year-to-year variability of the correlation coefficient of the coarse-resolution model.

It is not only that the strength is weaker in the SRM, but the sign also differs. In the coarse-resolution model, we observe widespread positive correlations, covering 97.9% of the points. In the SRM, those areas have shrunk to 65.8%. Part of this positive correlation could result from the fact that days with precipitation tend to have higher soil moisture simply because precipitation replenishes soil moisture. To filter out this effect, we follow Duerinck et al. (43) and compute the correlation coefficient between SMI and subsequent 9-d mean precipitation (Fig. 2). In the SRM, this changes the sign of the correlation coefficient to widespread negative (66.6% of the points), whereas the values in the coarse-resolution model remain mostly unaffected. Using different rain averaging periods (SI Appendix, Fig. S4) does not alter this behavior.

Although our study focuses on boreal summer, due to the larger frequency of convective events over land, repeating the analysis for austral summer (SI Appendix, Fig. S5) reveals broadly consistent results. The feedback is weaker in the SRM. Compared to boreal summer, the differences are reduced in the northern hemisphere and enhanced in the southern hemisphere. This is consistent with the shift of frontal activity to the northern hemisphere and of convective activity to the southern hemisphere during austral summer.

To understand why the soil moisture–precipitation feedback is weaker and even negative in the SRM, we decompose the feedback into its terrestrial and atmospheric segments by computing correlation coefficients between SMI and evapotranspiration as well as between evapotranspiration and precipitation.

The strength of the soil moisture–evapotranspiration feedback (Fig. 1D–F) is weaker in the SRM than in the coarse-resolution model at 70.4% of the points. Like with the soil moisture–precipitation feedback, but to a lesser extent, the sign has also changed. In addition, 86.6% of the points exhibit positive correlations in the coarse-resolution model, against 71.3% in the SRM. A weaker or even negative soil moisture–evapotranspiration feedback means that evapotranspiration is less often soil moisture limited and rather constrained by the amount of available energy. This is confirmed by determining to which soil moisture–evaporation regime grid points belong, following the method of Budyko (44, 45) (SI Appendix, Fig. S6). In the coarse-resolution model, the percentages are 6.5% for dry regime, 81.6% for transitional regime, and 11.9% for wet regime. The corresponding numbers are 3.5%, 59.5%, and 37% for the SRM.

The difference in feedback between the two models may stem from a difference in the functional relationship between soil moisture and evapotranspiration or from different soil moisture amounts. Fig. 3A indicates that the strength of the correlation between SMI and evapotranspiration exhibits very similar values across soil moisture indices in both models. This similarity is consistent with the fact that the two models employ the same land surface model, the same soil and vegetation properties except for the leaf area index, see methods and SI Appendix, Fig. S7. The existing differences in leaf area index between the two models nevertheless do not match well with the obtained differences in the soil moisture–evapotranspiration correlation coefficient between the two models. For instance, differences in the correlation coefficient over Europe or over Africa in the northern hemisphere are always of the same sign, whereas differences in leaf area index exhibit both positive and negative values within each domain. Hence, we attribute the weaker soil moisture–evapotranspiration feedback in the SRM primarily to the occurrence of higher soil moisture, as confirmed by Fig. 3A. The two soil moisture distributions are distinct, with the distribution from the SRM being shifted to higher values, independently of the years included in the coarse-resolution model (see thin lines in Fig. 3A). In the SRM, soil moisture indices are higher than 0.5 at 75.6% of the points, whereas this number decreases to only 37.6% for the 60-y mean in the coarse-resolution model. Likewise, if we take 75% of the field capacity as the critical soil moisture value for the start of the wet regime, we find that soil moisture is larger than this value 37% of the times in the SRM, versus 11.9% in the coarse-resolution model. Higher soil moisture values are already present at the beginning of the boreal summer season (SI Appendix, Fig. S8 A–C), and then, soil moisture remains higher in the SRM during the analysis period (SI Appendix, Fig. S1) and most regions even moistened during the summer season, ending up at a higher soil moisture value (SI Appendix, Fig. S8 D–F). This may indicate the tendency of the SRM to equilibrate at higher soil moisture than the coarse-resolution model. That the two models would equilibrate at a different soil moisture would be consistent with the two models having a distinct land–atmosphere coupling and with the fact that simulations with and without convective parameterization rain differently (see figure 9 in Prein et al. 46).

Looking now at the atmospheric segment, the coarse-resolution model exhibits expected positive correlations between evapotranspiration and precipitation. In contrast, in the SRM, those correlations are changed to negative ones (Fig. 1G–I). In the coarse-resolution model, the sign of the feedback is positive for 72.6% of the points. Negative signs can only be found over the regions in the wet regime (compare to SI Appendix, Fig. S6A), those being North-Eastern America and Eastern Asia. In the wet regime, evapotranspiration is primarily determined by net radiation and tends to decrease when precipitation increases due to increased cloud cover (SI Appendix, Fig. S9). In contrast, in the SRM, negative feedback exists at 56.2% of the points. Even over the region with remaining positive feedback, the strength is weaker, with a value of 0.22 compared to 0.35 in the coarse-resolution model on average.

Two mechanisms could explain the weaker and more negative correlations in the SRM: more wet regimes and/or a distinct relationship between evapotranspiration and precipitation. As already indicated in Section 2, wet regimes are much more widespread in the SRM than in the coarse-resolution model. But also the functional relationship between evapotranspiration and precipitation differs: for a given SMI, the correlation between evapotranspiration and precipitation is weaker in the SRM where SMI is smaller than 0.9 (Fig. 3B). More strikingly (Fig. 3C), whereas precipitation linearly increases with evapotranspiration in the coarse-resolution model, there is no monotonic increase in the SRM. This suggests that evapotranspiration is not the main control of precipitation in the SRM, also evidenced by the fact that it can still rain by very small evapotranspiration amounts, in contrast to the coarse-resolution model.

Hohenegger et al. (27) explained the distinct feedback sign between parameterized and explicit convection in their limited-area simulations over the Alpine region by the design of the triggering function of their default convective parameterization, making convection couples to moisture rather than to heat. A similar mechanism could be at play here. At least, the correlation between sensible heat flux and precipitation is indeed stronger in the SRM than in the coarse-resolution model for a given SMI (Fig. 3B). Moreover, in the presence of soil moisture heterogeneities, Taylor et al. (30) showed that afternoon precipitation is triggered over dry soils, while coarse-resolution models favor wet soils. In our SRM, afternoon precipitation is indeed triggered over dry soils (SI Appendix, Fig. S10), as in observations (see figure 1 in Taylor et al. 14 and figure 1A in Guillod et al. 47). These results could further explain the documented weaker evapotranspiration–precipitation feedback in the SRM. Finally, it may just be an indication that the SRM couples convection to circulation (48) rather than to surface fluxes.

In the previous sections, we demonstrated that the coupling between the land surface and the atmosphere is weaker and more negative in the SRM. Do observations agree with our findings? To determine which model better represents the coupling, we compare the simulated coupling to the one recorded by the FLUXNET2015 dataset (49) at 102 sites (Materials and Methods). Although being observations, it is important to remember that those observations still contain measurement errors and noise that will degrade the correlation (50). The comparison also suffers from the different scales considered: point observation is ideally representative for a larger fetch area versus a km-scale and a low-resolution model output downgraded to a 2${}^{°}$ grid analysis box.

Fig. 4 provides clear evidence that the SRM better represents the sign and strength of land–atmosphere coupling. In terms of the soil moisture–precipitation feedback, the SRM exhibits closer values to FLUXNET2015 at 83.3% of sites. Strikingly, the correlation coefficients derived from the coarse-resolution model rarely fall into the range of the correlation coefficients measured at the selected FLUXNET2015 sites. Similarly, for the soil moisture–evapotranspiration and evapotranspiration–precipitation feedback, the SRM indicates closer correlations to FLUXNET2015 at 71.6% and 70.7% of the sites, respectively.

We use two simulations integrated with different configurations of the ICON model. The first one is a fully coupled global simulation integrated at 5 km. This simulation was initialized on 20 January 2020 and integrated until 28 February 2021. It employs the new configuration of the ICON model, called Sapphire and is fully described in Hohenegger et al. (40), see the simulation called G_AO_5km. The second simulation is a fully coupled global simulation integrated at 160 km (42). It is part of the CMIP6 and we use the piControl simulation as it allows a better estimation of interannual variability than a transient simulation such as the historical simulation. The official simulation name is ICON-ESM-LR.piControl (51). We limit our analysis to a 60-y of the simulation as extending the time span of the data does not change the main findings.

The key difference between the two simulations is that convection is explicitly resolved at 5 km but parameterized at 160 km in ICON-ESM-LR.piControl. Both simulations use JSBACH (52) to represent land surface processes, with five soil layers, but slightly different configurations for the vegetation processes. Given the shortness of the simulation, the big leaf (BL) approach is used at 5 km, whereas plant functional types (PFT) approach is used in ICON-ESM-LR.piControl. The main differences between the two approaches are i) BL does not explicitly define a vegetation type, PFT up to eleven types of vegetation per grid box, and ii) BL prescribes leaf area index from monthly-mean observed climatological values, whereas PFT computes leaf area index interactively based on phenology. The other soil and vegetation properties are based on the same datasets.

To quantify the strength and the sign of land–atmosphere coupling, first, the simulation output is daily averaged and, second, interpolated to a common 2${}^{°}$ grid. Using a larger box of, for instance, 5${}^{°}$ grid or 10${}^{°}$ grid, does not affect the main conclusion of the study. Lastly, we compute three Pearson correlation coefficients between: i) SMI and precipitation, ii) SMI, and evapotranspiration, and iii) evapotranspiration and precipitation. The analysis is restricted to the boreal summer season (June–July–August, JJA). For the soil moisture analysis, we use the SMI as this is the variable employed in the computation of evapotranspiration in JSBACH:

where $\theta$ is soil moisture in the root zone, ${\theta }_{\text{pwp}}$ is the permanent wilting point, and ${\theta }_{\text{crit}}$ is 75% of the field capacity. These definitions follow what is done in JSBACH.

Given the 60-y time span of the coarse-resolution model, we compute the correlation coefficients for each year and then average them over 60 y. Subsequently, we only retain the computed correlations if the JJA mean precipitation is larger than 0.1 mm d${}^{-1}$ in both simulations and where differences in the correlation coefficient between SMI and precipitation between the two models may not be due to internal variability. We assess this by computing the SD of the correlation coefficient between SMI and precipitation in the coarse-resolution model based on the 60 y of simulation output and only retain the points where the difference in correlation coefficient between the two models is larger than one SD.

As a further metric, we compute the correlation coefficient between SMI and subsequent precipitation. It is computed following Duerinck et al. (43) by correlating daily mean SMI with precipitation averaged on subsequent days. We tested the sensitivity to the chosen averaging period by varying it between 1 and 120 d, see SI Appendix, Fig. S4.

We use the FLUXNET2015 dataset (49) to evaluate the simulated coupling. Among 206 available sites, we select 102 sites which have measurements of daily soil moisture, precipitation, and latent heat flux, needed to compute correlation coefficients. Days when at least one of the variables exhibits a quality control flag with a value below 0.75 are filtered out, conservatively. Since soil moisture in the root zone is not available at all sites, we compute correlation coefficients using surface soil moisture in FLUXNET2015 and simulated soil moisture in the soil layer closest to the measurement depth at each site for the simulations. Which measurement depths or soil layer depths are used to compute the correlation coefficients at each site are listed in SI Appendix, Table S1.

J.L. and C.H. designed research; performed research; analyzed data; and wrote the paper.

The authors declare no competing interest.