Weaker land–atmosphere coupling in global storm-resolving simulation

Significance The soil moisture–precipitation feedback plays an important role in shaping climate patterns over land. However, the sign of this feedback remains a topic of ongoing debate. We revisit the soil moisture–precipitation feedback using global, coupled simulations conducted at a grid spacing of 5 km. By avoiding the limitations associated with regional models or convective parameterization, our approach offers unique insights. The feedback in the 5-km simulation appears weaker and more negative compared to the one estimated from global coarse-resolution climate models. One potential implication of our findings is that the coarse-resolution climate model, with their strong positive feedback, may overestimate changes in droughts and heatwaves under future climate scenarios.


B. Preference for afternoon precipitation over soil moisture anomalies
Taylor et al. (3) and follow-up studies (4,5) have analyzed whether rain is more likely to occur over soils wetter or drier than their surrounding.We apply their approach to the output of our storm-resolving simulation.First, precipitation and soil moisture index are interpolated onto a common grid of 0.25 • .Then, an 1.25 • by 1.25 • event domain (Levt) is defined, centered at a cell location (Lmax) which coincides with a maximum in afternoon rainfall.Afternoon rainfall is defined as rainfall accumulated from 12:00 to 24:00.One or more cells are defined as Lmin where afternoon rainfall is at a minimum within Levt.In this step, we exclude Levt if it contains a cell with 1) morning precipitation (06:00-12:00) exceeding 1 mm day −1 , 2) topographic height variability exceeding 300 m, and 3) water bodies.We then compute morning (06:00-12:00) soil moisture index anomalies (S) by subtracting 31-day moving averaged soil moisture index centered at corresponding day.We compute the difference in soil moisture anomaly between Lmax and Lmin for event days as event sample, which is denoted as ∆Se.As the next step, we compute the difference in S between Lmax and Lmin using data from the same calendar month but for non-event days, denoted ∆Sc.This builds our control sample.
Afterward, all event and control samples within each 5 • by 5 • box are pooled together.The statistical significance is evaluated by assessing whether the differences between the two samples are significantly larger or smaller than the randomly generated differences.Before evaluating the significance, the climatology of the individual 5 • boxes is removed by subtracting June-July-August mean.Then, we compute δe = mean(∆Se) − mean(∆Sc) as the difference between event and control samples.
For the randomly generated difference, we compute δ = mean(∆ Ŝe) − mean(∆ Ŝc) where ∆ Ŝe is randomly pooled sample (the same size of the event sample) from ∆Se and ∆Sc, and ∆ Ŝc is randomly pooled sample from the remainder, and we repeat the computations 1,000 times (bootstrapping).Finally, we quantify the percentile of δe with respect to δ.For a more detailed description of the approach, please refer to previous studies (3)(4)(5).(g-h) the wettest year in LR, determined by the global mean soil moisture in June-July-August, and we use this to test whether using just one year, and one year which is wet, making it more similar to SRM, could explain the differences; and (i-j) a simulation (ICON-R2B6 AMIP) with the big leaf (BL) approach to test the effect of BL approach versus plant functional types (PFT) approach, but the simulation is Atmospheric Model Intercomparison Project (AMIP) type simulation with prescribed SST and higher grid spacing, which makes it more difficult to interpret.Top soil moisture is used to compute the correlation coefficient for two MPI-ESM simulations, as the field capacity and the permanent wilting point are not provided.Otherwise, soil moisture index in the root zone is used.Table S1.The descriptions of selected 102 FLUXNET2015 sites and corresponding soil layer depths of soil moisture for the coarse-resolution and storm-resolving models, used to compute correlation coefficients in Fig. 4.

Station name
Longitude(

Fig. S2 .
Fig. S2.Same as Fig. S1 but for time series of precipitation (P) and evapotranspiration (ET) for 60-year mean (thick orange for P and magenta for ET) and individual year (thin orange for P and magenta for ET) of the coarse-resolution model (LR), and the storm-resolving model (SRM, blue for P and green for ET).

Fig. S3 .
Fig.S3.The number of year when the correlation coefficient between soil moisture index-precipitation in the coarse-resolution model is smaller than that in the storm-resolving model.Areas where precipitation is smaller than 0.1 mm day −1 in both simulations are masked in grey.

Fig. S4 .
Fig. S4.Dependency of the correlation coefficient between soil moisture index (SMI) and subsequent mean precipitation (P) on the averaging period chosen to compute the mean precipitation for the coarse-resolution model (LR, red) and the storm-resolving model (SRM, blue).Dots indicate the averaging period of 9 days when the correlation coefficients for LR reach the maximum.

Fig. S5 .
Fig. S5.Correlation coefficients computed based on daily mean values for December-January-February between (a-c) soil moisture index and precipitation, (d-f) soil moisture index and evapotranspiration, and (g-i) evapotranspiration and precipitation for the 60-year mean of the coarse-resolution model (LR), the storm-resolving model (SRM), and the resulting difference in correlation coefficient.Areas where precipitation is smaller than 0.1 mm day −1 in both simulations or where the soil moisture index-precipitation correlation coefficient of the storm-resolving model is within one standard deviation of the year-to-year variability of the correlation coefficient in the coarse-resolution model are masked in grey.

Fig. S6 .
Fig. S6.Soil moisture-evaporation regime for (a) the coarse-resolution model (LR) and (b) the storm-resolving model (SRM).See definition in section A of SI Appendix.

Fig. S7 .
Fig. S7.(a-c) Leaf area index, (d-f) vegetation cover, and (g-i) field capacity for the coarse-resolution model (LR), storm-resolving model (SRM), and their difference.The averaging period for leaf area index is 60 years for LR and one year for SRM during June-July-August.With the exception of the leaf area index, two key soil and vegetation properties used for the computation of evapotranspiration in ICON are similar between the two models.

Fig. S9 .
Fig. S9.Correlation coefficient computed based on daily mean values for June-July-August between (a-c) precipitation and surface net radiation and (d-f) evapotranspiration and surface net radiation for (left) the 60-year mean of the coarse-resolution model (LR), (middle) the storm-resolving model (SRM), and (right) their difference.Gray areas as in Fig. S5.

Fig. S10 .
Fig. S10.Percentiles of δe with respect to δ indicating the preference for afternoon precipitation over soil moisture anomalies for June-July-August in the storm-resolving model (see definition in section B of SI Appendix).The lower percentile denotes that afternoon precipitation is triggered over dry soils and the higher percentile means afternoon precipitation is triggered over wet soils.Gray area denotes 5 • x 5 • cells containing fewer than 25 events sample.

Fig. S11 .
Fig. S11.Correlation coefficient between soil moisture and precipitation computed based on daily mean values for June-July-August and its difference from the storm-resolving model (SRM) for (a, b) ICON-ESM-LR.historical,which uses the same model version with the coarse-resolution model (LR, ICON-ESM-LR.piControl)but we use historical and the latest 15 years, to have years closer to present day, which is the year simulated in SRM and thus test whether the climate matters; (c-f) MPI-ESM1-2-LR at T64 (about 180 km) and MPI-ESM1-2-HR at T127 (about 100 km) for 30 years which use MPI-ESM that it is the predecessor of ICON-ESM and it shares the same land surface scheme and physical scheme, and we use them to test the effect of grid spacing but did not have two ICON-ESM runs we different resolution so that why we used MPI-ESM;(g-h) the wettest year in LR, determined by the global mean soil moisture in June-July-August, and we use this to test whether using just one year, and one year which is wet, making it more similar to SRM, could explain the differences; and (i-j) a simulation (ICON-R2B6 AMIP) with the big leaf (BL) approach to test the effect of BL approach versus plant functional types (PFT) approach, but the simulation is Atmospheric Model Intercomparison Project (AMIP) type simulation with prescribed SST and higher grid spacing, which makes it more difficult to interpret.Top soil moisture is used to compute the correlation coefficient for two MPI-ESM simulations, as the field capacity and the permanent wilting point are not provided.Otherwise, soil moisture index in the root zone is used.

Fig. S12 .
Fig. S12.Difference in the soil moisture index-precipitation coefficient between the new storm-resolving simulation (NGC3) for 5 years with a newer version of ICON model that uses difference land surface input parameters, which makes it less comparable, and the coarse-resolution model, computed across five years of the new storm-resolving model in comparison to the 60-year mean of the coarse-resolution model.