-
In this study, we examine soil moisture−atmosphere feedbacks in 18 historical CMIP6 simulations (Eyring et al., 2016). Table 1. summarises the details of each simulation, including atmospheric resolution and land surface model. CMIP6 experiments were selected based on the availability of sub-daily data of precipitation, surface-layer soil moisture and surface energy balance components. The CMIP6 model outputs used in this study are resolved at two temporal resolutions: precipitation, convective precipitation and surface fluxes are outputted as three-hourly means; whilst surface-layer soil moisture and surface air pressure are diagnosed instantaneously every three hours. Unless stated, we only analyse model outputs from a single ensemble member (r1i1p1f1 or lowest available), between years 1980 to 2014, and during boreal summer months (June to August; JJA).
Model name Institution Horizontal resolution Vertical resolution Land surface model ACCESS-CM2 CSIRO-ARCCSS N96; 192 × 144; 250 km 85 levels to 85 km CABLE 2.5 ACCESS-ESM1-5 CSIRO N96; 192 × 144; 250 km 85 levels to 85 km CABLE 2.4 with biogeochemistry BCC-CSM2-MR BCC T206; 320 × 160; 100 km 46 levels to 1.46 hPa BCC-AVIM2 CNRM-CM6-1 CNRM-CERFACS T127; 384 × 192; 150 km 91 levels to 0.01 hPa ISBA-CTRIP CNRM-ESM2-1 CNRM-CERFACS T127; 384 × 192; 150 km 91 levels to 0.01 hPa Surfex 8.0c EC-Earth3-Veg EC-Earth consortium TL255; 512 × 256; 100 km 91 levels to 0.01 hPa HTESSEL GFDL-CM4 NOAA-GFDL C96; 360 × 180; 100 km 33 levels to 1 hPa GFDL-LM 4.0.1 GISS-E2-1-G NASA-GISS C48; 144 × 90; 250 km 40 levels to 0.1 hPa GISS LSM HadGEM3-GC31-LL MOHC N96; 192 × 144; 250 km 85 levels to 85 km JULES-HadGEM3-GL7.1 HadGEM3-GC31-MM MOHC N216; 432 × 324; 100 km 85 levels to 85 km JULES-HadGEM3-GL7.1 HadGEM3-GC31-HM MOHC N512; 1024 × 768; 50 km 85 levels to 85 km JULES-HadGEM3-GL7.1 IPSL-CM6A-LR IPSL N96; 192 × 144; 250 km 79 levels to 40 km ORCHIDEE KACE-1-0-G NIMS-KMA N96; 192 × 144; 250 km 85 levels to 85 km JULES-HadGEM3-GL7.1 MIROC6 MIROC T85; 256 × 128; 250 km 81 levels to 0.004 hPa MATSIRO6.0 MPI-ESM1-2-HAM HAMMOZ consortium T63; 192 × 96; 250 km 95 levels to 0.01 hPa JSBACH 3.20 MPI-ESM1-2-HR MPI-M T127; 384 × 192; 100 km 95 levels to 0.01 hPa JSBACH 3.20 MPI-ESM1-2-LR MPI-M T63; 192 × 96; 250 km 47 levels to 0.01 hPa JSBACH 3.20 SAM0-UNICON SNU C96; 288 × 192; 100 km 30 levels to $ \approx$ 2 hPa CLM 4.0 Table 1. CMIP6 models used in this study. Third and fourth columns show the horizontal and vertical resolution of the model's atmospheric component. We follow the typical convention of the modelling institution in stating the model resolution. “T” and “TL” denote spectral models with a triangular truncation with an “L” signifying models with a linear Gaussian grid. “C” refers to a cubed-sphere finite volume model, whilst an “N” prefix is used before stating the total number of two-gridpoint zonal waves that can be represented. Following the grid specification, the dimensions of the model output on a Gaussian longitude/latitude grid is given alongside the stated nominal resolution from Taylor et al. (2017).
-
Building on the analysis performed by Talib et al. (2021), in section 4.1 we assess the behaviour of simulated surface fluxes during dry spells. To evaluate the simulated surface energy balance, we approximate real world radiative and turbulent fluxes through amalgamating weather station measurements and satellite-based observations. Here we provide a brief overview of our methodology to derive surface energy balance components, however more detail can be found in Talib et al. (2021).
Six-hourly data from 49 weather stations above 3000 m from the China Meteorological Administration (CMA), locations later shown in Fig. 2u, is used to approximate surface sensible heat flux (SHF, W m−2) and upward longwave radiation (LWup, W m−2). Using measurements of surface temperature (
$T_{\rm{s}}$ , K), near-surface air temperature ($T_{\rm {a}}$ , K) and 10 m wind speed (${v}_{10{\rm{m}}}$ , m s−1), we estimate the surface SHF using a bulk formula:where
$ C_p $ is the specific heat capacity of dry air at constant pressure (1005 J kg−1 K−1);$ \rho $ is density (kg m−3) and decreases exponentially with height; and$C_{\rm {DH}}$ is the drag coefficient for heat [assumed to be 4.0$ \times $ 10−3 for all stations following Duan and Wu (2008)]. We compute the outgoing surface LWup using the Stefan-Boltzmann equation:where
$ \epsilon $ is the surface emissivity (assumed here to be fixed at 0.95) and$ \sigma $ is the Stefan-Boltzmann constant (5.67 × 10−8 W m−2 K−4). We then combine computed surface SHF (Eqn. 1) and LWup (Eqn. 2) with radiative surface fluxes derived from the Clouds and the Earth's Radiant Energy System (CERES; Loeb et al., 2003) to partition the surface energy balance. Radiative surface fluxes are outputted on a$ 1^\circ $ latitude$ \times $ $ 1^\circ $ longitude grid. For each station, fluxes are selected from the nearest CERES grid point. The following equation is formulated after partitioning the surface energy balance into land surface forcings (left hand side) and surface fluxes that depend on land surface characteristics (right hand side):where SWnet denotes the net-downward shortwave radiation (W m−2); LWdown denotes the downward longwave radiation (W m−2); LHF denotes the surface latent heat flux (W m−2); and G denotes the ground heat flux (W m−2). To minimise errors associated with the spatial misalignment between in situ observations and gridded satellite products, we only analyse station-mean anomalies relative to monthly climatologies. If we assume that surface albedo changes are minimal, only components on the right-hand side of (Eq. 3) depend on changes in surface characteristics. Upon subtracting SHF and LWup from surface radiation (SWnet + LWdown), the remainder is assumed to be a combination of LHF and G.
Approximated observed surface fluxes are calculated instantaneously every six hours, whilst simulated fluxes are outputted as three-hourly means (i.e. 0900−1200, 1200−1500 UTC etc.; section 2.1). To enable a suitable comparison between observed and simulated surface fluxes during dry spells, we perform a temporal cubic interpolation of simulated surface fluxes. To do so we assume, for example, that the three-hourly mean between 0600 and 0900 UTC is an approximation of the instantaneous value at 0730 UTC. To then estimate the value at 0600 UTC, we perform a cubic interpolation of three-hourly mean surface fluxes centered at 0430 and 0730 UTC. With regards to referencing the local time of day, we conclude it is inappropriate to use Beijing time (BT) as a reference for local solar conditions on the TP as it covers a large longitudinal range. Instead, we define local solar time (LST) as six hours ahead of UTC as the eastern TP is situated at approximately
$ 90^\circ $ longitude. -
In this study, soil moisture-precipitation feedbacks are defined by the complete pathway with which soil moisture variations lead to precipitation changes through fluctuations in the partitioning of surface turbulent fluxes. Components of soil moisture-precipitation feedbacks are quantified through comparing covariances of evaporative fraction (EF, dimensionless), surface soil moisture (SM, m3 m−3) and precipitation (P, mm d−1). This follows multiple studies which have quantified components of land−atmosphere feedbacks through comparing covariances between surface and atmospheric fields (i.e. Dirmeyer, 2011; Dirmeyer et al., 2014; Müller et al., 2021b). Our coupling metrics are derived from Dirmeyer et al. (2014) and are the same as in Müller et al. (2021b), except for the use of EF instead of LHF. The use of EF removes the dependence of surface turbulent fluxes on radiation, and instead focuses on the partitioning of turbulent fluxes.
The terrestrial leg of soil moisture-precipitation feedbacks, which identifies areas where anomalous soil moisture drives surface flux variability, is quantified by a terrestrial coupling index (TCI, dimensionless):
where
${\rm{cov}}(x,y)$ and$ \sigma(x) $ denote the covariance between two variables and the temporal standard deviation of single variable respectively. Based on this definition, strong soil moisture-surface flux coupling is quantified in regions where soil moisture conditions drive evapotranspiration dynamics. Following this, the atmospheric component of soil moisture-precipitation feedbacks, which highlights regions where surface flux changes lead to a precipitation response, is defined by an atmospheric coupling index (ACI, mm d−1):ACI complements TCI through highlighting areas where the partitioning of surface turbulent fluxes impacts precipitation. Whilst previous studies have used more intermediate atmospheric variables to compute coupling indices, such as the lifting condensation level (Dirmeyer et al., 2014), we use precipitation as it ensures that the full cycle of land−atmosphere coupling is analysed given the direct feedback between precipitation and surface conditions (Müller et al., 2021b). Even though using precipitation will likely result in a weaker concluded coupling between surface fluxes and atmospheric conditions, our definition of ACI represents the full cycle in surface flux-precipitation coupling, which is not guaranteed when using intermediate atmospheric variables. Precipitation is also one of only suitable atmospheric diagnostics regularly outputted at a sub-daily temporal resolution.
Regions with strong soil moisture-precipitation feedbacks are identified using a two-legged coupling index (TLCI, mm d−1):
TLCI quantifies the anomalous precipitation influenced by moisture−driven surface flux variability and is derived through combining TCI and ACI. Hot spots with substantial soil moisture-precipitation coupling are regions where both TCI and ACI are large. In these regions we expect a feedback from the atmosphere to the land, completing the mechanistic loop (Guo et al., 2006). Partitioning TLCI into its two components highlights locations where the relationship between soil moisture and evaporative fraction is strong (TCI), and where anomalous evaporative fraction influences precipitation (ACI).
For calculating all three indices we use daily time series of anomalies relative to a monthly climatology. To ensure that we sample the impact of surface conditions on daytime precipitation, we compute coupling metrics using three-hourly means of surface SM and EF between 06 and 09 LST. We also only analyse precipitation accumulations between 09 and 18 LST. In principle, one could compare simulated indices with those computed using reanalysis data, however in practice, due to the use of complex land surface models and an inadequate representation of orographic precipitation (Tong et al., 2014; Hu and Yuan, 2021; Müller et al., 2021a), it is unreliable to evaluate simulated metrics with those calculated using reanalysis data. It is also challenging to obtain reliable surface flux observations across the whole of the TP. Therefore, we do not compute coupling metrics using observations.
-
To evaluate the simulated feedback between soil moisture and daytime rainfall, we use a metric derived by Taylor et al. (2012). For the rest of this study this metric is referred to as “T12” and denoted by
$\delta_{\rm e}$ . T12 quantifies soil moisture-rainfall coupling by assessing anomalous antecedent soil moisture differences between locations with daytime precipitation maxima and minima. The metric was originally developed for global applications and has been used to diagnose soil moisture-rainfall coupling in both observations and models (Taylor et al., 2012, 2013). T12 is computed using:where
$\overline{{\vartriangle}S_{\rm e}}$ is the composite-mean difference in pre-rainfall soil moisture anomalies between locations of maximum and minimum rainfall, and$\overline{{\vartriangle}S_{\rm c}}$ is a control sample of typical soil moisture anomaly mean differences between those locations. Locations of maximum and minimum rainfall are identified from accumulated convective precipitation between 0900 to 1800 LST. The inclusion of rainfall between 0900 to 1200 LST accounts for the early diurnal onset bias in simulated precipitation (Christopoulos and Schneider, 2021). A 3 × 3 pixel box is centered on an afternoon convective precipitation event with rainfall exceeding 2 mm. The minimum is located within the 3 × 3 pixel box. We decided to use a lower precipitation threshold than Taylor et al. (2012), 2 mm compared to 3 mm, due to low convective precipitation rates in several models. Where two boxes overlap, the box containing the more intense maxima is retained. If there is more than one minima within a 3 × 3 pixel box, the average soil moisture anomaly is taken. Soil moisture anomalies are sampled at 0600 to 0900 LST. If the total precipitation for sampled pixels exceeds 0.1 mm during this or the preceding time-step (0300 to 0900 LST), the event is excluded. This ensures that only pre-event soil moisture is sampled. Daily soil moisture anomalies are generated with respect to a 35-year (1980–2014) monthly climatology. For each event, soil moisture anomalies for the control sample are taken from the same day of year in non-event years.$\delta_{\rm e}$ is expressed as a percentile of typical$ \delta $ values derived from random reassignment of$ {\vartriangle}S $ values. Percentile values less than 10 denote a negative feedback at a 10% significance level or lower, whilst those greater than 90 indicate a positive feedback. Strong negative and positive feedbacks at a 1% signficance level or lower are indicated by percentile values less than 1 and greater than 99 respectively. Whilst a combination of high topographic complexity and poor quality remotely-sensed soil moisture data over the TP means that we cannot directly compare simulated T12 metrics with observations, we can compare the sign of the simulated metric with more detailed analysis from Barton et al. (2021). -
Coupling metrics illustrate large inter-model differences in simulated soil moisture-precipitation feedbacks across the TP (section 3.). Whilst these metrics provide a good overview of the coupling strength between soil moisture and precipitation, they can be influenced by atmospheric or rainfall persistence (Guillod et al., 2015). It is also challenging to understand the processes responsible for different coupling characteristics. In light of this, for the rest of this study we gain insight from examining observational metrics specifically designed for the TP. In the following subsection we evaluate the surface flux and atmospheric response to regional-scale dry spells. After this we investigate the sensitivity of convective precipitation to soil moisture heterogeneity (section 4.2).
-
To assess simulated surface fluxes during three-day dry spells, we first show the behaviour of the surface energy balance in the real world. In this study, we define a regional dry spell as a period of three consecutive days when the regional-mean precipitation rate is below the non-zero 20th percentile boreal summer daily rainfall accumulation. To identify real world regional-scale dry spells we use station-mean daily precipitation accumulations at 1200 UTC from CMA weather stations (Fig. 2u). Due to the time span of weather station data (section 2.2), we only composite real world dry spells between 2000 and 2014. Figure 3a shows surface flux variations across the TP during observed three-day dry spells. Day 0.0 is defined as the start of a three-day regional dry spell, whilst anomalies are only shown at 1200 LST as this is the time of day with the largest surface flux response. Unsurprisingly, a dry spell across the TP increases downward radiation into the surface due to reduced cloud cover. Surface drying during a dry spell changes the partitioning of this enhanced radiation with LHF decreasing by approximately 60 W m−2 and SHF increasing by approximately 40 W m−2 between days 0 to 2. We also observe increased LWup by approximately 30 W m−2 associated with increased surface temperatures.
We next compare this observed surface flux behaviour to that exhibited in CMIP6 experiments. To identify simulated dry spells, we use regional-mean daily precipitation accumulations on grid points above 3000 m between 25° to
$ 40^\circ $ latitude and 85° to$ 105^\circ $ longitude, region denoted by a red rectangle in Fig. 2u. Whilst for observations we were only able to composite dry spells between 2000 and 2014, for CMIP6 experiments we use data from 1980 to 2014 to increase the number of composited dry spells. Even though each GCM simulates a reasonable number of three-day dry spells when using its own 20th percentile precipitation rate, all of the simulated precipitation thresholds are larger compared to observations, which is unsurprising given positive precipitation biases (Fig. 2). When selecting simulated dry spells using the observed precipitation threshold, only eleven out the eighteen GCMs have a substantial number of dry spells ($\geqslant$ 20). Due to the smaller number of individual models with a substantial number of dry spells when using the observed threshold, we focus on the anomalous surface energy balance during dry spells that are defined using simulated 20th percentile precipitation rates. Given that simulated dry-spell precipitation rates are greater than observations, we expect simulated surface flux variations to be dampened.Figures 3b to 3e highlight the variety of model behaviours in the CMIP6 ensemble by focusing on BCC-CSM2-MR, HadGEM3-GC31-HM, MIROC6 and MPI-ESM1-2-HR. For CMIP6 experiments we are able to composite simulated latent heat fluxes, whilst for observations, we approximate the sum of latent and ground heat fluxes. Whilst all four chosen models simulate increased downward surface radiation, associated with clear skies, they all have a different surface energy balance response. All four chosen models simulate increased SHF and LWup, however, changes in these surface energy balance components are typically underestimated. Only BCC-CSM2-MR well represents changes in SHF, whilst anomalies in HadGEM3-GC31-HM, for example, are 50% smaller compared to observations. We also find that three of our chosen models exhibit small latent heat flux changes during a dry spell compared to observations, indicating that the surface dries more rapidly in the real world. Whilst we do see a latent heat flux decrease in MPI-ESM1-2-HR of a similar magnitude to observations, the reduction in evapotranspiration takes several days longer.
Figure 3. Anomalous surface fluxes (W m−2) and daily precipitation accumulations (mm d−1) preceding, during and after three-day regional dry spells in (a) observations, (b) BCC-CSM2-MR, (c) HadGEM3-GC31-HM, (d) MIROC6 and (e) MPI-ESM1-2-HR. A three-day regional dry spell is defined when the precipitation accumulation is smaller than the twentieth percentile of boreal summer daily precipitation, denoted by the blue dashed horizontal line, for three consecutive days. We show the following components of the surface energy balance: (orange) upward longwave radiation; (purple) sensible heat flux, (black) and sum of net-downward shortwave and longwave downward radiation. In panel (a) the red line denotes the sum of latent and ground heat fluxes, whilst for panels (b) to (e) it denotes the latent heat flux only. Panels (b) to (e) also include a (dashed grey) “residual” term which is the remainder when subtracting sensible and latent heat fluxes from net-downward radiation. Box-and-whisker plots show station-mean or regional-mean daily precipitation accumulations. The orange line within each box denotes the median; the top and bottom of the box denotes the upper and lower quartiles; and the blue whiskers denote the 10th and 90th percentiles. Filled blue circles denote outliers in precipitation rates.
Following on from focusing on four chosen models, Figs. 4a to 4c show the average anomalous surface SHF, LWup, and radiation reaching the surface during observed and simulated three-day dry spells. Consistent with the subset of models analysed in Fig. 3., the majority of models underestimate increases in SHF and LWup (Figs. 4a and 4b). For example, by day 2 of a three-day dry spell, the anomalous model-mean bias in SHF is approximately 10 W m−2 smaller than observations, whilst anomalous LWup is underestimated by approximately 20 W m−2. CMIP6 simulations better represent the amplitude of surface radiation anomalies across the TP during dry spells (Fig. 4c). This indicates that errors in cloud representation are less of a concern compared to surface flux dynamics. Given that it may be the case that changes in surface SHF and LWup are poorly simulated due to underestimated anomalous radiation, we compute the fraction of radiation inputted into the surface that is re-emitted as SHF or LWup. In this study, this term is referred to as the “fraction of downwelling radiation”:
Figure 4. Anomalous surface (a) sensible heat flux, (b) upward longwave radiation, (c) radiation inputted into the surface, and (d) fraction of downwelling radiation that is re-emitted as sensible heat and upward longwave radiation, preceding, during and after a three-day dry event. All models from the same model family are denoted by the same line colour with individual configurations distinguished by marker style. Observations are denoted by a green line whilst the model mean is denoted by a black line. The model mean for simulated dry spells with the observed precipitation threshold is denoted by a dashed black line. The values to the right of each model name in the legend include the regional-mean 20th percentile boreal summer rainfall rate and the number of three-day dry spells identified.
Figure 4d shows the change in the anomalous fraction of downwelling radiation during a dry spell in both observations and CMIP6 simulations. The increased fraction of downwelling radiation during a dry spell in observations illustrates that in the real world changes in surface characteristics lead to anomalous sensible heating and surface longwave emission. However, all CMIP6 simulations underestimate the change in partitioning of incoming radiation towards SHF and LWup. This difference between observations and CMIP6 simulations highlights that surface dynamics are poorly represented on the TP during a dry spell, and that errors in the surface energy balance are not solely due to biases in atmospheric radiation.
In comparison with observations, all CMIP6 simulations poorly represent the favouring of SHF over evapotranspiration during a dry spell. However, it may be the case that this weak surface response during simulated dry spells is associated with high dry-spell precipitation rates. For example, Fig. 3 shows larger anomalies in SHF and LWup in models with smaller dry-spell precipitation rates (BCC-CSM2-MR and HadGEM3-GC31-HM). To investigate the hypothesis that anomalous surface fluxes are small in CMIP6 experiments due to large dry-spell precipitation rates, Fig. 5a compares the average change in SHF and LWup during dry spells with the prescribed precipitation threshold. We also calculate the linear least-squares regression between simulated values and take note of the Pearson correlation coefficient and p-value for a single-sided t-test assuming a negative relationship. Whilst one would expect excessive dry spell rainfall to suppress the surface flux response, the slope of the linear regression is not significantly negative. This provides evidence that an improved representation of anomalous surface fluxes during a dry spell requires more than just a better representation of dry-spell precipitation intensities. To analyse our hypothesis further, Fig. A2 shows composited surface flux anomalies during simulated dry spells using the observed precipitation threshold. As discussed previously, only a selection of models simulate a substantial number of dry spells when using the observed precipitation threshold. The model-mean surface flux response during these dry spells is also denoted in Fig. 4 by black dashed lines. Whilst Figs. 4 and A2 illustrate that using the observed precipitation threshold leads to a better simulation of anomalous SHF and LWup, models still do not adequately capture the strong change in surface flux partitioning. This indicates that land surface schemes in GCMs are unable to represent soil moisture−driven short-term (
$ \approx $ few days) variability in evapotranspiration on the TP.Figure 5. (a) Comparison of boreal summer 20th percentile precipitation rate (mm d−1) against the change in surface sensible heat flux and upward longwave radiation (W m−2) at 12 LST between days −0.25 and 2.75 of a regional dry spell. The grey dashed line in panel (a) denotes the linear least-squared fit between simulated values. The line's Pearson correlation coefficient value (R) and p-value for a single-sided t-test assuming a negative relationship (P) is shown in the top right hand corner. (b) Anomalous surface pressure (hPa) between 25° to 40° latitude and 85° to 105° longitude preceding, during and after a three-day dry spell. Dashed grey horizontal and vertical lines in panel (b) denote the zeroth value. All models from the same model family are denoted by the same line colour with individual configurations distinguished by marker style. Observations are denoted by green circular markers, whilst the model mean is denoted by black stars.
The inter-model variability in the behaviour of surface fluxes during dry spells is consistent with differences in TCI values (Figs. 1 and 4). Models which simulate unrealistic large dry-spell precipitation rates, such as MIROC6 and ACCESS-ESM1-5 (Fig. 5a), simulate relatively weak TCI values. In these simulations it is likely that the dry-spell precipitation rate is greater than potential evapotranspiration. This may lead to an unrealistic representation of the land surface as it rarely dries out and uses all additional radiative energy to increase evapotranspiration. We also find that models with a relatively large change in the fraction of downwelling radiation during dry spells, such as MPI-ESM2-2-HR (Fig. 4d), are typically those with high TCI values. Whilst our analysis of surface fluxes during three-day dry spells cannot fully explain simulated TCI differences due to influence of variability on longer timescales, we find a good agreement between the magnitude of TCI values simulated and the response of surface fluxes during a dry spell.
In the real world surface drying on the TP favours sensible heat, a deeper planetary boundary layer, and a negative near-surface pressure tendency (Wan et al., 2017; Talib et al., 2021). Due to the lack of a surface energy balance response to regional dry spells in the majority of GCMs (Fig. 4), we predict that models underestimate the surface pressure response to surface drying. To illustrate the favouring of a heat low circulation across the TP during dry spells in observations, Fig. 5b shows regional-mean anomalous surface pressure from the European Centre for Medium-Range Weather Forecasts (ECWMF) Reanalysis version 5 (ERA5; Hersbach et al., 2020) at a 1° resolution between 25° to 40° latitude and 85° to 105° longitude. The negative pressure tendency associated with surface drying maximises after sunset which is indicative of a heat low circulation as negative pressure tendencies are limited until a stable boundary layer has formed. Fig. 5b also shows the anomalous surface pressure during simulated dry spells in each GCM. Whilst most models correctly simulate diurnal fluctuations in anomalous surface pressure, the magnitude of pressure tendencies during a dry spell are smaller compared to observations. For example, the model-mean surface pressure anomaly decreases by approximately 0.7 hPa from days 0.0 to 3.0, compared to 1.3 hPa in observations. We also observe distinct pressure anomalies in BCC-CSM2-MR and IPSL-CM6A-MR, which we associate with synoptic forcing dominating any effects from surface heating. The simulation of weaker pressure tendencies in CMIP6 experiments is consistent with underestimated changes in surface energy balance components. The weak surface pressure response is likely to limit the impact of soil moisture-atmospheric coupling on large-scale circulation anomalies (Wan et al., 2017; Talib et al., 2021).
-
In the previous subsection we show that the dampened behaviour of surface fluxes during dry spells in CMIP6 simulations leads to a misrepresentation of surface-induced atmospheric pressure fluctuations. In this subsection we investigate whether CMIP6 models correctly favour deep convection initiation over dry soils, as observed by Barton et al. (2021). To do this, we first analyse simulated pre-rainfall soil moisture anomalies before a strong convective precipitation event (section 2.3.2). We pool all events that occur within 80° to
$102^\circ$ longitude and 28° to$ 40^\circ $ latitude, as denoted by the grey box in Fig. 2v, where elevation exceeds 2500 m. The T12 metric for 12 out of the 18 GCMs is shown in Fig. 6, with the order of GCMs determined by horizontal resolution (increasing from left to right). For the remaining six CMIP6 simulations, an insufficient number of convective precipitation events (< 100) were identified due to either persistent early-morning rainfall (GISS-E2-1-G, IPSL-CM6A-LR, CNRM-CM6-1 and CNRM-ESM2-1) or minimal precipitation rates (GFDL-CM4 and BCC-CSM2-MR). These models are also those with relatively low values of deep convective precipitation (Fig. A3). Almost all remaining GCMs fail to capture the observed negative feedback between soil moisture and deep convection found in Barton et al. (2021). Only ACCESS-CM2 simulates a significant strong negative feedback, whilst six GCMs show a significant strong positive feedback.Figure 6. T12 metric (
$ \delta_e $ , percentile) for events within 80° to 102° longitude and 28° to 40° latitude for CMIP6 simulations. For clarity, bars are plotted as a deviation from 50 where values larger and smaller than 50 denote positive and negative feedbacks respectively. Blue and red filled bars denote a preference for afternoon convection over wet and dry soils respectively with a significance level below 10%. Grey hatched bars denote models with fewer than 100 events. Blue and orange horizontal dashed lines denote the 10% significance level for positive and negative feedbacks respectively, whilst the grey dashed vertical line partitions low- and medium-resolution models. Italicised values above each bar denote the number of afternoon convective events used to calculate the metric.Given that the majority of rainfall in CMIP6 experiments with a substantial number of events is associated with deep convection (Fig. A3) and most experiments simulate positive values of TCI and ACI (Fig. 1), it is unsurprising that a positive soil moisture-deep convection feedback is seen in most models (Fig. 6). The lack of a simulated negative feedback is similar in other semi-arid regions (Taylor et al., 2012) and can be explained by a typically strong dependence of convective parameterisations on low-level humidity, which is favoured across wet soils. In reality, convective initiation occurs later in the day than in climate models (Christopoulos and Schneider, 2021), which favours stronger daytime mesoscale circulations and more intense heavy precipitation. We might expect that increasing the horizontal resolution of a climate model improves the model's ability at simulating the formation of realistic circulations in response to soil moisture heterogeneity. However, if the model’s convection scheme is typically triggered before these circulations can develop, then a positive feedback may persist, irrespective of resolution, as illustrated in Taylor et al. (2013).
Considering all twelve GCMs for which we were able to compute the T12 metric, there is no obvious improvement with resolution (Fig. 6, left to right). However, if we compare different resolutions within the same model family, i.e. MPI-ESM1-2-HR with MPI-ESM1-2-LR and HadGEM3-GC31-HM with HadGEM3-GC31-MM, increased resolution decreases the value of the T12 metric and weakens the positive feedback. To examine this behaviour in more detail, we focus on HadGEM3-GC31 simulations for which we have a sufficient number of events at all three resolutions to subdivide the TP into four
$ 11^\circ $ longitude$\times \;6^\circ $ latitude quadrants (regions shown in Fig. 7). For the low (HadGEM3-GC31-LL) and medium (HadGEM3-GC31-MM) resolutions, a significant positive feedback is simulated in all four quadrants (Figs. 7a and 7b). HadGEM3-GC31-HM on the other hand, simulates varying behaviour across the TP with a negative/positive feedback in the south-east/north-east quadrant (Fig. 7c). Comparing simulated feedbacks with resolved topography (Figs. 7d to 7f) gives some indication that increased topographic complexity influences the sign of the feedback. To investigate whether topographic complexity in the model influences the feedback between soil moisture and deep convection in more detail, we partitioned all events in HadGEM3-GC31-HM into two groups based on the grid-scale topographic complexity. The grid-scale topographic complexity is calculated as the standard deviation in altitude across a 3 × 3 pixel which is centered on a rainfall event. As shown in Table 2, events centered where topographic complexity is low have a weak negative feedback, whilst where topographic complexity is high, a strong positive feedback is concluded. This indicates that when increasing resolution, and hence improving the representation of orographic convection, we begin to favour negative soil moisture-convection feedbacks across regions with low topographic complexity. It is known that convection-permitting resolutions are needed to fully capture soil moisture-convection feedbacks (Hohenegger et al., 2009; Taylor et al., 2013), but these configurations are currently too expensive to run globally across climate relevant time scales. The fact that HadGEM3-GC31-HM, a current medium-resolution global climate model, can begin to overcome a significant feedback bias on the TP is promising for future generations of ESMs.Figure 7. (a−c) T12 metric (
$ \delta_e $ , percentile) over 11° longitude × 6° latitude quadrants and (d−f) resolved topography (km) for (a, d) HadGEM3-GC31-LL, (b, e) HadGEM3-GC31-MM and (c, f) HadGEM3-GC31-HM. In panels (a) to (c) blue and red shading denotes a preference for afternoon convection over wet and dry soils respectively. Coloured shading is only applied in quadrants with a significance level below 10%. The grey outline of the TP denotes an elevation of 2500 m.Topographic complexity Number of events $ \delta_e $ Low 1842 10 High 1956 98 Table 2. Evaluation of topographic dependence on soil moisture-deep convection coupling over TP in HadGEM3-GC31-HM. Regions with low and high topographic complexity have a standard deviation in altitude below or above 100 m, respectively, over a 3 × 3 pixel box. Blue and red shading denotes a preference for afternoon convection over wet and dry soils respectively.
Model name | Institution | Horizontal resolution | Vertical resolution | Land surface model |
ACCESS-CM2 | CSIRO-ARCCSS | N96; 192 × 144; 250 km | 85 levels to 85 km | CABLE 2.5 |
ACCESS-ESM1-5 | CSIRO | N96; 192 × 144; 250 km | 85 levels to 85 km | CABLE 2.4 with biogeochemistry |
BCC-CSM2-MR | BCC | T206; 320 × 160; 100 km | 46 levels to 1.46 hPa | BCC-AVIM2 |
CNRM-CM6-1 | CNRM-CERFACS | T127; 384 × 192; 150 km | 91 levels to 0.01 hPa | ISBA-CTRIP |
CNRM-ESM2-1 | CNRM-CERFACS | T127; 384 × 192; 150 km | 91 levels to 0.01 hPa | Surfex 8.0c |
EC-Earth3-Veg | EC-Earth consortium | TL255; 512 × 256; 100 km | 91 levels to 0.01 hPa | HTESSEL |
GFDL-CM4 | NOAA-GFDL | C96; 360 × 180; 100 km | 33 levels to 1 hPa | GFDL-LM 4.0.1 |
GISS-E2-1-G | NASA-GISS | C48; 144 × 90; 250 km | 40 levels to 0.1 hPa | GISS LSM |
HadGEM3-GC31-LL | MOHC | N96; 192 × 144; 250 km | 85 levels to 85 km | JULES-HadGEM3-GL7.1 |
HadGEM3-GC31-MM | MOHC | N216; 432 × 324; 100 km | 85 levels to 85 km | JULES-HadGEM3-GL7.1 |
HadGEM3-GC31-HM | MOHC | N512; 1024 × 768; 50 km | 85 levels to 85 km | JULES-HadGEM3-GL7.1 |
IPSL-CM6A-LR | IPSL | N96; 192 × 144; 250 km | 79 levels to 40 km | ORCHIDEE |
KACE-1-0-G | NIMS-KMA | N96; 192 × 144; 250 km | 85 levels to 85 km | JULES-HadGEM3-GL7.1 |
MIROC6 | MIROC | T85; 256 × 128; 250 km | 81 levels to 0.004 hPa | MATSIRO6.0 |
MPI-ESM1-2-HAM | HAMMOZ consortium | T63; 192 × 96; 250 km | 95 levels to 0.01 hPa | JSBACH 3.20 |
MPI-ESM1-2-HR | MPI-M | T127; 384 × 192; 100 km | 95 levels to 0.01 hPa | JSBACH 3.20 |
MPI-ESM1-2-LR | MPI-M | T63; 192 × 96; 250 km | 47 levels to 0.01 hPa | JSBACH 3.20 |
SAM0-UNICON | SNU | C96; 288 × 192; 100 km | 30 levels to $ \approx$ 2 hPa | CLM 4.0 |