Constraining clouds and convective parameterizations in a climate model

19 Cloud and convective parameterizations strongly influence uncertainties in equilibrium 20 climate sensitivity (ECS). We provide a proof-of-concept study to constrain these 21 parameterizations in a perturbed parameter ensemble of atmosphere-only simulations by 22 evaluating model biases in the present-day runs using multiple satellite climatologies and 23 by comparing simulated δO of precipitation (δOp), known to be sensitive to 24 parameterization schemes, with a global database of speleothem δO records covering 25 the Last Glacial Maximum (LGM), mid-Holocene (MH) and pre-industrial (PI) periods. 26 Relative to modern, paleoclimate simulations show greater sensitivity to parameter 27 changes, allowing for an evaluation of uncertainties over a broader range of climate 28 forcing and the identification of parts of the world that are parameter sensitive. Certain 29 simulations reproduced LGM and MH δOp anomalies relative to the PI better than the 30 default parameterization. Not a single set of parameterizations worked well in all climate 31 states, thus improving simulations requires determining all plausible parameter 32

sensitive to parameter changes than traditional diagnostics such as precipitation and 85 temperature, likely related to cumulus entrainment strength (18). These models were 86 compared against modern water isotope observations from satellites (e.g., Aura 87 Tropospheric Emission Spectrometer (TES), (23)); Scanning Imaging Absorption 88 Spectrometer for Atmospheric Cartography (SCIAMACHY), (24)), providing a spatially 89 robust means of constraining model results. In a traditional PPE approach, models are not 90 typically re-tuned into radiative balance after altering a single tuning parameter (25), 91 which may have important implications in resolving or revealing biases from previous 92 compensating errors (26). However, not much is known whether this tuning approach 93 after each parameter change is preferable especially when considering a broader range of 94 climate states. 95 96 Variability in water isotopes may also be obtained from various paleoclimate archives 97 that are not only spatially well-distributed but are also available across timescales 98 drastically different from today, such as the Last Glacial Maximum (LGM; 21 ka, or kilo-99 years before present) and mid-Holocene (MH; 6 ka) periods. The LGM corresponds to a 100 time when global ice volume was at its maximum and greenhouse gas concentrations 101 were lower than today, both driving major changes in the atmosphere compared to 102 present conditions (27)(28)(29). During the MH, insolation is seasonally amplified in the 103 Northern Hemisphere, with larger winter-to-summer temperature differences and 104 associated changes in the hydrological cycle (30,31). Performing proxy-model 105 comparison across these contrasting time periods thus allows for evaluating model 106 performance over the full range of hydroclimatic variability in the Earth system. 107 108 One excellent source of past hydroclimatic information are speleothems. Speleothems are 109 secondary cave deposits that form from dissolution of carbonate bedrock through water 110 action. While their geographical distribution is largely constrained by the geology of a 111 region, speleothems form under a broad range of hydroclimatic regimes ideal for 112 investigating predominant regional patterns. Variations in speleothem 18 O largely 113 reflects the 18 O of soil ( 18 Os) and groundwater percolation, which in turn is heavily 114 influenced by 18 Op above the cave and other processes within the karst system (32, 33). 115 Early speleothem 18 O compilations and the more recently available Speleothem Isotope 116 Synthesis and Analysis (SISAL) database (34-36), a large global compilation of 117 speleothem isotope records since the last glacial, have aided in evaluating GCM 118 performance across the LGM and MH time periods (36-39) and have served as an 119 independent validation check in reconstructions of glacial temperature fields (40), 120 demonstrating their usefulness in benchmarking isotope enabled paleoclimate 121 simulations. However, not all parts of the world are equally influenced by cloud and 122 convective parameter changes, implying that proxy record locations may be more or less 123 constraining against simulations. This has not been fully quantified in existing 124 paleoproxy-model comparisons and/or analyses of model-satellite discrepancies both 125 globally and restricted to proxy sites only. 126 127 In this study, we explore cloud and convective parameterizations (Table 1) in the GISS-128 E2.1 climate model (41) that likely have a significant impact on water isotope distribution 129 and ECS. We use two sets of atmosphere-only simulations: one that has been re-tuned 130 This is a non-peer reviewed pre-print submitted to EarthArXiv. This manuscript has been submitted to Science Advances for peer review. into radiative equilibrium in the pre-Industrial (hereafter referred to as the balanced 131 version) and another which only changes the parameters (hereafter referred to as the 132 unbalanced version, see Materials and Methods), to evaluate whether this approach is 133 preferable in simulations of past climates with large differences in radiative forcing. We 134 investigate the variability and sensitivity of key climate variables to cloud and convective 135 changes and identify parameter-sensitive sites in the present-day (PD, year 2000) and 136 paleoclimate simulations covering the pre-industrial (PI, 0 ka), MH and LGM periods. 137 We also compare and evaluate the model simulations against multiple satellite 138 climatologies and assess the agreement between simulated 18 Op and speleothem 18  and LGM periods, respectively. From each of the models, we extracted the simulated 201 18 Op nearest each cave site. As shown in our proxy-model comparisons (Fig. 4), the 202 mean 18 Op distribution in all runs and time periods are in excellent agreement with the 203 proxies. In these comparisons, we prescribed weights to the simulated 18 Op, based on 204 Fig. 1, which gives importance to the spatial sensitivity of a particular site to parameter 205 changes. This significantly improved the overall proxy-model agreement compared to the 206 unweighted calculation ( fig. S6-a to -s and S7). 207 208 While these first order comparisons show excellent agreement, discrepancies remain; for 209 example, simulated 18 Op is more negative (positive) at low (mid-to high) latitude 210 speleothem sites compared to the proxies, with those from the LGM exhibiting the largest 211 offsets (Fig. 4). These discrepancies could be due to cave specific factors and model 212 limitations ( Nonetheless, certain simulations represent an improvement from the std run. The 224 entrainment rate for plume (entr20-80) parameterization exhibits the highest skill for the 225 PI period, whereas the convection adjustment time (tconvadjX2) parameterization best 226 represents cloud and convective processes for the MH and LGM periods. Considering 227 only the sites common across the time periods (i.e., limited by the number of LGM sites), 228 the entr20-80 parameterization became one of the poorest performing models for the PI 229 period. However, another entrainment rate scheme, entr60-40, emerged as the best 230 performing parameterization for PI. The tconvadjX2 parameterization remained the best 231 performing scheme for the MH, indicating that the reduced number of data points did not 232 affect the model evaluation for this time period. These results, broadly consistent with 233 best performers derived from satellite comparisons (considering only the proxy sites), 234 suggest that while different cloud and convective scheme settings do not necessarily 235 impose large changes on the model results for the sites considered, the best 236 parameterization for each time period varies depending upon the boundary conditions. 237 238 LGM and MH isotopic changes and model performance 239 To investigate the impact of parameter changes on the relative shift in 18 Op, we parameterization results (Fig. 7). Notably, this simulation also performed best in the 256 absolute value comparisons for the LGM period.

258
Compared to LGM variations, MH changes relative to PI are more modest. Interior South 259 America, India and Australia show positive 18 Op anomalies in the PPE mean (Fig. 6C). This is a non-peer reviewed pre-print submitted to EarthArXiv. This manuscript has been submitted to Science Advances for peer review.
found in North and Central America (South America) where positive (negative) 267 anomalies are not reflected in the simulated 18 Op changes. Overall, the magnitude of 268 change is consistently smaller in the simulations (Fig. 6D). Of the 19 simulations, only 9 269 PPE members show statistically significant (p < 0.04) relationship, outperforming the std 270 18 Op run (Fig. 7). The best performing parameterization is droprad130-50 (weighted r 2 271 = 0.11, Fig. 7), where 59% of the data points now share similar signs. Notable regions of 272 observed improvement are in Europe and Central Asia ( fig. S10). Reducing the number 273 of datapoints to match the sites from the LGM-PI changes shows a different result such 274 that the critQ2-4 parameterization now shows the highest skill (weighted r 2 = 0.45).

276 277
Discussion 278 In this study, we have identified parts of the world that are most sensitive to convective 279 and cloud parameterizations, which may provide the best opportunity for constraining 280 key metrics in climate models. Parameter-sensitive sites are different between the 281 balanced and unbalanced versions of the models with the latter showing more regions of 282 lower sensitivity scores. This is likely related to the greater variability among PPE 283 members induced by random changes in certain variable fields by the parameter 284 perturbations, affecting more indiscriminate regions in the world. This outcome from the 285 unbalanced version is less useful in constraining biases related to cloud and convective 286 parameterizations.

288
Our satellite-model analyses, stratified by global and proxy-specific skill scores, reveal 289 that the distribution of proxy sites here lie outside of the spatial domains most impacted 290 by cloud and convective parameterization choices. This suggests a need for additional 291 optimally suited sites distributed across more complex convection-cloud schemes to 292 constrain global simulations. Additionally, conducting these experiments using different 293 coupled atmosphere-ocean-vegetation models could provide an excellent framework for 294 targeted paleoclimate fieldwork to develop archives from these convective-and 295 parameter-sensitive areas across the world.

297
Though the proxy sites sample less complex atmospheric scenes, the first order spatial 298 pattern of 18 Op is in excellent agreement between proxy data and all PPE members 299 across all time periods. Also supported by the satellite analyses, two parameterizations 300 with highest model skill emerged: a 20:80 split of entrainment rate for plume (entr20-80) 301 for the PI period and doubled convection adjustment time (tconvadjX2) for the MH and 302 LGM periods. The simulations are able to capture broad scale LGM-PI 18 Op patterns 303 where tconvadjX2 parameterization performed best among parameterizations. On the 304 other hand, model skill is significantly reduced in the MH-PI runs where the magnitude 305 of change is consistently smaller in all simulations compared to the proxies. 306 307 It is highly likely that the coupled simulations of these same experiments will exhibit a 308 greater range of variability across simulations. The fixed SSTs in our runs allowed us the 309 ability to explore this approach with computationally inexpensive simulations; however, 310 it also throttles coupled feedbacks muting LGM and MH variability across ensemble 311 members and precluded us from calculating ECS for every perturbed parameter. Further, 312 This is a non-peer reviewed pre-print submitted to EarthArXiv. This manuscript has been submitted to Science Advances for peer review. these fixed surface ocean conditions limit the paleoclimate constraints to land-based 313 proxy archives. Other potential sources of model discrepancies are related to ice sheet 314 topography changes and dust concentrations (LGM), along with the lack of vegetation 315 and dust concentration feedbacks (LGM and MH) (46)(47)(48)(49), which may be best evaluated 316 using fully coupled atmosphere-ocean models. 317 318 Speleothem proxy climate records have their own set of uncertainties. Speleothem 18 O 319 primarily reflects local and regional climate signals controlling 18 Op. However, this 320 signal may be altered as it enters the soil zone and epikarst, a zone that stores infiltrated 321 rainwater, through mixing with existing waters, seasonality of recharge rates, and 322 fractionation by evaporation before reaching the cave system (50, 51 (70), and standard mean ocean water salinity and water isotopes (71) were made ( Table  372 2). All these runs were completed to surface equilibrium in GISS-E2. However, these experiments are computationally expensive, and beyond the scope of this 396 proof-of-concept study (but are planned in the future). The practical consequence is that 397 variability over the ocean especially is throttled, and the climate system during the 398 paleoclimate runs may no longer be in radiative equilibrium (a symptom the incomplete 399 climate response to the strong paleoclimate forcing perturbed parameter runs); we note 400 the net top of the atmosphere radiative balance of each simulation (Table 1).

402
The basic structure of the clouds and convection schemes are described in (72-74). We 403 have chosen here to explore six different parameters utilized in the cloud and convection 404 schemes that likely have a substantive impact on ECS as well as water isotope 405 distribution (Table 1). A total of 19 simulations were performed for each time period. 406 Parameters chosen are ones not directly constrained by current in situ or satellite 407 observing platforms. 408 409 Rain re-evaporation above the cloud base (rev) has been a parameter considered for 410 change previously because it improves convection and variability (e.g., Madden-Julian 411 Oscillation in (74)). This parameter makes the GISSE-2.2 model distinct from the 412 GISSE-2.1(75). Water isotopes are sensitive to changing this parameter (18). Increasing 413 this parameter results in additional atmospheric moistening and a subsequent increase in 414 precipitation over the Maritime Continent (i.e., increased bias); however, it does improve 415 isotopic matches between GISS-E2.1 simulations and satellite observations (23).

417
The entrainment rate (entr) parameters control how much environmental mass is 418 entrained into a less-and more-entraining convective plume. At most, two updraft 419 plumes are permitted to initiate at each model level in the GISS convective scheme, and 420 the only requirement is that they have different entrainment rates thus allowing a 421 representation of shallow (i.e., more entraining) and deep (i.e., less entraining) convective 422 towers within any convective cloud ensemble in the GCM grid box.

424
The convective adjustment time (tconvadj) is a parameter that controls how quickly 425 convective mass reaches the tropopause, and thus how quickly the environmental profile 426 of temperature and moisture adjusts to moist convective processes.

428
The convective trigger (ctrigger) parameter determines what environmental conditions 429 are necessary for initiating convection. Physically this parameter can be interpreted as 430 accounting for the multi-faceted role that the planetary boundary layer plays in 431 convective initiation (e.g., turbulent lifting of parcels, variations in near-surface stability 432 or moisture across a grid box), the role of vertical wind shear, the role of mesoscale 433 ascent causing local destabilization, or the role of gravity waves in the weakening of 434 convection-inhibiting stable layers.

436
The radius multiplier (droprad) is a parameter that governs the sizes of liquid droplets 437 and ice particles for a given condensate amount. Though there are some observational 438 estimates of sizes at cloud tops, within-cloud estimates are largely unconstrained (and 439 particularly within convection, where attenuation of radiometric signals are substantial). 440 In general, smaller sizes result in clouds reflecting more shortwave radiation coincident 441 with reduced outgoing longwave radiation.

443
Auto-conversion of cloud water content to precipitation is governed by a critical cloud 444 water content scaling parameter (critQ). Any liquid or ice water content above the scaled 445 critical threshold will be converted to precipitation via auto-conversion, thus affecting 446 cloud condensate, cloud fractions, and in turn, radiation. 447 448

Satellite data 449
This is a non-peer reviewed pre-print submitted to EarthArXiv. This manuscript has been submitted to Science Advances for peer review.  This is a non-peer reviewed pre-print submitted to EarthArXiv. This manuscript has been submitted to Science Advances for peer review. This is a non-peer reviewed pre-print submitted to EarthArXiv. This manuscript has been submitted to Science Advances for peer review.  LGM-PI.

871
This is a non-peer reviewed pre-print submitted to EarthArXiv. This manuscript has been submitted to Science Advances for peer review.