A dynamical downscaling projection of future climate change in the Laurentian Great Lakes region using a coupled air-lake model

Large water bodies such as the Laurentian Great Lakes have significant influences on local 1 and regional climate through their unique physical features. Due to the coarse spatial resolution of 2 general circulation models (GCMs), the Great Lakes are geometrically ignored in most GCMs. Thus, 3 the dynamical downscaling technique serves as a necessary and feasible solution to bridge the gap. 4 The Weather Research and Forecasting model (WRF) with an updated lake scheme is employed to 5 downscale from a GCM, GFDL-CM3. The WRF-Lake’s performance is evaluated against observations, 6 the GCM, as well as 23 other GCMs. Results show that the coupled air-lake model, with a fine spatial 7 resolution and realistic lake bathymetries, reproduced a more reasonable spatiotemporal climatology 8 than GCMs, as well as the lake-induced characteristics that were missed in GCMs. With lakes 9 present, the seasonal variability of air temperature was reduced in WRF-Lake relative to GFDL-CM3, 10 especially in summer. A reduced air temperature trend, about 4.5 ◦C/100 year in the 21st century, was 11 projected in WRF-Lake. The seasonal evolutions of lake surface temperature and lake ice coverage 12 were well captured by the lake model. The lake surface temperature was projected to be warming 13 by 3.5-4 ◦C and the lake ice diminishing by 58.9% 86%. Those results brought by the WRF-Lake 14 model suggest that a fine resolution of the topography and the incorporation of the lake-atmosphere 15 interactions are crucial to improve the understanding of the climate and climate change in the Great 16 Lakes region. 17

. Summary of prior studies with RCM-lake modeling systems in the Great Lakes Offline: the lake model was post-run; One-way: the lake model was forced by the RCM; Fully: the lake scheme was embedded in the RCM. In water-dominated regions such as the Great Lakes basin, the treatment of lake surface 62 temperature (LST) is a key point in the downscaling process (Table 1). If no lake component was 63 implemented in the Great Lakes, a "search" option in RCMs was employed to extrapolate LST from 64 the closest water point with valid data, e.g. Hudson Bay and the Atlantic Ocean [19,24,25,33]. Another In this study, we applied a regional model, the Weather Research and Forecasting (WRF) model  has a spatial resolution sufficient to satisfy the nesting ratio (about 6.5:1) between the parent GCM WRF used grid and subgrid moisture processes parameterized by WSM5 (WRF Single-Moment 114 5-class scheme) microphysics scheme [37] and KF (Kain-Fritsch) cumulus convection scheme [38].

115
Shortwave and longwave radiative processes were represented by the RRTMG schemes (Rapid

133
In the default lake model, the albedos of water and ice are specified with constant values, 134 0.08 and 0.6, respectively, ignoring solar diffusion, radiation spectrum, and snow effect. Before 135 the implementation of the coupled WRF-Lake model, some modifications have been added to the lake 136 model, including a dynamic lake surface albedo, calibrated vertical diffusivities, and a sophisticated 137 treatment of snow cover over lake ice [45]. At the current grid spacing, the Great Lakes are well 138 represented in our WRF-Lake model (Fig. 3).    air temperature (T2) was compared with GFDL-CM3, CMIP5 MME, and two reanalysis datasets.

168
The annual, winter (December-January-February), and summer (June-July-August) climatologies from lakes in winter, the air temperature over Lake Superior and Lake Michigan is warmer than the 174 surrounding land area at the same latitude, and vice versa in summer. Lake Erie acts as a warming 175 pool to the overlying atmosphere in the summer. Some discrepancies also exist in the WRF-Lake model.

176
There is a systematic cold bias in the RCM, which was partially induced by the parent GCM. The lake 177 model magnified the cooling effect from lakes, especially in summer time.  In addition to air temperature, precipitation from the RCM and GCMs is also assessed against 179 observations (Fig. 5). In winter, the general pattern (southeast-northwest gradient) of precipitation

191
[56]). Our downscaling result is quite consistent with the observations that less precipitation was 192 produced over lake than over land in summer.

193
It has been demonstrated that the MME is able to reduce model uncertainties and agrees better 194 with observations than a single model [57]. However, if there is a systematic discrepancy, such as no 195 lakes being represented, in CMIP5 models, the MME result doesn't change the overall shortcoming.

196
This discrepancy even exists in the reanalysis data when it is too coarse to resolve individual lakes.

197
On this point, the downscaling technique becomes particularly important to mitigate this problem.

198
The above assessment of T2 and precipitation indicates that the historical simulation from GFDL-CM3 199 has been significantly improved by the WRF downscaling which is coupled with a lake model. In the 200 following subsections, we will analyze the future projection in our WRF-Lake model and compare it 201 with GFDL-CM3 and CMIP5 MME. to have a larger increase in the RCM than that in the GCM, especially in Lake Superior where the 211 overlaying atmosphere could be warmed as much as 5 • C after the lake was introduced in WRF.

212
The future change of the summertime T2 was characterized with a domain-wide strong warming in 213 GFDL-CM3 and WRF, but the magnitude was reduced in WRF, agreeing with the MME's projection.

214
In the downscaling procedure, the WRF didn't alter the overall pattern of T2 change projected by GFDL-CM3, but highlighted the Great Lakes' influence on the atmosphere at local and regional scales, 216 which was missed in CMIP5 GCMs. According to the Clausius-Clapeyron expression, the saturation vapor pressure is regulated 218 by the air temperature. The column-integrated water vapor increases by roughly 7.5% K -1 , and 219 precipitation by 2.2% K -1 [58]. The local precipitation change, which can be attributed to multiple 220 factors: water vapor, instability, and topography, exhibits a strong spatial heterogeneity (Fig. 7), 221 relative to the air temperature change (Fig. 6). In the annual mean precipitation, the region southeast

235
In the previous two subsections, we compared the climatology of historical simulation and future 236 projection. In this subsection, their interannual variabilities and trends will be examined. Fig. 8 shows 237 the area-mean T2 in HIS and RCP simulated by WRF, GFDL-CM3 and MME. In both HIS and RCP, the 238 RCM has maintained the intraseasonal and interannual variabilities that were produced by the GCM.

239
The difference between RCP and HIS (Figs. 8c and 8f) suggests that even in a much warmed climate 240 in the late 21st century we will still experience some cold winters. In CMIP5 MME with interannual of T2 in WRF, GFDL-CM3 and MME were calculated (Table 4). In the RCP4.5 scenario, the T2 will 253 have a rapid increase in the first half-century and the rate of increase will slow down gradually in the     Generally, the projection for the GCM we selected ranks in the middle among the CMIP5 models, 265 suggesting that our downscaling results are representative. The GCM projected a stronger warming 266 of T2 in summer than in winter over all lakes, especially Lake Superior. The lakes' impact on T2 is 267 projected to be enhanced in the future. For example, T2 was projected to warm by 6.7 • C on Lake February, because too much ice was produced in the lake model (to be elaborated in section 3f), an  downscaling result had a systematically cooler LST than GLSEA, except for Lake Superior, which is 286 the coldest lake. However, the seasonal variability of LST has been well reproduced by the 1-D lake 287 model. In the RCP4.5 scenario, LST was projected to increase by roughly 3.5 • C in Lake Superior and 288 up to 4.0 • C in Lake Ontario. Similar to T2 (Fig. 10), LST will rise more in summer than winter. The 289 magnitude of the LST change is smaller than that of the T2 change. ice coverage was projected to decrease by 58.9% (Lake Superior) to 86% (Lake Ontario). Given that 300 the ice coverage in the historical simulation was substantially overestimated, the future change of ice 301 coverage is likely to be amplified by the lake model. In addition to the lake-mean ice coverage, the 302 seasonal change of the spatial ice coverage was further examined (Fig. 13). Heavy ice was produced in 303 north coastal Lake Superior. The lower lakes will experience more dramatic ice losses under global 304 warming. In April in the last decades of 21 century, the ice will almost disappear in all lakes.  lakes is capable of reproducing a more reasonable spatial-temporal climatology in the Great Lake 320 region, as well as the lake-induced characteristics that were missed in the GCMs. the WRF-Lake model, especially in summer. Even in an overall warming climate, we could still 323 experience some colder-than-20th-century winters in the late 21st century.