Global dust cycle and uncertainty in CMIP5 models

Abstract. Dust cycle is an important component of the Earth system and have been implemented into climate models and Earth System Models (ESMs). An assessment of the dust cycle in these models is vital to address the strengths and weaknesses of these models in simulating dust aerosol and its interactions with the Earth system and enhance the future model developments. This study presents a comprehensive evaluation of global dust cycle in 15 models participating in the fifth phase of the Coupled Model Intercomparison Project (CMIP5). The various models are compared with each other and with an aerosol reanalysis as well as station observations of dust deposition and concentrations. The results show that the global dust emission in these models ranges from 735 to 8186 Tg yr−1 and the annual mean dust burden ranges from 2.5 to 41.9 Tg, both of which scatter by a factor of about 10–20. The models generally agree with each other and observations in reproducing the dust belt that extends from North Africa, Middle East, Central and South Asia, to East Asia, although they differ largely in the spatial extent of this dust belt. The models also differ in other dust source regions such as North America and Australia, where the contributions of these sources to global dust emissions vary by a factor of more than 500. We suggest that the coupling of dust emission with dynamic vegetation can enlarge the range of simulated dust emission. For the removal process, all the models estimate that wet deposition is a smaller sink than dry deposition and wet deposition accounts for 12–39 % of total deposition. The models also estimate that most (77–91 %) of dust particles are deposited onto continents and 9–23 % of them are deposited into oceans. A linear relationship between dust burden, lifetime, and fraction of wet deposition to total deposition from these models suggests a general consistency among the models. Compared to the observations, most models reproduce the dust deposition and dust concentrations within a factor of 10 at most stations, but larger biases by more than a factor of 10 are also noted at specific regions and for certain models. These results cast a doubt on the interpretation of the simulations of dust-affected fields in climate models and highlight the need for further improvements of dust cycle especially on dust emission in climate models.


spatial extent of this dust belt. The models also differ in other dust source regions 23 such as North America and Australia, where the contributions of these sources to 24 global dust emissions vary by a factor of more than 500. We suggest that the 25 coupling of dust emission with dynamic vegetation can enlarge the range of 26 simulated dust emission.

27
For the removal process, all the models estimate that wet deposition is a 28 smaller sink than dry deposition and wet deposition accounts for 12-39 % of total 29 deposition. The models also estimate that most (77-91 %) of dust particles are 30 deposited onto continents and 9-23 % of them are deposited into oceans. A linear 31 relationship between dust burden, lifetime, and fraction of wet deposition to total 32 deposition from these models suggests a general consistency among the models. 33 Compared to the observations, most models reproduce the dust deposition and dust 34 concentrations within a factor of 10 at most stations, but larger biases by more 35 than a factor of 10 are also noted at specific regions and for certain models. These

42
Dust cycle is an important component of the Earth system as it has strong impacts 43 on the Earth environment and climate system (Shao et al., 2011). Dust aerosol in the 44 atmosphere significantly impacts the climate systems via various pathways, such as 45 scattering and absorbing the solar and terrestrial radiation, modifying cloud radiative 46 forcing by acting as cloud condensation nuclei and ice nucleating particles, and reducing 47 the snow albedo when depositing onto snow (Boucher et al., 2013;Forster et al., 2007;48 Liu, et al., 2012a;Mahowald et al., 2011;Wu et al., 2018a;Rahimi et al., 2019). Dust 49 affects the biogeochemical cycle by delivering the nutrients (e.g., mineral, nitrogen, and 50 phosphorus) from dust sources to the oceans/other continents (Jickells et al., 2005;51 Mahowald et al., 2011). Dust aerosol is also one of the main contributors to air pollution 52 that is hazardous to human health (Bell et al., 2008;Lin et al., 2012). 53 To quantify the dust impacts on Earth system, dust cycle including dust emission, 54 transport, and dry and wet deposition has been incorporated in climate models and Earth 55 System Models (ESMs) since 1990s. These models have the capability to reproduce the 56 general patterns of global dust distribution (e.g., Ginoux et al., 2001;Zender et al., 2003;57 Yue et al., 2009;Huneeus et al., 2011;Liu et al., 2012b). However, large uncertainties 58 still exist in the simulated global dust budgets in these models, as revealed by a wide 59 range of model results. A comparison of 14 different models from the Aerosol 60 that it includes the assimilation of AOD (Randles et al., 2017), which is not included in 157 MERRA and other commonly-used reanalysis datasets such as ECWMF Reanalysis 158 (ERA5) and NCEP/DOE Reanalysis II (R2). The aerosol fields (including dust) in 159 MERRA-2 are significantly improved compared to an identical control simulation that 160 does not include the AOD assimilation (Randles et al., 2017;Buchard et al., 2017). It 161 should be noted that as only AOD is taken into account in the aerosol assimilation, there 162 may be discrepancies in the related aerosol fields such as aerosol concentration and 163 deposition. In addition, dust emission is calculated directly from surface wind speed and 164 soil wetness based on the dust emission scheme of Ginoux et al. (2001), and there is no 165 direct impact on emission from aerosol assimilation. Therefore, there may be 166 inconsistence between dust emission, burden, and deposition. In fact, as shown in the 167 Section 4, there is imbalance between total dust emission and deposition globally and 168 adjustment of dust emission to fit the dust burden is still needed. Despite the limitation, 169 MERRA-2 provides a well-constrained global dust dataset, which is very useful for 170 model evaluations. We will use MERRA-2 as a referential data but with the knowledge 171 of its limitation. We will use the long-term means of dust-related variables during the 172 whole period when data is available (i.e., 1980-2018). Dust in MERRA-2 is treated by 173 five size bins spanning from 0.2 to 20 μm, which are summed to provide the total values. 174 MERRA-2 is provided at the resolution of 0.5º ×0.625º , which is similar to one CMIP5 175 model (MIROC4h) and finer than other CMIP5 models. 176 177 3. Observations such as circulation and precipitation (e.g., Wu and Lin, 2013). The estimated global dust 248 burden ranges from 2.5 to 41.9 Tg, and from 8.1 to 36.1 Tg when MIROC4h and 249 HadGEM2-CC/ES are excluded. The lifetime of global dust particles ranges from 1.3 to 250 4.4 days. The dust burden (lifetime) in MERRA-2 is 20.3 Tg (4.1 days), which is larger 251 (longer) than most CMIP5 models. The fraction of wet deposition to total deposition in 252 MERRA-2 is 38.6%, which is in the upper end of CMIP5 results. There is a linear 253 relationship (with the correlation coefficient R=0.67, above the statistically significant 254 level of 0.01) between global dust burden and lifetime in CMIP5 models (excluding 255 HadGEM2-CC/ES; Figure 2a), indicating a longer lifetime of dust is generally associated 256 with a larger dust burden. Linear relationship (R=0.46, above the statistically significant 257 level of 0.05) is also found between lifetime and fraction of wet deposition (Figure 2b), 258 which indicates that a longer lifetime corresponds to a larger fraction of wet deposition in 259 the total deposition. 260 261

Global dust emissions 262
Dust emission is the first and the foremost process in the dust cycle and determines 263 the amount of dust entrained into the atmosphere. Figure 3 shows the spatial distribution 264 of dust emission fluxes from 15 CMIP5 models and MERRA-2 reanalysis. In general, all 265 the models can reproduce the main dust sources, known as the "dust belt" that extends 266 from North Africa, Middle East, Central Asia, South Asia, to East Asia and that can be 267 seen from satellite observations (Prospero et al., 2002;Ginoux et al., 2012). However, the 268 models differ significantly in the extent of this "dust belt". Although a large group of 269 CMIP5 models (CSIRO-Mk3-6-0, GFDL-CM3, GISS-E2-H/S, MIROC5, MIROC-ESM, 270 emission regions mostly over deserts and adjacent arid/semi-arid regions, two of the 272 models (CESM1-CAM5 and MIROC4h) simulate much smaller areas of dust emission 273 and a few others (ACCESS1-0, CanESM2, HadGEM2-CC/ES) simulate more extended 274 dust emission regions. CESM1-CAM5 simulates isolated dust emission regions with "hot 275 spots" of dust emissions larger than 500 g m -2 yr -1 , and dust emission in MIROC4h 276 concentrates only over the centers of deserts. In contrast, ACCESS1-0, CanESM2, and 277 HadGEM2-CC/ES not only simulate the dust emissions in deserts and adjacent regions, 278 but also produce a considerable amount of dust emissions over the Eastern Africa 279 (Somalia, Ethiopia, and Kenya), East India, and northern part of Indo China Peninsula, 280 which are rarely regarded as potential dust sources (Formenti et al., 2011;Shao, 2008). 281 Dust sources also exist in Australia, North America, South America, and South 282 Africa, as evident from surface observations (e.g., Shao, 2008) and satellite observations 283 (Prospero et al., 2002;Ginoux et al., 2012), although the emission fluxes are smaller than 284 those in the aforementioned "dust belt". In these regions, most models produce a 285 considerable amount of dust emissions (>5 g m -2 yr -1 ), while a small group of models 286 simulate much less or even negligible dust emissions. The models differ greatly in these 287 regarded as a potential dust source (Formenti et al., 2011;Shao, 2008). 299 The contributions of dust emissions in nine different regions to global dust emission 300 is summarized in Table 4. The models consistently simulate the largest dust emission in 301 North Africa, which accounts for 36-79% of the global total dust emission. The models 302 also estimate large dust emissions in Middle East and East Asia, which account for 7-20% 303 and 4-19% of global dust emission, respectively. The contributions from Central Asia and 304 South Asia in CMIP5 models range from 1-14% and 0.9-10%, respectively. The 305 contributions from other sources (North America, South Africa, Australia, South America) 306 are much less consistent among the models, and the largest difference is in North 307 America (0.008-4.5%) and Australia (0.02-28%) by three orders of magnitude. 308 Particularly, HadGEM2-CC/ES simulate 25-28% of global dust emission from 309 Australia, which is comparable to that from sum of all Asian sources (Middle East, 310 Central Asia, South Asia, and East Asia). This estimate is unrealistically high, as will be 311 Among the CMIP5 models, CESM-CAM5 and MIROC4h simulate the smallest dust 331 emission area, which are 2-3% of the global surface area, while CanESM2 simulates the 332 largest dust emission area (18% of the global surface area; Figure 4 and Table 3). The 333 maximum normalized dust emission flux is also the largest at 2682 and 3635 in CESM1-334 CAM5 and MIROC4h, respectively, indicating the "hot spots" with extremely high dust 335 emission flux in the two models. The maximum normalized dust emission flux is 336 generally between 100 and 300 in other CMIP5 models and is approximately 200 in 337 MERRA-2 reanalysis. 338 The smallest dust emission area in CESM1-CAM5 is mainly because the model 339 adopts a geomorphic source erodibility with a threshold value of 0.1 for the dust emission 340 occurrences (Zender et al., 2003;Wu et al., 2016). Small dust emission area in MIROC4h 341 may be partly due to the higher horizontal resolution of the model (0.56º ) than other 342 models (1º -3º ) including MIROC5 (Table 1). The higher model resolution may change 343 the patterns of wind speeds and precipitation as well as the occurrence frequency of 344 strong winds and heavy precipitation and thus affect the dust emission regions. The 345 largest dust emission area in CanESM2 may be due to its prescribed land cover map,

Dust deposition flux 382
Dust deposition is a vital process in the dust cycle which removes dust particles 383 from the atmosphere and provides nutrients to the terrestrial and marine ecosystems. 384 Figure 6 shows the comparison of dust deposition flux at 84 selected stations between the 385 models and observations. Only seven CMIP5 models provide total dust deposition flux 386 (sum of dry and wet deposition), which are used here. The global dust emission in these 387 seven models ranges from 1600 to 3500 Tg yr -1 , which is at the medium level of all the 388 CMIP5 models. Observed annual mean dust deposition flux ranges from 10 -4 to 10 3 g m -2 389 yr -1 , indicating large spatial variabilities of dust deposition. In general, six of seven 390 CMIP5 models (excluding ACCESS1-0) reproduces the observed dust deposition flux 391 within a factor of 10 in most regions except over the Southern Ocean, Antarctica, and 392 Pacific. Over the Southern Ocean and in the Antarctica, all the models except CESM1-393 CAM5 overestimate the dust deposition flux by more than a factor of 10 at two stations. Dust cycle can deliver nutrients from continents to oceans. Table 5  in continents (oceans), and this estimation is smaller (larger) than all seven CMIP5 448 models, indicating MERRA-2 transport dust more efficiently to oceans. This is consistent 449 with the comparison of dust deposition flux shown in Figure 6 and may be related to the 450 assimilation of both meteorology and aerosols in MERRA-2. The fractions of wet 451 deposition (with respect to total deposition) in seven CMIP5 models are 8-33% and 49-71% 452 over continents and oceans, respectively. MERRA-2 estimates the fraction of wet deposition (with respect to total deposition) 26% and 69% over the continents and oceans, 454 respectively, which lie within the range of CMIP5 models. 455 456

Dust concentration 457
Dust concentration is an important variable for its cycle. Figure  observations is 0.91, which is larger than all the CMIP5 models, and mean bias (MBlog) is 507 close to zero (0.01). 508 509

510
In this study we examine the present-day global dust cycle simulated by the 15 511 climate models participating in the CMIP5 project. The simulations are also compared 512 with a dataset MERRA-2 and observations of dust deposition and concentration. The 513 results show that the global dust emission in these models ranges from 735 to 8186 Tg yr -514 1 and the global dust burden ranges from 2.5 to 41.9 Tg. The differences are larger than 515 those from models participating in the AeroCom project (Huneeus et al., 2011), which is 516 a result of enhanced model complexities in modeling both climate and dust emission in 517 the CMIP5 models. 518 The simulated dust emission regions also differ greatly accounting for a global 519 surface area of 2.9%-18%. The models agree most with each other in reproducing the Asia, but there are large uncertainties in the extent of this "dust belt" and other source 522 regions including Australia, North America, South America, and South Africa. and fraction of wet deposition to total deposition is present in the CMIP5 models, 534 suggesting a general consistency among these models. The models also estimate that 77-535 91% of emitted dust are deposited back to continents and 9-23% of them are deposited to 536 the oceans. The fraction of wet deposition is smaller in most CMIP5 models and dust 537 lifetime is shorter compared to MERRA-2 reanalysis, indicating a shorter distance for 538 dust transport from its sources in most CMIP5 models. Compared to the observations, the 539 can be ascribed to the fact that most dust emissions in MIROC4h are concentrated over 551 the desert centers, which limits the long-range transport of dust particles to the remote 552

regions. 553
These results show large uncertainties of global dust cycle in ESMs. In fact, these 554 models are fully-coupled atmosphere-land-ocean models and some of them also include 555 the dynamic vegetation. As a result, uncertainties are larger compared to those in 556 previous models participating in the AeroCom intercomparison project where sea surface 557 temperature is prescribed, and more strictly, in some models, meteorological fields are 558 prescribed from reanalysis (Huneeus et al., 2011). Larger uncertainties in the CMIP5 559 models with dynamic vegetation is expected, as a prognostic vegetation would depart 560 from the observed or constructed vegetation and may also lead to a large bias in soil 561 moisture, which may thus lead to an additional bias in dust emissions in these models. 562 Uncertainties of dust simulations also vary with regions, and a smaller uncertainty is 563 found in the deserts over the "dust belt" in the North Hemisphere, but a larger uncertainty 564 exists in other regions including Australia and North America. The large uncertainties of 565 global dust cycle in the CMIP5 models would cast a doubt on the reliability of dust 566 radiative forcing estimated in these models. 567 Because the dust lifecycle involves various processes with the scales from 568 micrometers to tens of thousands of kilometers and consists of lots of parameters, the 569 representation of dust cycle in climate models is a big challenge for the model 570 community. Dust emission is the first and foremost process for model improvements of 571 dust cycle (Shao, 2008;Shao et al., 2011). Improving dust emission not only lies in the 572 development of dust emission scheme but also in its implementation into climate models 573 (e.g., Shao, 2008;Wu et al., 2016;Wu et al., 2019). For example, different dust emission 574 schemes with specific land cover datasets and criteria for the occurrence of dust emission 575 are adopted in the models (Table 1 and references therein). Therefore, different results of 576 dust emission among the CMIP5 models reflect in many aspects the differences in 577 meteorology, land cover data, and dust emission parameterizations. A close look at these 578 factors in each model will help to unravel reasons behind the biases in these models. In It should be mentioned that dust size distribution is an important parameter for dust 587 cycle (e.g., Shao, 2008;Mahowald et al., 2014), and it is not included in this study as the 588 model data are not available. Evolution of dust size distribution during dust transport and 589 deposition is critical to our understanding of the model bias in dust cycle. We suggest that 590 the size-resolved dust emission, concentration, and deposition should be outputted and 591 provided in the latest CMIP6 project (Eyring et al., 2016). Moreover, observations of 592 size-resolved dust concentration and deposition is urgently needed. A compile of 593 available observations of dust size distribution (e.g., Mahowald et al., 2014: Ryder et al., 594 2018 are also required for model evaluation.    emission. The dust emission area is defined as the region with the annual mean dust 969 emission flux larger than 1% of global mean annual dust emission flux. 970 b : The ratio of wet deposition to total deposition is given in parenthesis next to wet 971 deposition. 972 b : The global dust deposition is 1692 Tg, which is larger than dust emission because 973 of no adjustment done with dust emission after aerosol assimilation (Section 2). 974 https://doi.org/10.5194/acp-2020-179 Preprint. Discussion started: 3 April 2020 c Author(s) 2020. CC BY 4.0 License. Table 5. Total dust deposition and wet deposition in the global surface, continents, 977 and oceans, respectively from CMIP5 models and MERRA-2 reanalysis. Only the 978 seven CMIP5 models with both dry and wet depositions provided are used here.