RESEARCH ARTICLE SUMMARY◥

CLIMATE

Aerosol-driven dropletconcentrations dominate coverageand water of oceanic low-level cloudsDaniel Rosenfeld*†, Yannian Zhu†, Minghuai Wang*, Youtong Zheng,Tom Goren, Shaocai Yu*

INTRODUCTION: Human-made emissionsof particulate air pollution can offset part ofthe warming induced by emissions of green-house gases, by enhancing low-level cloudsthat reflectmore solar radiation back to space.The aerosol particles have this effect becausecloud droplets must condense onpreexisting tiny particles in thesame way as dew forms on coldobjects; more aerosol particlesfrom human-made emissions leadto larger numbers of smaller clouddroplets.Onemajorpathway for low-level cloud enhancement is throughthe suppression of rain by reducingclouddroplet sizes. This leavesmorewater in the cloud for a longer time,thus increasing the cloud cover andwater content and thereby reflect-ing more solar heat to space. Thiseffect is strongest over the oceans,where moisture for sustaining low-level clouds over vast areas is abun-dant. Predicting global warmingrequires a quantitative understand-ing of how cloud cover and watercontent are affected by human-made aerosols.

RATIONALE: Quantifying theaerosol cloud–mediated radiativeeffects has been a major challengeand has driven the uncertainty inclimate predictions. It has beendifficult to measure cloud-activeaerosols from satellites and to iso-late their effects on clouds frommeteorological data. The develop-ment of novel methodologies toretrieve cloud droplet concentra-tions and vertical winds from satellites rep-resents a breakthrough that made this quan-tification possible. The methodologies wereapplied to the world’s oceans between theequator and 40°S. Aerosol and meteorologicalvariables explained 95% of the variability inthe cloud radiative effects.

RESULTS: The measured aerosol cloud–mediated cooling effect was much largerthan the present estimates, especially via theeffect of aerosols on the suppression of pre-cipitation, which makes the clouds retain morewater, persist longer, and have a larger frac-

tional coverage. This goes against most previousobservations and simulations, which reportedthat vertically integrated cloud water may evendecrease with additional aerosols, especially inprecipitating clouds. The major reason for thisapparent discrepancy is because deeper cloudshave more water and produce rainfall more

easily, thus scavenging the aerosols more effi-ciently. The outcome is that clouds with feweraerosols have more water, but it has nothing todo with aerosol effects on clouds. This fallacy isovercome when assessing the effects for cloudswith a given fixed geometrical thickness.The large aerosol sensitivity of thewater con-

tent and coverage of shallow marine cloudsdispels another belief that the effects of addedaerosols are mostly buffered by adjustmentof the cloud properties, which counteracts the

initial aerosol effect. Forexample, adding aerosolssuppresses rain, so theclouds respond by deep-ening just enough to re-store the rain amount thatwas suppressed. But the

time scale required for the completion of thisadjustment process is substantially longerthan the life cycle of the cloud systems, whichismostly under 12 hours. Therefore,most of themarine shallow clouds are not buffered for the

aerosol effects, which are inducingcooling to a much greater extentthan previously believed.

CONCLUSION: Aerosols explainthree-fourths of the variability in thecooling effects of low-level marineclouds for a given geometrical thick-ness. Doubling the cloud dropletconcentration nearly doubles thecooling. This reveals a much greatersensitivity to aerosols than previ-ously reported, meaning too muchcooling if incorporated into pres-ent climate models. This argumenthas been used to dismiss such largesensitivities. To avoid that, the aero-sol effects in some of the modelswere tuned down. Accepting thelarge sensitivity revealed in thisstudy implies that aerosols haveanother large positive forcing, pos-sibly through the deep clouds,which is not accounted for in cur-rent models. This reveals additionaluncertainty that must be accountedfor and requires a major revision incalculating Earth’s energy budgetand climate predictions. Paradoxical-ly, this advancement in our knowl-edge increases the uncertainty inaerosol cloud–mediated radiativeforcing. But it paves the way toeventual substantial reduction of

this uncertainty.▪

RESEARCH

Rosenfeld et al., Science 363, 599 (2019) 8 February 2019 1 of 1

Coverage and droplet concentrations (Nd) of shallow marineclouds over the northeast Pacific. Smoke particles emitted fromship smokestacks form cloud droplets and elevate Nd.The smoke-freeclouds (Nd < ~30 cm−3) precipitate and break up. The fraction ofcloud cover increases with more Nd that suppresses precipitation.Thesolid cloud cover is maintained by smoke that was spread from oldship tracks, crossed by newer ones.

The list of author affiliations is available in the full article online.*Corresponding author. Email: daniel.rosenfeld@huji.ac.il (D.R.);minghuai.wang@nju.edu.cn (M.W.); shaocaiyu@zju.edu.cn (S.Y.)†These authors contributed equally to this work.Cite this article as D. Rosenfeld et al., Science 363, eaav0566(2019). DOI: 10.1126/science.aav0566

ON OUR WEBSITE◥

Read the full articleat http://dx.doi.org/10.1126/science.aav0566..................................................

Erratum 20 June 2019. See erratum. on M

arch 10, 2021

http://science.sciencemag.org/

Dow

nloaded from

RESEARCH ARTICLE◥

CLIMATE

Aerosol-driven dropletconcentrations dominate coverageand water of oceanic low-level cloudsDaniel Rosenfeld1,2*†, Yannian Zhu3†, Minghuai Wang2,4*, Youtong Zheng5,Tom Goren6, Shaocai Yu7,8,9*

A lack of reliable estimates of cloud condensation nuclei (CCN) aerosols over oceans hasseverely limited our ability to quantify their effects on cloud properties and extent ofcooling by reflecting solar radiation—a key uncertainty in anthropogenic climate forcing.We introduce a methodology for ascribing cloud properties to CCN and isolating the aerosoleffects from meteorological effects. Its application showed that for a given meteorology,CCN explains three-fourths of the variability in the radiative cooling effect of clouds, mainlythrough affecting shallow cloud cover and water path. This reveals a much greatersensitivity of cloud radiative forcing to CCN than previously reported, which means too muchcooling if incorporated into present climate models. This suggests the existence ofcompensating aerosol warming effects yet to be discovered, possibly through deep clouds.

Marine stratocumulus clouds (MSCs) areresponsible for reflecting much of thesolar radiation received by Earth backto space. Therefore, understanding thecauses for the variability in their radi-

ative effects is necessary for understanding thenatural and anthropogenic controls of Earth’senergy budget and the resultant climatic impli-cations. The effects of MSCs on the albedo of agiven ocean area are determined by cloud frac-tion (Cf), droplet concentration (Nd), and liquidwater path (LWP). Cloud droplets must form oncloud condensation nuclei (CCN) aerosols, andNd is determined by the CCN activation spec-trum as a function of vapor supersaturation (S),which in turn is driven by cloud base updraftvelocity (Wb) (1). The overall properties of thecloud system are determined by the meteoro-logical setting, which includes vertical thermo-

dynamic and wind profiles. However, radiationemitted by the clouds and precipitation fromthem may modify their state. Because such feed-backs can depend on the aerosol effects on cloudmicrostructure and precipitation, it is very dif-ficult to isolate the effects of aerosols on thecloud properties (2). Here, we addressed this chal-lenge by calculating, for a given meteorologicalcondition, the amount of variability in Cf, LWP,and cloud radiative effect (CRE) that can be ex-plained by variability in CCN. The meteorologyis encapsulated by the satellite-retrievedWb, thecloud geometrical thickness (CGT), and the ther-modynamic structure of the lower troposphere.

Known effects of aerosols on shallowwater clouds

This study is limited to cumulus and MSC cloudswith geometrical thickness of up to 800m.Mostcloud droplets in shallow convective clouds format their base. In adiabatic cloud parcels, the clouddropmass grows nearly linearly with height abovecloud base while the Nd mixing ratio remainsconstant. This leads to a nearly linear increasein cloud liquidwater content (LWC) with heightabove base. Therefore, integration of LWC withheight results in LWP º CGT2. Aircraft obser-vations have shown that the cores of actual shal-low convective clouds and MSCs do not deviatesubstantially from this ideal behavior (3). Thecloud drop coalescence rate is proportional toLWC2re

5, where re is the cloud drop effective ra-dius (4). When re exceeds 14 mm, the clouds startprecipitating through the quick formation ofdrizzle (5–10). The drizzle leads to accretionthat forms raindrops from the full column ofthe clouds (10). The CGT for reaching that valueof re depends linearly on Nd (4, 11, 12).

Overcast decks of MSCs usually break up whenthey begin to produce substantial amounts ofprecipitation (rain rate > ~2 mm day−1) inresponse to the thermodynamic effects. Theclouds break up mainly because rain-drivendowndrafts form miniature gust fronts at thesea surface, which trigger convective cloud for-mation when they collide with each other (13–15).This leads to a large decrease in Cf from nearunity to ~0.6 (16). The precipitation scavengingof the CCN leads to a decrease of Nd at theconvective cores from an average of 55 cm−3 to15 cm−3 from just before to after the transitionfrom closed to open cells of MSCs, whereasLWC does not change significantly during thetransition (16, 17).Aerosol optical depth (AOD) and aerosol index

(AI) have been used as proxies for CCN (18–20).However, direct comparisons of satellite-observedrelationships of AOD and cloud properties dis-agree with model simulations, as the modelshave a much larger sensitivity to aerosols thanobservations indicate (21). This is caused byissues that decorrelate the aerosol propertiesretrieved by satellite from boundary layer CCNconcentrations, such as aerosol swelling due tohigher relative humidity in the vicinity of clouds[see review in (22)]. Furthermore, aerosols can-not be retrieved in cloudy conditions, where theyare most needed. Even under simulated idealconditions, AOD can explain only a small frac-tion of the variability of CCN in the boundarylayer (23). This might explain the findings ofGryspeerdt et al. (24), who showed that ~80%of the indicated positive relationship betweenAOD and Cf is explained by factors that cannotreadily be interpreted as a causal effect of AODvariability on Cf. Much of this problem is causedby the poor relationship between AOD and CCN,especially in areas with small CCN concentrations,where a small absolute change in the concentra-tion is a large fractional change. Because theeffects of aerosols on cloud properties are loga-rithmic (25), the largest aerosol effects on cloudsare poorly detected. The situation is worst overthe Southern Oceans, where the clouds preventretrieval of AOD most of the time. Furthermore,when AOD is retrieved, its value is often withinthe detection limit of 0.06 over ocean [figure 11of (26), which translates to >100 to 200 CCN cm−3

at S = 0.4% (27)]. Moreover, according to figure 14of (26), the median satellite-indicated AOD overocean is only 0.05, which is within the range ofmeasurement error.This situation requires an approach that does

not depend onAOD. In response, we useNd andWb as proxies for CCN at cloud base (28). Thishas become possible as a result of recentlydeveloped methodologies for retrieving Nd (29)andWb (30, 31) of the convective cores of clouds,which best reflect CCN concentrations ingestedinto cloud bases.

Known effects of meteorology onshallow water clouds

Marine stratocumulus clouds form by radiativecooling of the cloud-topped marine boundary

RESEARCH

Rosenfeld et al., Science 363, eaav0566 (2019) 8 February 2019 1 of 9

1Institute of Earth Sciences, The Hebrew University ofJerusalem, Jerusalem 91904, Israel. 2School of AtmosphericSciences, Nanjing University, China. 3Meteorological Instituteof Shaanxi Province, Xi’an, China. 4Joint InternationalResearch Laboratory of Atmospheric and Earth SystemSciences and Institute for Climate and Global ChangeResearch, Nanjing University, China. 5Earth System ScienceInterdisciplinary Center, University of Maryland, College Park,MD, USA. 6University of Leipzig, Leipzig, Germany. 7ResearchCenter for Air Pollution and Health; Key Laboratory ofEnvironmental Remediation and Ecological Health, Ministry ofEducation, College of Environmental and Resource Sciences,Zhejiang University, Hangzhou, Zhejiang, P.R. China. 8Divisionof Chemistry and Chemical Engineering, California Instituteof Technology, Pasadena, CA 91123, USA. 9Center forExcellence in Regional Atmospheric Environment, Institute ofUrban Environment, Chinese Academy of Sciences, Xiamen,Fujian 361021, P.R. China.*Corresponding author. Email: daniel.rosenfeld@huji.ac.il (D.R.);minghuai.wang@nju.edu.cn (M.W.); shaocaiyu@zju.edu.cn (S.Y.)†These authors contributed equally to this work.

Erratum 20 June 2019. See erratum. on M

arch 10, 2021

http://science.sciencemag.org/

Dow

nloaded from

layer. This is promoted by the moist marineboundary layer that is capped by a pronouncedinversion with dry air above it, which allows astrong radiative cooling of the cloud tops. Thedryness of the free troposphere air is often aresult of subsidence in anticyclones. Klein et al.(32) showed that lower tropospheric stability

(LTS, the difference between potential temper-ature at 700 hPa and at the surface) explainsmost of the variability in the seasonally averagedamount of MSCs at five oceanic regions, with a6% increase of Cf for each 1 K increase of LTSwithin the LTS range of 14 to 22 K. They alsoshowed that the cloud top radiative cooling rate

(CTRC) explains similarly most of the variabilityin Cf with the same data. This can be the casebecause the LTS and CTRC are not independent.A recent similar analysis using an artificial neu-ron network (33) reconfirmed the major role ofLTS in the determination of Cf. Andersen et al. (33)also showed that LWP is dominated by boundary

Rosenfeld et al., Science 363, eaav0566 (2019) 8 February 2019 2 of 9

Fig. 1. The MODIS Terra overpass on 7 December 2017, 18:25 UT. (A) MODIS natural projection, divided into 110 km × 110 km scenes. The red areaswere excluded from the analysis. (B) A geographically rectified image. The contours show the LTS from the reanalysis data.

Fig. 2. Distribution of Nd and LTS for the analyzed scenes. The color represents the 1° × 1° averaged Nd (A) and LTS (B).

RESEARCH | RESEARCH ARTICLE

Erratum 20 June 2019. See erratum. on M

arch 10, 2021

http://science.sciencemag.org/

Dow

nloaded from

layer height. This makes physical sense becausea deeper boundary layer allows a proportionallylarger CGT, whereas adiabatic LWP is propor-tional to CGT2.We isolated the effects of aerosols and me-

teorology on the components of CRE—Nd, Cf,and LWP—of boundary layer marine water cloudsby testing their dependence on meteorologicalfactors (LTS, CGT, CTRC) and on aerosols, asapproximated by Nd and Wb. The value of Nd

is determined by CCN and by Wb; in turn, Wb

is determined by cloud base height (30) andCTRC (31). A surrogate for the CCN is calcu-lated by accounting for the effect of Wb on Nd

and obtaining Nd/Wb0.5, as described in the

methodology. Whereas previous analyses weredone on large areas at monthly to seasonalscales, Nd varies at much smaller scales in timeand space. The fraction of variability in CREand its components that is explained by each

of these meteorological and aerosol propertiesis quantified.

Data and methodology

MODIS data were analyzed over the SouthernOceans between the equator and 40°S. Thedata were collected for the southern summers,for themonths of November and December 2014,January and February 2015, November andDecember 2015, January and February 2016,November and December 2016, and Januaryand February 2017. A total of 4620 MODISgranules were analyzed. Each granule was di-vided into 306 scenes of 110 km × 110 km (about1° × 1° near the equator). The two external sceneson each edge were truncated for data quality(Fig. 1A). A scene was selected for analysis ifit included only liquid water–phase cloudsthat were not obscured by higher cloud layersin any part of the scene. We analyzed 664,128

(Fig. 2) out of 1,413,868 scenes (47% of totalscenes viewed by the satellites).The following cloud properties were calcu-

lated for each 110 km × 110 km scene:1) Cloud drop concentrations (Nd) in units

of cm−3. To obtain Nd that is most relevant toCCN below cloud base, we used the methodologyof Zhu et al. (29) that was developed specifi-cally for application to this study. The input isobtained from MODIS collection 6 level-2 cloudproducts of re and cloud optical depth (t). Thepixels with highest 10% of t within the scene areused for retrieving Nd (29) because the brightestclouds are the convective cores which are closestto adiabatic. Its high value (median t = 21) fur-ther reduces the retrieval bias of Nd due to up-welling radiation from below cloud tops (34).This methodology minimizes the retrieval biasof Nd in broken clouds relative to full cloudcover to less than 5% (29). The geographicaldistribution of average Nd is shown in Fig. 2A.2) Liquid water path at cloud cores (LWPcore)

in units of g m−2. It was averaged over thehighest 10% of t within the scene from pixelvalues obtained from MODIS cloud products.3) Geometrical thickness of the cloudy layer

(CGT). Both cloud top and base heights areuniquely defined for any combination of LWP,cloud top temperature (CTT), and sea surfacetemperature (SST) when assuming dry adiabaticlapse rate from SST to cloud base, and moistadiabatic from cloud base to cloud top. We solvefor CGT and cloud base height (CBH) underthese assumptions. This is done for the convec-tive cores by averaging CTT and LWPcore overthe highest 10% of t for the scene. The dailymean SST data at a spatial resolution of 0.25° ×0.25° were provided by the National Oceanic andAtmospheric Administration (NOAA). Cloudswith CGT > 800mwere excluded to restrict theanalysis for shallow clouds. This excluded only2% of the total number of scenes with single–boundary layer water clouds.4) Cloud fraction (Cf), defined as the frac-

tion of pixels of the scene that have 0.64 mmreflectance greater than the reflectance that ismatched to the minimum detectable t. Thisallows counting pixels as cloudy even when t isnot calculable in the MODIS products, as shownin fig. S12.5) Cloud top radiative cooling rate (CTRC) in

units of K day−1. CTRC was calculated by a ra-diative transfer model, SBDART (Santa BarbaraDISORT Atmospheric Radiative Transfer), withinput as CTT and the vertical profile of temper-ature and relative humidity above cloud tops, asobtained from the National Centers for Envi-ronmental Prediction (NCEP) reanalysis data (31).A full description of the methodology is providedin the supplementary materials.6) Lower tropospheric stability (LTS) in units

of K. The LTS was calculated according to thedifference of the geopotential temperature be-tween the sea surface and 700 hPa. Both cal-culations rely on atmospheric profiles from theNCEP reanalysis data. The geographical distri-bution of average LTS is shown in Fig. 2B.

Rosenfeld et al., Science 363, eaav0566 (2019) 8 February 2019 3 of 9

Fig. 3. Dependence of cloud properties on drop concentrations. (A) Dependence of Cf onNd for intervals of LWPcore. (B to D) Dependence of Cf (B), LWP (C), and CRE (D) on Nd for fixedintervals of CGT. The data are for all the scenes over the Southern Oceans between 0° and 40°S.Cloud top drop effective radius of each line (i.e., for a given LWPcore) decreases with increasingdroplet concentration, reaching 14 mm at the dotted line. The values are the medians for each bin.(The equivalent fig. S1 shows the means of the bins.) The box plot distributions for each bin areshown in figs. S5 to S8.

RESEARCH | RESEARCH ARTICLE

Erratum 20 June 2019. See erratum. on M

arch 10, 2021

http://science.sciencemag.org/

Dow

nloaded from

Rosenfeld et al., Science 363, eaav0566 (2019) 8 February 2019 4 of 9

Fig. 4. Dependence of cloud properties on drop concentrations as a function of LTS. (A, D, G, and J) LTS > 18 K; (B, E, H, and K) 18 K ≤ LTS ≤ 14 K;(C, F, I, and L) LTS < 14 K. (The equivalent fig. S2 shows the means of the bins.)

RESEARCH | RESEARCH ARTICLE

Erratum 20 June 2019. See erratum. on M

arch 10, 2021

http://science.sciencemag.org/

Dow

nloaded from

7) The strength of the inversion at the cloudtops (Dq), calculated from the reanalysis dataas the difference in the potential temperature(q) above and below the inversion.8) Neutralized Nd forWb (Nd/Wb

0.5), which isdetermined primarily by CCN. Twomey (1) showedthat for a given CCN(S) =N0S

k,Nd depends onWb

according to

Nd‐Twomey ¼ 0:88N02=ðkþ2Þð0:07Wb

1:5Þk=ðkþ2Þ ð1Þ

whereN0 is CCN concentration at supersaturationS = 1.01, and k is the coefficient of the powerlaw. We estimated Wb for convective cloudsaccording to Zheng and Rosenfeld (30), whoapproximatedWb = 0.9 CBH, whereWb is in ms−1

and CBH is in km. Clouds were considered asconvective for LTS ≤ 16 K. For MSCs (LTS > 16 K),we used the relationships of Zheng et al. (31),who showed thatWb of MSCs is driven by CTRCand is approximated as Wb = [–0.37(CTRC) +26.1]/100 ms−1. Clouds in scenes with CBH >1 km were considered decoupled and usedWb =[–0.38(CTRC) + 8.4]/100 ms−1 (35). The Nd sen-sitivity to Wb was calculated according to Eq. 1,while assuming k = 1, as Wb

0.5. Thus, Nd-Twomey

neutralized to the effect of Wb was calculatedas Nd/Wb

0.5.

Results

Figure 3 shows the dependence of Cf on Nd forvarious intervals of LWPcore and the dependenceof Cf, LTS, CGT, and CTRC onNd for fixed inter-vals of CGT. The data were classified into four-dimensional bins of Nd, LTS, CGT, and CTRC.Binning is necessary to average out variabilityin Cf at the scale of the scene (110 km × 110 km)that is not related to meteorology, such as self-organization of cloud clusters, as is evident inFig. 1B.The median values of Cf for the individual

scenes in each bin for Nd, LWPcore, and CGT areshown in Fig. 3. The box plot distributions of thecases that compose each of the points of Fig. 3are shown in figs. S5 to S8. For a given CGT, Cfincreases with both LWPcore and Nd until reach-ing nearly full cloud cover for Nd > 100 cm−3 andLWPcore≥ 240 gm−2 (Fig. 3, A andB). This result isqualitatively similar to theNd-Cf relationshipshown in figure 5 of Gryspeerdt et al. (24). How-ever, they assumed that CCN correlates with AODand concluded that themuch poorer correlationof Nd with AOD meant that only 20% of theobservedNd-Cf relationships could be interpretedas related to CCN, whereas the remaining 80%was caused by retrieval errors and meteorolog-ical factors. We question the validity of such aninterpretation, because low CCN concentrationstypical of the boundary layer of the open oceanpoorly correspond toAOD (23) and thereforeAODalso poorly corresponds to Nd. It follows that thepoor correlation between AOD and CCN weak-ens theNd-AOD relationships. Instead, we ascribethe variability in Nd directly to CCN by using Nd

of the convective cores and correcting it forWb,while accounting formost of the variability inme-teorology as manifested by CGT, Dq, and CTRC.

Because Nd and LWPcore represent the cloudcores, which are the closest to adiabatic, in-creasingNd for a given LWPcore decreases re ina known way under an assumption of a fixedadiabatic fraction. The points where re reachesthe precipitation initiation value of 14 mm areconnected by the dotted line in Fig. 3A. Theclouds to the left of the dotted line have smallerNd and therefore re > 14 mm. Because the coa-lescence rate is proportional to re

5, the cloudswith re > 14 mm precipitate substantially (rainrate > ~2 mm/day), whereas precipitation in

clouds to the right of the dotted line is mostlysuppressed as a result of the small cloud dropletsize whenNd becomes large. It is noteworthy thatthe nearly full cloud cover for LWPcore≥ 240 gm−2

breaks up abruptly with the onset of precipita-tion, as indicated by re exceeding 14 mm atNd <100 cm−3. The abruptness of the breakup isattributed to the positive feedback of scavengingthe CCNby the precipitation (13). This, alongwiththe fact that MSC decks break up when startingto precipitate heavily, indicates that Cf is deter-mined to a large extent by the CCN control onNd.

Rosenfeld et al., Science 363, eaav0566 (2019) 8 February 2019 5 of 9

Table 1. The susceptibility of Cf, LWP, and CRE to Nd, based on the slopes of the lines of figs.S1 and S2.The numbers are based on the means and standard deviations of the slopes of the lines

for the different CGTs in each panel.

LTS All LTS > 18 K 18 K ≥ LTS ≥ 14 K LTS < 14 K

@ln(Cf)/@ln(Nd) 0.38 ± 0.06 0.32 ± 0.07 0.34 ± 0.06 0.23 ± 0.06. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .

l = @ln(LWP)/@ln(Nd) 0.03 ± 0.07 0.03 ± 0.08 0.01 ± 0.08 0 ± 0.04. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .

@ln(CRE)/@ln(Nd) 0.76 ± 0.08 0.68 ± 0.08 0.69 ± 0.07 0.56 ± 0.04. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .

Table 2. Multiple regression of Cf, LWP, and CRE with log10(Nd) or log10(Nd/Wb0.5), CGT, LTS

or Dq, and CTRC.The regression is applied to the means of the bins, with direct Nd or with theupdraft normalized Nd/Wb

0.5, which serves as a proxy to CCN concentrations. The contribution to R2

for each of the independent variables is the decrease in total R2 with that variable omitted. The sums

of individual contributions of R2 are adjusted to equal the total R2.

Total R2 RMS error log10(Nd) CGT LTS CTRC

Cf 0.92 0.064 0.31 0.41 0.13 0.07. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .

LWP 0.95 9.27 g m−2 0.01 0.91 0.01 0.02. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .

CRE 0.93 11.0 W m−2 0.37 0.40 0.09 0.07. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .

. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .

Total R2 RMS error log10(Nd) CGT Dq CTRC

Cf 0.93 0.059 0.39 0.37 0.12 0.05. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .

LWP 0.97 7.13 g m−2 0.02 0.91 0.01 0.03. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .

CRE 0.94 9.74 W m−2 0.45 0.36 0.07 0.06. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .

. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .

Total R2 RMS error log10(Nd/Wb0.5) CGT LTS CTRC

Cf 0.92 0.060 0.29 0.43 0.13 0.07. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .

LWP 0.95 8.80 g m−2 0.01 0.91 0.01 0.03. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .

CRE 0.93 9.81 W m−2 0.37 0.40 0.09 0.07. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .

. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .

Total R2 RMS error log10(Nd/Wb0.5) CGT Dq CTRC

Cf 0.93 0.054 0.40 0.38 0.10 0.05. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .

LWP 0.97 6.71 g m−2 0.01 0.91 0.01 0.04. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .

CRE 0.94 8.33 W m−2 0.46 0.36 0.06 0.06. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .

RESEARCH | RESEARCH ARTICLE

Erratum 20 June 2019. See erratum. on M

arch 10, 2021

http://science.sciencemag.org/

Dow

nloaded from

The meteorological control of Cf is known tobe related mostly to LTS at seasonal time scalesand regional space scales, because the formationof MSCs depends on a strong capping inversionwith dry air above it (32). Indeed, fig. S3A showsthat for a givenNd, Cf increases by a factor of ~3from low to high LTS. Dry air above cloud topsallows strong CTRC, which helps tomaintain the

cloud cover. This is captured by the strong indi-cated increase of Cf with CTRC (fig. S3B), espe-cially for the weakly precipitating clouds (largeNd). The Cf of such clouds depends strongly onmeteorological factors depicted here by CGT,LTS, and CTRC. However, Cf and LWP decreasesubstantially when clouds start to precipitatesubstantially (> ∼2 mm/day), as indicated by

re > 14 mm, especially when they are deeperthan ~300m (Fig. 3, A and B). A full cloud coverrequires large LTS, CTRC rate < ~ –50 W m−2,CGT > ~300 m, and sufficiently large Nd to sup-press precipitation (>60 to 100 cm−3, more fordeeper clouds).Although aerosol effects on MSC cloud cover

are relatively well documented (13, 36, 37), aero-sol effects on cumulus (Cu) clouds have beenmore elusive. To explore the role of cloud re-gimes in the aerosol effects, we repeated theanalysis shown in Fig. 3 for Cu (LTS < 14 K),MSC (LTS > 18 K), and a transition cloud regimefor 18 K ≥ LTS ≥ 14 K (Fig. 4). The overall shapeof the relationships of cloud properties with Nd

shown in Fig. 3 remains similar for both MSCand Cu, although the magnitude of the slopesdecreases gradually to slightly more than halfwhen transitioning from MSC to Cu. Quantifi-cation of the slopes tells the same story. Accord-ing to Table 1, the susceptibility of Cu propertiesto Nd is similar to the susceptibility of MSC,although somewhat weaker. A salient feature isthat Cf ofMSC reaches unity at largeNd and CGT(Fig. 4D), whereas Cf of Cu cannot exceed 0.7 atanyNd and CGT. The transitions fromMSC to Cuis gradual and depends on LTS, whereas thedependence of the average LWP, Cf, and CREis scaled down with smaller LTS. This situa-tion calls for a quantitative formulation of thisdependence.ExpandingFigs. 3 and4 to the four-dimensional

dependence of CF on Nd, CGT, LTS, and CTRCrequires a multiple regression procedure, whichallows a polynomial fit of second order. BecauseCGT, LTS, and CTRC are not completely inde-pendent, the regression procedure computes onlythe independent component that each variablecontributes. The procedure was applied to thebinned data. The central values of the individualbins are shown by the dots on the curves andclassifier intervals in Fig. 3 and fig. S3. To averageout the effect of cloud self-organization on Cf,we incorporated into the regression only binsthat contained at least 10 scenes (see Fig. 1B).Replacing LTS with Dq yielded larger R2 values,probably because Dq is much more focused onthe cloud tops than is LTS. The results witheither LTS or Dq are shown in Table 2.Table 2 shows that four parameters explain

up to 95% of the variability in CRE.When usingDq, Nd alone explains 46% of the variability inCRE. Its importance is higher than that of CGT(36%), despite thicker clouds having larger Cf,when everything else is held constant. Dq andCTRC, whose high values support the formationof MSC decks, explain 7% and 6% in the var-iability of CRE, respectively. Less than 6% ofCRE remains unexplained by the combinationof Nd, CGT, CTRC, and Dq or LTS.Canmeteorology explain some of the variabil-

ity in Nd, thus serving as an alternative expla-nation to the implied high contribution of CCNto R2 of Cf? That can happen by meteorology in-curring stronger Wb for greater Cf, which thenincreasesNd for the sameCCN. This question canbe addressed by accounting for the effect of Wb

Rosenfeld et al., Science 363, eaav0566 (2019) 8 February 2019 6 of 9

Clean Polluted

CGT=700

CGT= 300

CGT=150

Nd=20 Nd=100 Nd=250

Fig. 5. A conceptual representation of the relationships between Nd and cloud properties formarine boundary layer clouds of varying thicknesses. Meteorology determines CGT, whereasCCN determines Nd. Cloud fraction and LWP become larger for thicker and more polluted clouds.Thedevelopment of precipitation is associated with the decrease in Cf and LWP, likely due to raindepleting cloud water and causing evaporative downdrafts that break up the cloud cover wheninteracting with each other.The rain intensity increases with cloud drop effective radius at its top andwith cloud depth. Therefore, deeper clouds start precipitating and breaking up at larger Nd thanshallower clouds.

Table 3. Multiple regression with polynomial fit of the second order for Cf with log10(Nd/Wb0.5),

Dq, and CTRC for different intervals of CGT.The values in columns 4 to 6 are the partial R2 (or

fraction of explained variability of Cf) attributed to Nd/Wb0.5, Dq, and CTRC. The sums of individual

contributions of R2 are adjusted to equal the total R2.

CGT (m) Total R2 RMS error log10(Nd/Wb0.5) Dq (K)

CTRC

(W m−2)Number of scenes

Cf.. .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. .

0 to 150 0.91 0.04 0.74 0.11 0.06 51,935.. .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. .

150 to 300 0.93 0.04 0.71 0.13 0.09 138,626.. .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. .

300 to 450 0.95 0.04 0.62 0.20 0.13 193,831.. .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. .

450 to 600 0.91 0.06 0.68 0.15 0.08 181,566.. .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. .

600 to 800 0.88 0.06 0.62 0.15 0.11 98,230.. .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. .

LWP (g m−2).. .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. .

0 to 150 0.59 1.5 0.33 0.15 0.11 51,935.. .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. .

150 to 300 0.85 2.7 0.12 0.04 0.69 138,626.. .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. .

300 to 450 0.76 5.3 0.13 0.12 0.51 193,831.. .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. .

450 to 600 0.74 7.2 0.10 0.13 0.51 181,566.. .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. .

600 to 800 0.75 11.0 0.08 0.19 0.48 98,230.. .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. .

CRE (W m−2).. .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. .

0 to 150 0.93 3.2 0.82 0.09 0.02 51,935.. .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. .

150 to 300 0.93 6.7 0.67 0.10 0.16 138,626.. .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. .

300 to 450 0.96 6.8 0.66 0.12 0.18 193,831.. .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. .

450 to 600 0.94 9.1 0.76 0.06 0.12 181,566.. .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. .

600 to 800 0.94 9.3 0.74 0.04 0.16 98,230.. .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... .. ... ... .. ... ... .. ... ... .. .

RESEARCH | RESEARCH ARTICLE

Erratum 20 June 2019. See erratum. on M

arch 10, 2021

http://science.sciencemag.org/

Dow

nloaded from

onNd. The possibleWb-induced enhancement ofNd may be neutralized by dividing Nd by theupdraft enhancement factor of Nd. The outcomeisNd/Wb

0.5, which serves as a proxy for the CCNconcentrations. The results of the applicationof the regression to Nd/Wb

0.5 are shown in thelower half of Table 2. The overall R2 values ofCf, LWP, or CRE with Nd/Wb

0.5 are practicallyequal to the R2 with Nd. This shows that themeteorological factors that affectWb (CBH andCTRC) cannot provide an alternative explana-tion to the large implied aerosol effects on cloudproperties.The polynomial regressions described inTable 2

are in fact an observationally based simplifiedformulation of parameterization of the prop-erties of marine boundary layer clouds. The co-efficients of the equations are given in table S2.Because CGT reflects almost three-fourths

of the variability in CRE due to meteorology,the analysis was extended to classes of CGT,

as shown in Table 3. About 73% of the variabilityin CRE for clouds with a given CGT is explainedby the CCN surrogate Nd/Wb

0.5, and up to 82%for the shallowest clouds. The total explainedvariability in CREwhen also considering Dq andCTRC can be as much as 95%. Using Nd insteadofNd/Wb

0.5 (table S1) yields systematically lowerR2 by a few percent. This underscores the largerrelevance of the retrieved proxy to CCN relativeto the retrieved Nd.

Discussion and summary

The emerging capability to ascribe variability inCCN to satellite-retrieved Nd of the convectivecores allows the use of this Nd as a faithful sur-rogate for CCN. This was applied to analyses ofboundary layer shallow (<800m) clouds over theoceans between 0° and 40°S. Themeteorologicalfactors that were found to drive variability in Cf,LWC, and CRE are CGT, CTRC, and Dq. An illus-tration of the conceptual model for the main

findings is shown in Fig. 5. The quantitativefindings are as follows:1) Almost all (95%) of the variability in CRE

and its components (Cf and LWP) is explainedby CGT, Nd, Dq (or LTS), and CTRC.2) Aerosols, encapsulated byNd, explain 45%

of the variability in CRE; 50% is explained bymeteorology as represented by CGT, Dq (or LTS),and CTRC, leaving only 5% of CRE unexplained.3) Variability in Nd can explain ~25% of the

variability in LWP and nearly 40% of the varia-bility of Cf.4) For a given CGT, which encapsulates most

of the meteorological effect, nearly three-fourthsof the variability in CRE is explained by varia-bility in Nd.5) For a given CGT,Nd explains nearly half of

the variability in LWP and nearly two-thirds ofthe variability in Cf.6) The effect of Nd on Cf appears to be medi-

ated mainly by its control on coalescence and

Rosenfeld et al., Science 363, eaav0566 (2019) 8 February 2019 7 of 9

Fig. 6. Conventional representation of the relationships between average LWP and average Nd for equatorial and high-latitude areas. Latituderanges are shown at tops of panels. (A and B) A slight negative slope in the equatorial areas and a slight positive slope in high latitudes are indicated, incontrast to the large positive slopes in Figs. 3 and 4. (C and D) CGTdecreases systematically with increasing Nd for reasons explained in the text.

RESEARCH | RESEARCH ARTICLE

Erratum 20 June 2019. See erratum. on M

arch 10, 2021

http://science.sciencemag.org/

Dow

nloaded from

precipitation, which breaks up the cloud decksand depletes LWP.The amount of explained variability in CRE

and its components increases even further bya few percent when using the CCN proxy Nd/Wb

0.5 instead of Nd.The susceptibility of Cf, LWP, and CRE to Nd

is quantified in Table 1. The values are far largerthan those previously determined by observa-tions and simulations (38–40). Possible devia-tions from the retrieval assumption of adiabaticcores due to precipitation are more likely formore heavily precipitating cloudswith lowerNd.In such cases, lower Nd values have a smalleradiabatic fraction and a respectively greaterunderestimate relative to largerNd values. Cor-recting such bias inNd would increase the slopeseven further, rendering the numbers in Table 1 aslower bounds. This apparent disagreement withprevious reports has at least two causes: (i) Ourmethodology selects Nd of the convective coresthat is shown to best represent CCN, whereas pre-vious studies used area-averaged Nd; (ii) the sen-sitivity to CCNmust be enhanced after isolating itfrom meteorology, which dilutes its magnitude.The study encompasses a wide range of cloud

regimes, fromMSC under the stable subtropicalhighs to tropical Cu in unstable air masses, asis evident in Fig. 2B. The analysis here showsa large susceptibility of Cf, LWP, and CRE tochanges in Nd and CCN. The susceptibility be-haves qualitatively similarly in all cloud regimes,with a continuous quantitative transition fromtheMSC regime to the Cu regime, as captured bythe relationships shown in Fig. 3 and Table 1.According to Table 3, this allows prediction ofup to 95%of the clouds’ Cf, LWP, and CRE basedon knowledge of the CCN andmeteorology. Themeteorology is quantified by CGT, CTRC, LTS,and/or Dq. The quantitative relationships canserve as a simplified parameterization formarineboundary layer clouds. The relationships aregiven in tables S2 and S3 as the coefficients ofsecond-order polynomial equations.Observations (38, 40, 41) and simulations

(42–44) reported a generally negative responseof LWP to Nd in the region of precipitatingshallow convective clouds, as quantified by l =@ln(LWP)/@ln(Nd), where l is the LWP sus-ceptibility to Nd. This neutralizes to someextent the cloud albedo effect (45). The samestudies showed that l becomes more positivein less precipitating clouds. This portrayed arather weak overall aerosol effect, which standsin contrast to the results of this study.The apparent difference with respect to pre-

vious reports is most conspicuous with respectto l. The reported values of l are negativemostlyin convective precipitating clouds (40, 44), incontrast to the highly positive values in this study(Table 2 and Fig. 4F). To explain this apparentdiscrepancy, the data used in this studywere alsoused to reproduce the relationships in figure 1 of(44), shown here as Fig. 6. The average LWP wasrelated to the averageNd of all cloudy pixels withvalid MODIS retrievals having re product uncer-tainty < 10%. Similar values of l ≈ –0.14 and –0.2

were obtained for the equatorial central Pacificin the analysis here (Fig. 6A) and that of (44),respectively. The clouds at high latitudes ofthe central South Pacific have similar andslightly positive values, with l ≈ 0.2 and 0.1 inFig. 6B and (44), respectively. This captures amajor element of the spatial distribution of l in(44) and adds credibility to our reproduction oftheir results.Evidently, the strong positive l occurs only

while stratifying the data by CGT, as done herein Figs. 3 and 4. Not constraining l by CGT canlead to large changes in l that are not necessarilyrelated to aerosol effects, because for the sameNd, LWP increases stronglywith increasing CGT(Fig. 3B and Fig. 4, D to F). Indeed, for the dataof Fig. 6, A and B, CGT decreases systematicallywith increasing Nd (Fig. 6, C and D). It incurs atrend of decreasing LWP with increasing Nd—that is, a negative bias to l that is not related toaerosol effects but rather to a systematic decreas-ing cloud depth with increasing Nd.The trend of smaller Nd with deeper clouds

(i.e., greater CGT) is systematic and occurs evenat high latitudes, where l is slightly positive(Fig. 6, B and D). This requires an explanation.We hypothesize that the following mechanismmaintains this relationship, according to thefollowing logic:1) Deeper clouds have larger LWP, because

LWP º CGT2.2) Therefore, for a given Nd, cloud top re in-

creases with increasing CGT.3) Because the coalescence rate is proportional

to re5 (4), rain scavenging of aerosols increases

strongly with increasing CGT.4) Therefore, deeper clouds (larger CGT) due

tometeorological causes are likely to precipitatemore and to have smaller Nd.5) Possible deviations from the retrieval as-

sumption of adiabatic cores due to precipitationwould cause a larger underestimate of CGT formore heavily precipitating clouds with lowerNd.Therefore the indicated decreasing trend of CGTin Fig. 6, C and D, is the lower bound for themagnitude of this trend.This reasoning breaks down at very small

values ofNd, which are caused by excessive aero-sol scavenging by precipitation. An additionalfactor that comes into play in these ultracleanconditions is the suppression of convection dueto scarcity of aerosols that incurs large vaporsupersaturation on the expense of condensationand latent heat release. This was documented tooccur from marine boundary layer clouds (46)to deep tropical convective clouds (47) and likelyexplains the much smaller CGT and LWP valuesin the lowest bin of Nd in Fig. 6.This implies that the abundance of negative

l has beenmisinterpreted as a negative aerosolcloud–mediated radiative effect. In analogy tothe aerosol cloud albedo (Twomey) effect beingdefined only for a fixed LWP (45), the LWP andCf effects must be defined only for a fixed CGT.This is, in fact, what was done in this study. Thismakes it possible to go beyond the Twomeyeffect (45) and properly quantify the aerosol-

mediated CRE on LWP and Cf, and eventually onthe total solar CRE.In principle, a negative lmakes physical sense

only when the aerosol-induced entrainment ef-fect (consumption of LWP) outweighs the aero-sol lifetime effect. Such a condition is unlikely tobe met in precipitating cloud regimes becausethe precipitation considerably suppresses theentrainment rate by stabilizing the planetaryboundary layer (48, 49). The susceptibility of CftoNd is likely to be larger thanpreviously thought,as found here for l. The reported negative sus-ceptibilities were ascribed to faster evaporation ofsmaller drops when Nd is larger (40–44). Thisprobably happens when Nd is sufficiently largeto suppress rain. This is evident in Fig. 4, F andI, where LWP and Cf of nonprecipitating cumu-lus clouds (Nd > ~130 cm−3) stop increasing oreven slightly decrease with added Nd.Assessments of the susceptibility by running

large eddy simulations (LESs) have shown rathersmall to negative susceptibility, especially forconvective clouds [e.g., (42, 43)], which is at oddswith our results. LESs are usually analyzed afterequilibrium is reached with respect to all theadjustment processes, and the field of cloudproperties becomes steady-state or oscillatingaround a time-invariantmean state. By that time,the effects of added aerosols aremostly bufferedby changes in the cloud system. For example,adding aerosols suppresses rain, so the cloudsrespond by deepening just enough to restore therain amount that was suppressed [see the con-ceptual model in figure 12 of Seifert et al. (43)].But the steady state with high aerosol concen-trations cannot exist, because many of the aero-sol particleswould have been cleansed by the timethat equilibriumwould have been reached after2 to 3 days, as shown by Goren and Rosenfeld(17). Furthermore, clouds in nature rarely reachequilibrium. Dagan et al. (50) state that “Theenvironmental conditions and the cloud fieldproperties change faster than the time it takesto reach an equilibrium state. Therefore, cloudfields are likely to be in a transient state andcan be highly susceptible to changes in aerosolproperties.”This is evident, for example, in figure 3of (43), where adding drop concentrations inseveral steps from 35 to 105 cm−3 is followed bya respectively larger suppression of rain duringthe first 20 hours, but all converge to the samemean rainfall after 40 hours. However, in realitythe typical lifetime of fields of warm convectiveclouds is shorter than 12 hours (50). Nonetheless,using the methodology of this study to analyzethe simulated fields even after equilibriumwouldlikely still showpositive l for a fixed CGTbecausethe adjustment processes affect CGT, amongothercloud properties.If the reported observed large sensitivity of

shallow marine clouds to aerosols were incor-porated in GCMs, they likely would simulateglobal cooling, whereas the world is actuallywarming. This argument has been used to dis-miss such large sensitivities. For example, thesensitivity in the European ECHAM model istuned down by setting a minimumNd to 40 cm−3,

Rosenfeld et al., Science 363, eaav0566 (2019) 8 February 2019 8 of 9

RESEARCH | RESEARCH ARTICLE

Erratum 20 June 2019. See erratum. on M

arch 10, 2021

http://science.sciencemag.org/

Dow

nloaded from

because it was perceived that a lower value wouldmake the model overly sensitive to aerosols asa result of too much cooling (51). Accepting thelarge sensitivity revealed in this study impliesthat aerosols incur another large positive forcing,possibly through the deep clouds. However, sucha forcing cannot be accounted for in currentmod-els because of a lack of computational capacity,nor are observations sufficiently capable of relat-ing aerosols to forcing mediated by deep clouds.

REFERENCES AND NOTES

1. S. Twomey, The nuclei of natural cloud formation part II: Thesupersaturation in natural clouds and the variation of clouddroplet concentration. Geofis. Pura Appl. 43, 243–249 (1959).doi: 10.1007/BF01993560

2. B. Stevens, G. Feingold, Untangling aerosol effects on cloudsand precipitation in a buffered system. Nature 461, 607–613(2009). doi: 10.1038/nature08281; pmid: 19794487

3. D. Painemal, P. Zuidema, Assessment of MODIS cloud effectiveradius and optical thickness retrievals over the SoutheastPacific with VOCALS-REx in situ measurements. J. Geophys.Res. 116, D24206 (2011). doi: 10.1029/2011JD016155

4. E. Freud, D. Rosenfeld, Linear relation between convectivecloud drop number concentration and depth for rain initiation.J. Geophys. Res. 117, D02207 (2012). doi: 10.1029/2011JD016457

5. H. Gerber, H. Gerber, Microphysics of Marine StratocumulusClouds with Two Drizzle Modes. J. Atmos. Sci. 53, 1649–1662(1996). doi: 10.1175/1520-0469(1996)053<1649:MOMSCW>2.0.CO;2

6. M. B. Pinsky, A. P. Khain, Effects of in-cloud nucleationand turbulence on droplet spectrum formation in cumulus clouds.Q. J. R. Meteorol. Soc. 128, 501–533 (2002). doi: 10.1256/003590002321042072

7. M. C. vanZanten, B. Stevens, G. Vali, D. H. Lenschow,Observations of Drizzle in Nocturnal Marine Stratocumulus.J. Atmos. Sci. 62, 88–106 (2005). doi: 10.1175/JAS-3355.1

8. R. Chen, R. Wood, Z. Li, R. Ferraro, F.-L. Chang, Studying thevertical variation of cloud droplet effective radius usingship and space-borne remote sensing data. J. Geophys. Res.113, D00A02 (2008). doi: 10.1029/2007JD009596

9. T. V. Prabha et al., Microphysics of Premonsoon and MonsoonClouds as Seen from In Situ Measurements during theCloud Aerosol Interaction and Precipitation EnhancementExperiment (CAIPEEX). J. Atmos. Sci. 68, 1882–1901 (2011).doi: 10.1175/2011JAS3707.1

10. D. Rosenfeld, H. Wang, P. J. Rasch, The roles of cloud dropeffective radius and LWP in determining rain properties inmarine stratocumulus. Geophys. Res. Lett. 39, L13801 (2012).doi: 10.1029/2012GL052028

11. M. Konwar et al., Aerosol control on depth of warm rain inconvective clouds. J. Geophys. Res. 117 (D13), D13204 (2012).doi: 10.1029/2012JD017585

12. R. C. Braga et al., Further evidence for CCN aerosolconcentrations determining the height of warm rain and iceinitiation in convective clouds over the Amazon basin.Atmos. Chem. Phys. 17, 14433–14456 (2017). doi: 10.5194/acp-17-14433-2017

13. D. Rosenfeld, Y. J. Kaufman, I. Koren, Switching cloud coverand dynamical regimes from open to closed Benard cellsin response to the suppression of precipitation by aerosols.Atmos. Chem. Phys. 6, 2503–2511 (2006). doi: 10.5194/acp-6-2503-2006

14. H. Wang, G. Feingold, Modeling Mesoscale Cellular Structuresand Drizzle in Marine Stratocumulus. Part I: Impact ofDrizzle on the Formation and Evolution of Open Cells. J. Atmos.Sci. 66, 3237–3256 (2009). doi: 10.1175/2009JAS3022.1

15. G. Feingold, I. Koren, H. Wang, H. Xue, W. A. Brewer,Precipitation-generated oscillations in open cellular cloudfields. Nature 466, 849–852 (2010). doi: 10.1038/nature09314; pmid: 20703303

16. T. Goren, D. Rosenfeld, Decomposing aerosol cloud radiativeeffects into cloud cover, liquid water path and Twomeycomponents in marine stratocumulus. Atmos. Res. 138, 378–393(2014). doi: 10.1016/j.atmosres.2013.12.008

17. T. Goren, D. Rosenfeld, Extensive closed cell marinestratocumulus downwind of Europe-A large aerosol cloudmediated radiative effect or forcing? J. Geophys. Res. 120,6098–6116 (2015).

18. T. Nakajima, A. Higurashi, K. Kawamoto, J. E. Penner,A possible correlation between satellite-derived cloud andaerosol microphysical parameters. Geophys. Res. Lett.28, 1171–1174 (2001). doi: 10.1029/2000GL012186

19. V. N. Kapustin et al., On the determination of a cloudcondensation nuclei from satellite: Challenges and possibilities.J. Geophys. Res. 111, D04202 (2006). doi: 10.1029/2004JD005527

20. A. Clarke, V. Kapustin, Hemispheric aerosol vertical profiles:Anthropogenic impacts on optical depth and cloud nuclei.Science 329, 1488–1492 (2010). doi: 10.1126/science.1188838;pmid: 20847262

21. J. Quaas, B. Stevens, P. Stier, U. Lohmann, Interpreting thecloud cover–aerosol optical depth relationship found insatellite data using a general circulation model. Atmos. Chem.Phys. 10, 6129–6135 (2010). doi: 10.5194/acp-10-6129-2010

22. D. Rosenfeld et al., Global observations of aerosol-cloud-precipitation-climate interactions. Rev. Geophys. 52, 750–808(2014). doi: 10.1002/2013RG000441

23. P. Stier, Limitations of passive remote sensing to constrainglobal cloud condensation nuclei. Atmos. Chem. Phys. 16,6595–6607 (2016). doi: 10.5194/acp-16-6595-2016

24. E. Gryspeerdt, J. Quaas, N. Bellouin, Constraining the aerosolinfluence on cloud fraction. J. Geophys. Res. Atmos. 121,3566–3583 (2016). doi: 10.1002/2015JD023744

25. I. Koren, J. V. Martins, L. A. Remer, H. Afargan, Smokeinvigoration versus inhibition of clouds over the Amazon.Science 321, 946–949 (2008). doi: 10.1126/science.1159185;pmid: 18703738

26. R. C. Levy et al., The Collection 6 MODIS aerosol products overland and ocean. Atmos. Meas. Tech. 6, 2989–3034 (2013).doi: 10.5194/amt-6-2989-2013

27. Y. Shinozuka et al., The relationship between cloud condensationnuclei (CCN) concentration and light extinction of driedparticles: Indications of underlying aerosol processes andimplications for satellite-based CCN estimates. Atmos. Chem.Phys. 15, 7585–7604 (2015). doi: 10.5194/acp-15-7585-2015

28. D. Rosenfeld et al., Satellite retrieval of cloud condensationnuclei concentrations by using clouds as CCN chambers.Proc. Natl. Acad. Sci. U.S.A. 113, 5828–5834 (2016).doi: 10.1073/pnas.1514044113; pmid: 26944081

29. Y. Zhu, D. Rosenfeld, Z. Li, Under what conditions can we trustretrieved cloud drop concentrations in broken marinestratocumulus? J. Geophys. Res. Atmos. 123, 8754–8767(2018). doi: 10.1029/2017JD028083

30. Y. Zheng, D. Rosenfeld, Linear relation between convectivecloud base height and updrafts and application to satelliteretrievals. Geophys. Res. Lett. 42, 6485–6491 (2015).doi: 10.1002/2015GL064809

31. Y. Zheng, D. Rosenfeld, Z. Li, Quantifying cloud base updraftspeeds of marine stratocumulus from cloud top radiativecooling. Geophys. Res. Lett. 43, 11407–11413 (2016).doi: 10.1002/2016GL071185

32. S. A. Klein, D. L. Hartmann, The Seasonal Cycle of LowStratiform Clouds. J. Clim. 6, 1587–1606 (1993). doi: 10.1175/1520-0442(1993)006<1587:TSCOLS>2.0.CO;2

33. H. Andersen, J. Cermak, J. Fuchs, R. Knutti, U. Lohmann,Understanding the drivers of marine liquid-water cloudoccurrence and properties with global observations usingneural networks. Atmos. Chem. Phys. 17, 9535–9546 (2017).doi: 10.5194/acp-17-9535-2017

34. D. P. Grosvenor, O. Sourdeval, R. Wood, Parameterizing cloudtop effective radii from satellite retrieved values, accountingfor vertical photon transport: Quantification and correctionof the resulting bias in droplet concentration and liquid waterpath retrievals. Atmos. Meas. Tech. 11, 4273–4289 (2018).doi: 10.5194/amt-11-4273-2018

35. Y. Zheng, thesis, University of Maryland, College Park (2018).doi: 10.13016/M2RV0D40V

36. B. A. Albrecht, Aerosols, cloud microphysics, and fractionalcloudiness. Science 245, 1227–1230 (1989). doi: 10.1126/science.245.4923.1227; pmid: 17747885

37. T. Goren, D. Rosenfeld, Satellite observations of ship emissioninduced transitions from broken to closed cell marinestratocumulus over large areas. J. Geophys. Res. 117, D17206(2012). doi: 10.1029/2012JD017981

38. F. F. Malavelle et al., Strong constraints on aerosol-cloudinteractions from volcanic eruptions. Nature 546, 485–491(2017). doi: 10.1038/nature22974; pmid: 28640263

39. S. Zhang et al., On the characteristics of aerosol indirecteffect based on dynamic regimes in global climate models.Atmos. Chem. Phys. 16, 2765–2783 (2016). doi: 10.5194/acp-16-2765-2016

40. T. Michibata, K. Suzuki, Y. Sato, T. Takemura, The source ofdiscrepancies in aerosol–cloud–precipitation interactionsbetween GCM and A-Train retrievals. Atmos. Chem. Phys. 16,15413–15424 (2016). doi: 10.5194/acp-16-15413-2016

41. Y.-C. Chen, M. W. Christensen, G. L. Stephens, J. H. Seinfeld,Satellite-based estimate of global aerosol–cloud radiativeforcing by marine warm clouds. Nat. Geosci. 7, 643–646(2014). doi: 10.1038/ngeo2214

42. Z. J. Lebo, G. Feingold, On the relationship between responsesin cloud water and precipitation to changes in aerosol.Atmos. Chem. Phys. 14, 11817–11831 (2014). doi: 10.5194/acp-14-11817-2014

43. A. Seifert, T. Heus, R. Pincus, B. Stevens, Large-eddysimulation of the transient and near-equilibrium behavior ofprecipitating shallow convection. J. Adv. Model. Earth Syst. 7,1918–1937 (2015). doi: 10.1002/2015MS000489

44. Y. Sato et al., Aerosol effects on cloud water amounts weresuccessfully simulated by a global cloud-system resolving model.Nat. Commun. 9, 985 (2018). doi: 10.1038/s41467-018-03379-6;pmid: 29515125

45. S. Twomey, The Influence of Pollution on the Shortwave Albedoof Clouds. J. Atmos. Sci. 34, 1149–1152 (1977). doi: 10.1175/1520-0469(1977)034<1149:TIOPOT>2.0.CO;2

46. I. Koren, G. Dagan, O. Altaratz, From aerosol-limited toinvigoration of warm convective clouds. Science 344,1143–1146 (2014). doi: 10.1126/science.1252595;pmid: 24904161

47. J. Fan et al., Substantial convection and precipitationenhancements by ultrafine aerosol particles. Science 359,411–418 (2018). doi: 10.1126/science.aan8461;pmid: 29371462

48. R. Wood, Cancellation of Aerosol Indirect Effects in MarineStratocumulus through Cloud Thinning. J. Atmos. Sci. 64,2657–2669 (2007). doi: 10.1175/JAS3942.1

49. M. D. Lebsock, G. L. Stephens, C. Kummerow, Multisensorsatellite observations of aerosol effects on warm clouds.J. Geophys. Res. 113, D15205 (2008). doi: 10.1029/2008JD009876

50. G. Dagan, I. Koren, O. Altaratz, Y. Lehahn, Shallowconvective cloud field lifetime as a key factor for evaluatingaerosol effects. iScience 10, 192–202 (2018).doi: 10.1029/2008JD009876

51. C. Hoose et al., Constraining cloud droplet numberconcentration in GCMs suppresses the aerosol indirect effect.Geophys. Res. Lett. 36, L12807 (2009). doi: 10.1029/2009GL038568

ACKNOWLEDGMENTS

Funding: This research was supported by the Joint NSFC-ISFResearch Program (No. 41561144004), the National NaturalScience Foundation of China (No. 41575136). Y.Z. is supportedby the U.S. Department of Energy (DOE) Atmospheric SystemResearch program (DE-SC0018996). T.G. received funding fromthe European Union Horizon 2020 research and innovationprogram under the Marie Sklodowska-Curie grant agreement703880. M. W. is supported by the National Natural ScienceFoundation of China (No. 41575073, 41621005, and 91744208) andby the Jiangsu Collaborative Innovation Center of Climate Change.S.Y. is supported by the Department of Science and Technologyof China (No. 2016YFC0202702, 2018YFC0213506, and2018YFC0213503), National Research Program for Key Issuesin Air Pollution Control in China (No. DQGG0107), and NationalNatural Science Foundation of China (No. 21577126 and41561144004). Author contributions: D.R. designed this study;Y. Zhu acquired and processed the data; D.R. and Y. Zhu carriedout analyses, interpreted data, and wrote the manuscript;M.W. contributed to the comparisons with other studies and tothe discussion; Y. Zheng supported the calculations of cloudbase updrafts and contributed to the discussion; and T.G.and S.Y. contributed to the discussion. Competing interests:All authors have no competing interests. Data and materialsavailability: The MODIS data for this study are obtained fromNASA (https://search.earthdata.nasa.gov).

SUPPLEMENTARY MATERIALS

www.sciencemag.org/content/363/6427/eaav0566/suppl/DC1Supplementary TextFigs. S1 to S13Tables S1 to S3

19 September 2018; accepted 3 December 2018Published online 17 January 201910.1126/science.aav0566

Rosenfeld et al., Science 363, eaav0566 (2019) 8 February 2019 9 of 9

RESEARCH | RESEARCH ARTICLE

Erratum 20 June 2019. See erratum. on M

arch 10, 2021

http://science.sciencemag.org/

Dow

nloaded from

cloudsAerosol-driven droplet concentrations dominate coverage and water of oceanic low-level

Daniel Rosenfeld, Yannian Zhu, Minghuai Wang, Youtong Zheng, Tom Goren and Shaocai Yu

DOI: 10.1126/science.aav0566 (6427), eaav0566.363Science 

, this issue p. eaav0566; see also p. 580Scienceof other compensating warming effects.these clouds is much more sensitive to the presence of CCN than current models indicate, which suggests the existence abundance explained most of the variability in the radiative cooling. Thus, the magnitude of radiative forcing provided byreflect much of the solar radiation received by Earth back to space (see the Perspective by Sato and Suzuki). The CCN

analyzed how CCN affect the properties of marine stratocumulus clouds, whichet al.global temperatures? Rosenfeld How much impact does the abundance of cloud condensation nuclei (CCN) aerosols above the oceans have on

Reflections on cloud effects

ARTICLE TOOLS http://science.sciencemag.org/content/363/6427/eaav0566

MATERIALSSUPPLEMENTARY http://science.sciencemag.org/content/suppl/2019/01/16/science.aav0566.DC1

CONTENTRELATED

http://science.sciencemag.org/content/sci/364/6446/eaay4194.fullhttp://science.sciencemag.org/content/sci/363/6427/580.full

REFERENCES

http://science.sciencemag.org/content/363/6427/eaav0566#BIBLThis article cites 50 articles, 6 of which you can access for free

PERMISSIONS http://www.sciencemag.org/help/reprints-and-permissions

Terms of ServiceUse of this article is subject to the

is a registered trademark of AAAS.ScienceScience, 1200 New York Avenue NW, Washington, DC 20005. The title (print ISSN 0036-8075; online ISSN 1095-9203) is published by the American Association for the Advancement ofScience

Copyright © 2019, American Association for the Advancement of Science

on March 10, 2021

http://science.sciencem

ag.org/D

ownloaded from