Freshwater transport from warm to cold ocean regions amplifying faster than all model estimates

9 Warming-induced global water cycle changes pose a signiﬁcant challenge to global ecosystems and human society. The magnitude 10 of historical water cycle change is uncertain due to a dearth of direct rainfall and evaporation observations, particularly over 11 the ocean where 80% of global evaporation occurs. Air-sea ﬂuxes of freshwater and river run-oﬀ imprint on ocean salinity at 12 diﬀerent temperatures, such that warmer regions tend to be saltier and cooler regions tend to be fresher. In this work, we track 13 observed salinity trends in the warm, salty fraction of the ocean from 1970 to 2014, and infer the global poleward transport 14 of freshwater over this period. Since 1970, 46 – 77 × 10 12 m 3 of freshwater has been transported poleward from the warmest 15 fraction of the ocean. No model in the current generation of climate models (the 6th Climate Model Intercomparison Project; 16 CMIP6) replicates this transport, with the closest model underestimating transport by 2 – 4 times. We trace the climate model 17 biases to a weaker than expected surface freshwater ﬂux intensiﬁcation, just 0 – 4% in CMIP6 models compared to an estimated 18 3 – 7.5% in observations. 19

to preparing for and mitigating the impacts of climate change. However, direct observations of rainfall are 23 sparse, particularly over the ocean, which acts as a conduit for 80% of global freshwater transport [1,2]. Lim-24 ited rainfall and evaporation observations, combined with substantial variance across atmospheric reanalysis 25 products [3,4] have made direct measurements of long-term changes to the global water cycle challenging. 26 As a result, much of our understanding of changes to the water cycle over the past fifty years has been 27 shaped by proxies for rainfall measurements over the ocean including salinity. As water associated with the 28 hydrological cycle transits through the ocean, it generates a signature of ocean salinification (where evapo- 29 ration changes exceed precipitation and river runoff changes) and freshening (where precipitation and river 30 runoff changes exceed evaporation changes), both at the surface [5, 6] and the interior [7,8]. Using in-situ 31 observations and modelling of ocean salinity, a growing body of research has identified a 'wet gets wetter, 32 dry gets drier' paradigm, wherein the global water cycle intensifies due to global warming [9,5,10], with 33 some estimates suggesting a water cycle intensification of 2 -4% relative to the 1950 ocean state [8,6,11]. 34 This intensification manifests as increasingly salty sub-tropical oceans and increasingly fresh tropical and 35 sub-polar oceans. 36 The observed distribution of salinity change (or tendency) in the global ocean has a large scale coherent Stippling in b) and d) indicates regions where more than 2/3 of the CMIP6 models agree on the sign of the salinity tendency. Black contour lines indicate zonally averaged 'temperature-percentiles' with the warmest 2% of the ocean bound by the 2% contour, the warmest 6% by the 6% contour and so on. Here we track salinity changes within these percentile layers.  figure 1). Based on this realisation, past research has used water mass-based diagnostics to frame 49 ocean salinity changes, for instance, by quantifying mean salinity in temperature (or density) layers (known 50 as the 'T-S' curve of the ocean) [12,7,13].

51
Tracking the T-S curve over time provides information about global salinity changes while maintaining 52 the ability to differentiate between climatic regions using temperature. However, simply computing trends 53 of absolute salinity (S) at constant conservative temperature (T) ignores ocean warming which moves the 54 underlying T-S curve. [14]. Indeed, as the ocean warms, the coldest temperatures in the sub-polar and polar 55 oceans will disappear, while new tropical waters will be created, so a trend at constant T can be misleading. In 56 this work, we analyse salinity changes at constant 'temperature-percentile' rather than constant temperature 57 [15]. That is, we organise the ocean from warmest to coldest and track changes at fixed accumulated volume.

58
For example, by tracking salinity changes within the warmest 6% of the ocean volume, we are able to track a 59 region which retains the same volume and remains in roughly in the same climatic region even as the ocean 60 warms. This framework allows us to track changes in the T-S curve of the ocean while eliminating much 61 of the direct influence of ocean warming. The overlaid black contours in figure 1 illustrate the geographical 62 distribution of temperature-percentiles in the observations, from hot (0%) to cold (100%). Past research has 63 used temperature-percentiles to analyse heat content change, revealing that approximately 50% of surface 64 heat flux changes enter 90% of the ocean volume via the sub-polar ocean since 1970 [15].

65
Changes to the T-S curve may be attributed to changes in the build up of greenhouse gases (GHGs) and 66 aerosols in the atmosphere. GHGs tend to accelerate the warming of the climate system, thereby amplifying 67 the strength of the water cycle in a 'wet-gets-wetter-dry-gets-drier' pattern [10,4]. Anthropogenic and natural 68 aerosols, on the other hand, tend to cool the climate system on a global scale, leading to the opposite response 69 in the water cycle, with wet areas getting drier and dry areas getting wetter [10]. In essence, aerosols tend  In addition to the global T-S changes presented here, we may also refer to the basin-scale salinity tendencies 105 in figure 1 to explore hemispheric shifts in salinity. Indeed, figure 1 reveals a pattern of freshwater transport 106 from the North Atlantic to the North Pacific in both models and observations, with a stronger transport 107 in CMIP6 models compared to observations. Based on figure 2, the warmest 2% of the ocean experiences 108 salinfication in both the CMIP6 model and observations, and the warmest 6% of the ocean experiences 109 salinification in the observations. Consequently, we focus hereafter on tracking changes in the warmest 2% 110 and warmest 6% of the ocean.  In order to account for the volume of ocean water that experiences salinification in the observations and 116 models, we multiply mean salinity by layer volume to obtain a global freshwater content (as in equation 117 (6) in Methods). In the warmest 2% ocean by volume (figure 3a), the freshwater content decreases in both 118 the observations and the CMIP6 models, relative to a 1970-1980 baseline. By the end of the historical 119 period, the warmest 2% of the ocean has lost 1.7 -5.2 times more freshwater in the observations than in 120 the CMIP6 ensemble mean. That said, the rate of freshwater change in some CMIP6 models (calculated as 121 the slope of the linear regression over the historical period) is the same as that in three out of four of the 122 observational data sets (histograms in figure 3a). In the warmest 6% of the ocean (figure 3b), the observed 123 ocean loses substantially more freshwater than the CMIP6 models. Overall, the observations suggest a 124 poleward redistribution of freshwater of between 46 and 77×10 12 m 3 . On the other hand, many CMIP6 125 models (and the CMIP6 ensemble mean) experience a freshening over the historical period, with the closest 126 CMIP6 model experiencing a freshwater loss of about 1/3 of observations (18×10 12 m 3 ). There is also larger 127 spread between individual model realisations, with the models coalescing into two distinct groups (histograms 128 in figure 3b). A 'salty' group experiences little to no freshwater change over the historical period, while a 129 'fresh' group experiences freshening over the same period. The specific models which correspond to each 130 group are listed in Table S1 in the supplementary information.

131
Note that the difference between models and observations in the warmest 6% of the ocean may be due to a 132 difference in the pattern of salinity change rather than a difference in magnitude of salinity change. Figure 1 133 shows the magnitude of salinity change in the CMIP6 models is in fact greater than that in the observations, agricultural changes are based in part on estimates of water cycle change from CMIP6 models. That CMIP6 138 models cannot accurately capture historical salinity trends is therefore a cause of concern for future climate 139 change projections.

140
The systematic fresh bias in CMIP6 models must be urgently addressed in order to provide accurate future 151 climate projections. In order to sample a wider range of perturbations to the water cycle, we compare a set 152 of six DAMIP models that each performed a historical greenhouse gas forcing only (GHG-only) and a his-  The tracer-percentile framework 200 We begin by touching upon the formulation for the tracer-percentile framework in temperature-percentiles, 201 as laid out in detail by [15]. The volume of ocean warmer than a given conservative temperature Θ * is 202 expressed as: where Θ is the global temperature field and Θ * is the binning temperature chosen for the integration. We 204 invert equation (1) to express temperature as a function of volume, V, or equivalently, temperature-percentile, 205 p: where p = V/V T × 100 and V T is the total volume of the ocean. Θ(V, t) therefore represents the mean 207 temperature bounding a given volume, and Θ(p, t) represents the mean temperature bounding a given where S(x, y, z, t) is the three-dimensional absolute salinity field in the ocean. To find mean salinity at 216 temperature-percentiles, we interpolate S(Θ * , t) onto Θ(p, t) surfaces: 217 S(Θ * , t) =⇒ S(Θ(p, t), t) = S(p, t).
Mapping mean salinity changes at constant temperature-percentiles ensures we maintain the benefits of 218 adiabatic-invariant watermass methods while also being able to trace freshwater fluxes directly to climatic 219 regions (e.g., tropics, sub-tropics and sub-polar regions) at the ocean's surface.

220
The resulting variable S(p, t) provides a characteristic p − S curve for the ocean which, like T-S curves, In order to account for the volume of water in a given percentile layer experiencing salinity changes, we 226 calculate a freshwater content, F (p, t) by multiplying salinity by layer volume: where S(t) is the volume-averaged salinity that serves as the baseline differentiating 'fresh' and 'salty' water 228 in the ocean, and S 0 is the reference salinity, assumed to be S 0 = 35 g/kg. The global change in freshwater, 229 or freshwater flux F(p, t), can be expressed as: In practice, ∂S/∂t is approximated over the time period of interest as the slope of the linear trend in S(p, t). we begin by quantifying surface freshwater fluxes into volumes bounded by isotherms before interpolating 235 onto temperature-percentiles: where A is the surface area bounded by contours of temperature Θ * , ρ 0 is the reference density, assumed

255
Note that the IAP dataset is specifically formulated to reduce errors associated with sampling biases due to 256 the rapid introduction of Argo in the early 2000s (see [21] for more details), so we opt to use it where available.

257
The composite dataset therefore makes use of the IAP data in the upper ocean and fills the deep ocean with 258 EN4 data. All observed temperature and salinity measurements are converted to Conservative Temperature The analysis presented here enables direct comparisons between observations and climate models. We assess 272 the difference between observations and climate models using twenty climate models from the Climate Model

273
Intercomparison Project phase 6 [CMIP6; [28]] (model names and references are provided in the supplemen-274 tary information, Table S1). We focus on the historical experiments in the CMIP6 models (with masked 275 marginal seas) which branch off from the pre-industrial control (piControl ) runs from 1850 onwards. We GHGs in an effort to attribute the model response to either forcing field (see [29] for more details of the 284 DAMIP protocol). The list of DAMIP models used and relevant references (which include further details on 285 the model setup) may be found in the supplementary information (Table S2). As with the CMIP6 models,