We use an atmospheric general circulation model (GCM) developed at Laboratoire de Météorologie Dynamique, Paris, France with isotopes-tracking implement, called LMDZ-iso (Risi et al., 2010). LMDZ-iso is derived from the LMDz model (Hourdin et al., 2006) that has been used for numerous future and palaeoclimate studies (Ladant et al., 2014; Pohl et al., 2014; Sepulchre et al., 2006). Water in a condensed form and its vapour are advected by the Van Leer advection scheme (Van Leer, 1977). Isotopic processes in LMDZ-iso are documented in (Risi et al., 2010). Evaporation over land is assumed not to fractionate, given the simplicity of the model surface parameterisation (Risi et al., 2010). Yao et al. (2013) have provided a precise description of rainfall patterns over the TP, and showed LMDZ-iso ability to simulate atmospheric dynamics and reproduce rainfall and δ18O patterns consistent with data over this region.

LMDZ-iso is also equipped with water tagging capabilities, allowing us to quantify different moisture contributions from continental and oceanic evaporation sources. The advantage of this technique compared to typical back-trajectories methods is that it tracks the water rather than air masses, thus taking into account effects of phase changes. In our simulations five potential moisture sources are considered: (1) continental sources, (2) Indian Ocean, (3) Atlantic Ocean, (4) Mediterranean Sea, and (5) Pacific Ocean.

We use a model configuration with 96 grid points in longitude, 72 in latitude and 19 vertical layers, with the first four layers in the first kilometre above the surface. LMDZ-iso has a stretchable grid that allows increased spatial resolution over a defined region. In our case, it gives an averaged resolution of ∼ 100 km over central Asia, which is a good trade-off between a reasonable computing time and a spatial resolution that adequately represents main features of TP topography.

Here we report results from three experiments designed to isolate the influence of Asian topography on climate and isotopic composition of precipitation. Topography is derived from a 10 min US Navy dataset and interpolated to the model grid. The control run (MOD) is a pre-industrial run, i.e. initialised with boundary conditions (insolation, greenhouse gases, sea surface temperatures (SSTs), topography) kept at pre-industrial values. For the two other experiments, we keep all boundary conditions (including albedo, rugosity, and vegetation distribution) similar to those in MOD run, except for the topography. We reduce the altitude over the area covering the Tibetan Plateau, Himalayas and a part of surrounding mountains: Tian Shan, Pamir, Kunlun and Hindu Kush to 50 % of modern elevations (intermediate, INT case) and to 250 m elevation (low, LOW case) (Fig. 1). SSTs for all runs come from the AMIP dataset (monthly SSTs averaged from 1979 to 1996; Taylor et al., 2000). Each experiment has been run for 20 years. We analyse seasonal means over the last 18 years, as the two first years are extracted for spin-up.

Our goal is to understand to what extent topography changes explain the precipitation δ18O signal over TP (i.e. the direct topography effect) and what part of this signal depends on other climate processes. To do so, we develop a theoretical expression for the precipitation composition.

To the first order, the δ18O composition of the precipitation Rp follows that of the vapour Rv. Deviations from the vapour composition, ε=Rp-Rv, are associated with a local condensational or post-condensational process. Rp=Rv+ε In an idealized framework of an isolated air parcel transported from an initial site at low altitude to the site of interest (Fig. 2), the vapour composition can be predicted by Rayleigh distillation: Rv=Rv0×fα-1+dRv, where Rv0 is the initial composition of the vapour at the initial site; α is the fractionation coefficient that depends on temperature and on the water phase (Majoube, 1971; Merlivat and Nief, 1967); and f is the residual fraction of the vapour at the site of interest relatively to the initial site of an air mass ascent. We take the initial site as characterised by a temperature and humidity T0 and q0. Under these conditions, Rv0 is the theoretical isotopic composition of vapour that it would have if all the vapour originated from the local evaporation over quiescent oceanic conditions. Depending on the atmospheric circulation, on deep convective and mixing processes and on the source region of water vapour, the isotopic composition of vapour may deviate from the Rayleigh distillation by dRv.

The residual fractionfdepends on the specific humidity q at the site of interest: f=q/q0.

The air is not always saturated near the surface, therefore: q=h×qs(Ts), where h and Ts are the relative humidity and air temperature near the surface of the site of interest. The air can be under-saturated because it can be considered as air that has been transported adiabatically from the area of minimum condensation temperature, T* (Galewsky and Hurley, 2010; Galewsky et al., 2005; Sherwood, 1996): q=qs(T*).

The surface temperature can be predicted to first order by the adiabatic lapse rate, Γ, and is modulated by the non-adiabatic component dTs that represents processes such as large-scale circulation or radiation: Ts=T0+Γ×z-z0+dTs, where z and z0 are the altitudes at the site of interest and at the initial site. We use an adiabatic lapse rate equal to -5 km-1 based on the measurements of modern observed mean temperature lapse rate on the southern slope of the central Himalayas, that ranges from -4.7 to -6.1 km-1 for the monsoon season and from -4.3 to -5.5 km-1 for the rest of the year (Kattel et al., 2015).

If we combine Eqs. (1) to (5), we get that Rv is a function of ε, dRv, h, dTs and z: Rp=Rv0×h×qsT0+Γ×(z-zo+dTs)/q0α-1+dRv+ε. Or in a simpler form: Rp=Rpε,dRv,h,dTs,z. Parameters z0, q0, T0 are reference values that are common to all sites of interest, all climates and geographies. Even if initial conditions for the Rayleigh distillation vary depending on the atmosphere circulation, on deep convective processes and on the site of interest, we keep the same reference values and we consider all variations in initial conditions are accommodated by dRv.

This model is equivalent to that of Rowley et al. (2001) for dRv=0 (i.e. neglecting the effects of mixing and deep convection on the initial water vapour), ε=(α-1) ×Rv (i.e. neglecting post-condensational effects), and h=1 (i.e. assuming the site of interest is inside the precipitating cloud).

Our goal is to understand why Rp varies from one climatic state to another. We refer to these climatic states using subscript 1 and 2 and to their difference using the Δ notation. Differences between INT and LOW and between MOD and INT climatic states correspond to the initial and the terminate stages of the TP uplift respectively. We decompose ΔRp=Rp2-Rp1 into contribution from ΔdRv, Δε, Δh, ΔdTs, and Δz: ΔRp=ΔRp,Δε+ΔRp,ΔdRv+ΔRp,Δh+ΔRp,ΔdTs+ΔRp,Δz+N, where ΔRp,Δε, ΔRp,ΔdRv, ΔRp,Δh, ΔRp,ΔdTs, and ΔRp,Δz are the contributions of ΔdRv, Δε, Δh, ΔdTs, and Δz to ΔRp. Non linear terms of decomposition are gathered into the residual term N. Contributions are estimated using Eq. (7) (see also Table 1): Rp,Δε=Rp(ε2,dRv′,h′,dTs′,z′)-Rp(ε1,dRv′,h′,dTs′,z′)Rp,ΔdRv=Rp(ε′,dRv2,h′,dTs′,z′)-Rp(ε′,dRv1,h′,dTs′,z′)Rp,Δh=Rp(ε′,dRv′,h2,dTs′,z′)-Rp(ε′,dRv′,h1,dTs′,z′)Rp,ΔdTs=Rp(ε′,dRv′,h′,dTs2,z′)-Rp(ε′,dRv′,h′,dTs1,z′)Rp,Δz=Rp(ε′,dRv′,h′,dTs′,z2)-Rp(ε′,dRv′,h′,dTs′,z1). In order to decrease the sensitivity of the decomposition to the state at which it has been calculated we take z′, dTs′, h′, dRv′, and ε′ as centred differences: z′=(z2+z1)/2dTs′=(dTs2+dTs1)/2h′=(h2+h1)/2dRv′=(dRv2+dRv1)/2ε′=(ε2+ε1)/2.

Note that ε′ in Eqs (10) to (13) and dRv′ in Eqs. (9) and (11) to (13) can be replaced by 0 without changing the result. Parameters z, dTs, h, dRv, and ε are diagnosed for the climatic states 1 and 2 from LMDZ-iso simulations (e.g. for pairs of experiments, MOD and INT cases). Parameter ε is estimated as ε=Rp-Rv, where Rp and Rv are isotopic ratios simulated by LMDZ-iso. Parameter h is the relative humidity simulated by LMDZ-iso. Altitude z is a prescribed boundary condition of the simulations. Parameter dRv is estimated by calculating the difference between the water vapour isotopic ratio simulated by LMDZ-iso (Rv,LMDZ) and that predicted by Rayleigh distillation if the initial water vapour isotopic ratio is Rv0: dRv=Rv,LMDZ-Rv0×(q/q0)α-1, where q is the specific humidity simulated by LMDZ-iso and α is the isotopic fractionation as a function of the near-surface air temperature Ts simulated by LMDZ-iso. Parameter dTs is estimated from Eq. (5) by calculating the difference between the near-surface air temperature simulated by LMDZ-iso and that predicted by the adiabatic lapse rate: dTs=Ts-T0-Γ×(z-z0). All the isotopic decomposition terms computed are weighted by the precipitation amount.

First, to check whether the linear decomposition is a good approximation of the total Rp change, we estimate the non-linear term N as a residual, i.e. for each pair of states, we calculate the deviation of ΔRp=Rp(ε2, dRv2, h2, dTs2, z2) - Rp(ε1, dRv1, h1, dTs1, z1) from LMDZ-simulated isotopic differences between the two experiments. N represents less than 17 % of the total Rp change for both stages of TP uplift.

Our method to estimate the terms in Eq. (7) is equivalent to first order approximation of partial derivatives, i.e. we neglect the sensitivity of the partial derivatives to the state at which they are calculated. We tested this sensitivity by using Eqs. (9) to (13) changing z′ to z1 or z2, dTs′ to dTs2 or dTs1 and so on. For example, in Eq. (13), replacing h′ with h1 changes the resulting Rp,Δz by 0.03 ‰, replacing h′ with h2 has an impact of 0.09 ‰. In the same equation, replacing dTs′ with dTs1 and with dTs2 contributes to Rp,Δz with 0.005 and 0.039 ‰ respectively. As it was highlighted earlier, replacing ε′ and dTs′ with ε1 or ε2 and dRv1 or dRv2 respectively has no impact to the resulting Rp,Δz Thus, our method shows low sensitivity to the state.

Second, to check the influence of initial conditions Rv0, T0 and q0 on the decomposition, we estimate the sensitivity of the different contributions to changes in Rv0, T0, and q0, of 1 %, 1 K and 10 % respectively (Table 2). Rv0 is the parameter that influences most of the decomposition terms, with a maximal sensitivity of 0.9 ‰ obtained for ΔRp,Δz for a change of 1 ‰ in Rv0. Sensitivity to temperature and humidity is lower, ranging from 0 to 0.6 ‰. Overall, all the decomposition terms show a sensitivity < 1 ‰ with most (82 %) of them < 0.5 ‰, making our decomposition method robust.

LMDZ has been used for numerous present-day climate and palaeoclimate studies (Kageyama et al., 2005; Ladant et al., 2014; Sepulchre et al., 2006), including studies of monsoon region (e.g. Lee et al., 2012; Licht et al., 2014). Yao et al. (2013) showed that LMDZ-iso has the best representation of the altitudinal effect compared to similar GCM and RCM models. These authors have also provided a detailed description of rainfall patterns over the Tibetan Plateau, and showed LMDZ-iso ability to simulate atmospheric dynamics and reproduce rainfall and δ18O patterns consistent with data over this region. For the purpose of our experiments validation, we compare MOD experiment outputs with rainfall data from the Climate Research Unit (CRU) (New et al., 2002) (Fig. 3a, b, c). When compared to CRU dataset, MOD annual rainfalls depict an overestimation over the high topography of the Himalayas and the southern edge of the Plateau, with a rainy season that starts too early and ends too late in the year. Over central Tibet (30–35∘ N), the seasonal cycle is well captured by LMDz-iso, although monthly rainfall is always slightly overestimated (+0.5 mm day-1). CRU data show that the northern TP (35–40∘N) is drier with no marked rainfall season and a mean rainfall rate of 0.5 mm day-1. In MOD experiment, this rate is overestimated (1.5 mm day-1 on annual average). Despite these model data mismatches, the ability of LMDZ-iso to represent the seasonal cycle in the south and the rainfall latitudinal gradient over the TP allows its use for the purpose of this study.

Our MOD simulation is pre-industrial, consequently a comparison with modern data is expected to provide differences driven by the pre-industrial boundary conditions. Still comparing LMDZ-iso outputs with mean annual temperatures from CRU dataset (New et al., 2002) (Fig. 3d, e, f) and relative humidity from NCEP-DOE Reanalysis (Kanamitsu et al., 2002) (Fig. S1 in Supplement) shows that LMDZ-iso model captures these variables reasonably well.

Theoretically, the Tibetan Plateau has both mechanical and thermal effects on atmospheric dynamics that induce increased monsoon activity to the south and drive arid climate to the north (Broccoli and Manabe, 1992; Sato and Kimura, 2005). Thus, modifying TP height is expected to alter these large-scale atmospheric dynamics and associated climate variables (namely temperature, precipitation, relative humidity (hereafter RH), cloud cover), and in turn to affect the isotopic signature of rainfall.

In LOW experiment, strong summer heating leads to the onset of a “Thermal Low” at the latitude of maximal insolation (ca. 32∘ N), similar to the present-day structure existing over the Sahara (Fig. S2). This structure is superimposed by large-scale subsidence linked to the descending branch of the Hadley cell, and both factors act to drive widespread aridity over the TP area between ca. 30 and 40∘ N, associated with very low (< 40 %) RH values (Fig. S2). Subsidence also prevents the development of South Asian monsoon over the north Indian plane and favours aridity over this region. In winter, large-scale subsidence induces high surface pressures and creates an anticyclonic cell that prevents convection and humidity advection, resulting in low RH and annual rainfall amount ranging from 50 to 500 mm over the TP area (Fig. 4f).

Uplifting TP from 250 m above sea level (ASL) to half of its present-day altitude (INT case) initiates convection in the first tropospheric layers, restricting large-scale subsidence to the upper levels (Fig. 4c, e). In turn, South Asian monsoon is strengthened and associated northward moisture transport and precipitation increase south of TP (Figs. 5, 6). As a consequence the hydrological cycle over TP is more active, with higher evaporation rates (Fig. 7d). Together with colder temperatures linked to higher altitude (adiabatic effect) (Fig. 7b), the stronger hydrological cycle drives an increase in RH (Fig. 7a) and cloud cover (Fig. S3). Another consequence of increased altitude is higher snowfall rates in winter and associated rise of surface albedo (Fig. S4). When added to the increased cloud cover effect, this last process contributes to an extra cooling of air masses over the Plateau. To the north of TP, the initial stage of uplift results in increased aridity (i.e. lower RH and rainfall) over the Tarim Basin region (Fig. 6). This pattern can be explained both by a barrier effect of southern topography and by stationary waves strengthening, which results in subsidence to the north of TP. This latter mechanism is consistent with pioneer studies which showed that mountain-related activation of stationary waves prevented cyclonic activity over Central Asia and induced aridity over this region (Broccoli and Manabe, 1992).

The impact of the terminal stage of TP uplift also drives an increase in RH over the Plateau, especially during summer time, when a very active continental recycling (Fig. S6) makes RH rise from 40 (INT) to 70 % (MOD). Precipitation amount also increases significantly (Fig. 6), driven both by increased evaporation and water recycling during summer, and intense snowfall during winter. The latter contributes to the increase in the surface albedo and associated surface cooling during winter. Conversely, the uplift to a modern-like Plateau reduces RH (down to 30 %) north of the Plateau, and allows the onset of large arid areas. We infer that this aridification is linked to a mechanical blocking of moisture transport, both by Tian Shan topography for the winter westerlies, and the eastern flanks of TP for summer fluxes; despite changes in stationary waves structure and sensible heat (not shown), no marked shift in subsidence between INT and MOD experiments is simulated. This result is consistent with recent studies (Miao et al., 2012; Sun et al., 2009) that have suggested the potential contribution of Pamir and Tian Shan rainshadow effect to aridification in Qaidam Bassin and the creation of Taklamakan Desert.

RH alters rainfall isotopic signature through two steps, during and after condensation. As mentioned earlier, the first effect of RH, as shown in Eq. (4) and expressed as ΔRp,Δh, occurs during condensation through Rayleigh distillation and induces that Rp increases with increasing RH. Our model shows that RH increases over TP with the initial stage of uplift, driving precipitation δ18O towards less negative values. This mechanism is more efficient for the terminal stage of uplift, when RH is increased in summer as a response of a more active water cycle. South of TP, RH direct effect on δ18O is noticeable, as efficient moisture transport is activated with the uplift-driven strengthening of monsoon circulation (Fig. 4). Interestingly, this mechanism is not active for the second stage of the uplift, during which rainfall increases through more effective convection, not through higher advection of moisture. As a consequence, negligible RH and Rp changes are simulated south of the Plateau when it reaches its full height. This suggests that an altitudinal threshold might trigger South Asian monsoon strengthening, and ultimately precipitation δ18O signature, a hypothesis that should be explored in further studies. Conversely, the negative values of ΔRp,Δh over and northeast of the Tarim basin are related to a decrease in RH during both stages. Our analysis suggests that the first uplift stage is sufficient to create both barrier effects to moisture fluxes and large-scale subsidence that ultimately drive aridity over the region.

The second effect of RH on δ18O concerns very dry areas (ca. < 40 %), where raindrop re-evaporation can occur after initial condensation, leading to an isotopic enrichment of precipitation compared to water vapour (Lee and Fung, 2008) (Fig. S2). Such an effect is implicitly included in the post-condensational term of our decomposition that shows opposite sign when compared to ΔRp,Δh. Over the Plateau, this mechanism is effective only for the first uplift stage, where TP area transits from very low precipitation amounts and very low RH values to wetter conditions (Fig. S7).

Over TP, the opposed effects of RH almost compensate each other for the early stage of the uplift (Fig. 10d, e), but this is not the case for the final stage, since RH post-condensational effect is similar between INT and MOD experiments. Since absolute values of the impact of RH through condensation and post-condensational processes can reach 5 ‰, it is crucial to consider RH variation when inferring palaeoaltitudes from carbonates δ18O.

Our results also show a substantial increase in precipitation amount over northern India, the Himalayas and TP with the growth of topography for both uplift stages (Fig. 14). The inverse relation between the enrichment in heavy isotopes in precipitation and precipitation amount, named the “amount effect” (Dansgaard, 1964) is largely known for oceanic tropical conditions (Risi et al., 2008; Rozanski et al., 1993) and for Asia monsoonal areas (Lee et al., 2012; Yang et al., 2011). Over South Tibet recent studies have shown the role of deep convection in isotopic depletion (He et al., 2015). For the two stages of uplift, the residual component of the isotopic signal depicts negative values over southern TP, where annual rainfall amount is increased. Thus we infer that this anomaly can be driven, at least partly, by the amount effect that increases with growing topography.

Various climate studies have suggested that the appearance of the monsoonal system in East Asia and the onset of central Asian desertification were related to Cenozoic Himalayan–Tibetan uplift and withdrawal of the Paratethys Sea (An et al., 2001; Clift et al., 2008; Guo et al., 2002, 2008; Kutzbach et al., 1989, 1993; Ramstein et al., 1997; Raymo and Ruddiman, 1992; Ruddiman and Kutzbach, 1989; Sun and Wang, 2005; Zhang et al., 2007), although the exact timing of the monsoon onset and its intensification remains debated (Licht et al., 2014; Molnar et al., 2010). Although our experimental setup, which does not include Cenozoic palaeogeography, was not designed to assess the question of monsoon driving mechanisms nor its timing, our results suggest that uplifting the Plateau from 250 m ASL to half of its present height is enough to enhance moisture transport towards northern India and strengthen seasonal rainfall. Nevertheless, massive increase of rainfall over TP between INT and MOD experiments indicates that the second phase of uplift might be crucial to activate an efficient, modern-day-like, hydrological cycle over the Plateau. The decrease in simulated precipitation north of the Plateau also suggests that terminal phase of TP uplift triggered modern-day arid areas.

Although precipitation amount change explains well the residual isotopic anomaly (Figs. 10f, 11f), additional processes could interplay. Continental recycling can overprint original moisture signature and shifts the isotopic ratios to higher values due to recharging of moisture by heavy isotopes from soil evaporation (Lee et al., 2012; Risi et al., 2013). In our simulation, we detect an increasing role of continental recycling in the hydrological budget of the TP (Fig. S6), especially in its central part, that likely shifts the δ18O to more positive values and partially compensates for the depletion linked to the “amount effect” over the central plateau. Another process frequently invoked to explain the evolution of precipitation δ18O patterns over TP is changes in moisture sources (Bershaw et al., 2012; Dettman et al., 2003; Quade et al., 2007; Tian et al., 2007). Except for the continentally recycled moisture, southern Himalayas precipitation moisture originates mainly from the Indian, the Atlantic and the Pacific oceans (Fig. S6). Proximate oceanic basins are known to be sources of moisture with a more positive signature than remote ones (Chen et al., 2012; Gat, 1996). Supplemental analyses with water-tagging feature of LMDZ-iso show that contribution of continental recycling to rainfall over TP increases with the uplift, at the expanse of Pacific and Indian sources (Fig. S6). Although we have no mean to decipher between sources and amount effect in the residual anomaly, it seems that the change of sources is not sufficient to yield a strong offset of δ18O values.

Quantitative palaeoelevation reconstructions using modern altitude-δ18O relationship will succeed only if ΔRp corresponds mainly to the direct topography effect. Modern palaeoaltimetry studies cover almost all regions of the Plateau for time periods ranging from Palaeocene to Pleistocene-Quaternary (see data compilation in Caves et al., 2015). Most of these studies consider changes in δ18O as a direct effect of the topography uplift. Palaeoelevation studies locations (see Caves et al., 2015 for a synthesis) plotted over the anomaly maps (Figs. 12a, 13a) show for what geographical regions restored elevations should be used with an additional caution. Numerous palaeoelevation data points were located either over the northern part of the TP (from 34 to 38∘ N and from 88 to 100∘ E) or over the southern region (from 27 to 33∘ N and from 75 to 95∘ E).

Our model results show that when TP altitude is increased from half to full, considering topography as an exclusive controlling factor of precipitation δ18O over the southern (northern) region likely yields overestimations (underestimations) of surface uplift, since the topography effect is offset by RH and amount effects. Projecting our modelling results to each locality where palaeoelevation studies have been published (Table 4) reveals that topography change explains simulated total isotopic change reasonably well for only few locations (Linzhou Basin, Lunpola Basin, Kailas Basin, Huaitoutala). Indeed topography appears to be the main controlling factor for only 40 % of the sites, while 30 % is dominated by RH effects, 20 % by residual effects and 5 % by post-condensational and non-adiabatic temperature changes, respectively. Nevertheless such figures have to be taken carefully, since we ran idealized experiments testing only the impact of uplift, neglecting other factors like horizontal palaeogeography or pCO2 variations, the latter being known to influence δ18O as well (Jeffery et al., 2012; Poulsen and Jeffery, 2011).

For the initial uplift stage apparent consistency occurs between the topography impact and the total isotopic composition, in relation to counteracting effects RH and post-condensational processes. For the southern region RH impact appears to be the main controlling factor for the isotopic composition of precipitation, surpassing the direct topography impact. Nevertheless, these processes have a different contribution for initial and terminal stages of uplift. Precipitation changes lead to overestimating altitude changes for both stages, but for the terminal stage its contribution is bigger. This effect dominates in the southern part, and more generally where the isotopic composition of precipitation strongly depends on convective activity. RH changes dominate over the western part of TP and Northern India for initial uplift stage and over the northern TP for the terminal. Differences between both stages could be partly explained by non-linearities in qs – temperature relationships, as well as in Rayleigh distillation processes (Fig. S8). Determining whether other processes contribute to this difference would be of interest, but it is out of the scope of the present study.