Introduction

Silicate weathering influences Earth’s climate and habitability by transferring carbon dioxide (CO2) from the atmosphere to the lithosphere1. However, understanding how this process has varied in the past and what has controlled it (e.g., through changes in climatic or tectonic forcing) remains a major challenge2,3,4. Lithium isotopes (reported as δ7Li) have been widely used as a tracer for silicate weathering5,6,7: Li is hosted mainly in silicate minerals8, whose Li contents are orders of magnitude higher than those of carbonate8,9 and vegetation10,11,12,13, and its isotopes (7Li and 6Li) fractionate during secondary mineral formation associated with weathering9,14. The δ7Li value of river water (dissolved Li) is controlled by the congruency of silicate weathering, i.e., the ratio of primary mineral dissolution relative to secondary mineral formation7,15,16. Congruent release of Li results in riverine δ7Li values that match primary rock compositions with no isotopic fractionation16,17,18,19. Incongruent release of Li due to uptake of Li into secondary minerals will increase dissolved δ7Li values by preferentially incorporating light 6Li into clays6,16,19,20,21,22,23. Thus, riverine δ7Li does not necessarily directly track silicate weathering rate or intensity (defined as the ratio of silicate weathering to total denudation rate19,24).

Past records of the seawater δ7Li are characterized by large excursions (by up to 15‰) over 106 year timescales25 as well as gradual changes over 107 years25, such as occurred during the Cenozoic when δ7Li increased by ~9‰6. Different interpretations exist for the Cenozoic δ7Li increase6,24,26,27, but most studies have interpreted it as reflecting a shift from more congruent continental weathering ~60 million years ago (typical of flat lowland settings) to more incongruent weathering in the present-day (typical of mountains), as a result of increased global denudation. This Cenozoic seawater δ7Li rise, as with 87Sr/86Sr and 187Os/188Os changes, was thought to be driven by mid to late Cenozoic tectonic uplift of major mountain ranges28,29,30,31, which would have led to a change of the global weathering rate or the weathering feedback strength6,24,26,27,32.

A major issue with this interpretative framework is that present-day riverine δ7Li data do not follow the expected pattern: some rivers draining flat lowlands have high δ7Li9,14, whereas others draining active mountains (e.g., Himalayas, Andes and New Zealand) have generally lower δ7Li values relative to downstream areas7,16,19,33. Besides, nearly the entire range of riverine δ7Li values can be found in fluids within a single weathering profile34. These observations, contrary to expectations from shift in weathering regimes, illustrate the difficulty of interpreting modern-day riverine δ7Li, and complicate our understanding of past δ7Li records7,9,15,16,19,33,35,36,37,38. Weathering in large river floodplains has been proposed as a source of high δ7Li7,16,19,35, but the existence and extent of this process is debated36. Instead, reactive-transport modelling approaches suggest that fluid residence time may impose a major control on dissolved δ7Li and weathering congruency26,34,38,39,40,41, but this idea needs to be confirmed by observations from modern rivers globally.

Here we explore the role of hydrology as a control on present-day global river δ7Li variations. We first characterize the seasonal variability in δ7Li values by reporting new time-series data from the Tibetan and the Pamir Plateaus. We then revisit published seasonal data of individual rivers from the Arctic to the equator (Supplementary Fig. 1) and compare the spatial average δ7Li of global 64 rivers of various sizes and climatic conditions, but from similar geomorphic settings, to assess the influence of hydrology. Finally, we propose a unifying interpretation for past δ7Li changes on geological timescales.

Results and discussion

Seasonality of global riverine δ7Li

The precipitation regime affects the residence time of waters in river catchments, across storm events and over seasonal timescales42,43,44. Thus, timeseries of riverine δ7Li can provide an insight into Li isotopic fractionation as a function of varying duration and/or degree of water-rock interaction35. Our weekly-sampled time-series data from the northeastern (NE) Tibetan Plateau (Supplementary Fig. 2) provide a case study for in-depth understanding of the seasonal behaviour of riverine δ7Li (Fig. 1a). In the Buha River (BH), riverine δ7Li values are the highest (up to 22.4‰) during winter dry conditions, when river water discharge (Qw) is fed by baseflow (minimum Qw of 1.5 m3/s), corresponding to slow flow and thus a long water-rock contact time. In contrast, at the onset of the summer monsoon, with a sharp increase in Qw (up to 221 m3/s), BH riverine δ7Li decreases to its lowest value of ~12.0‰. Fast river flow during the summer monsoon corresponds to relatively short water-rock interaction times, which is consistent with an increase in the relative contribution of carbonate versus silicate weathering giving rise to a decrease in 87Sr/86Sr in this carbonate-dominated catchment45. After the summer monsoon, riverine δ7Li returns to high values concurrent with decreasing Qw. We observe similar seasonal δ7Li variations in the adjacent but silicate-dominated Shaliu River (SL), which shows a seasonal riverine δ7Li variation (δ7Lidry − δ7Liwet) of ~8.5‰ (Fig. 1b). Each river shows a negative relationship between δ7Li and Qw (Fig. 1c), and when the weekly data are considered altogether, there is a significant (r2 = 0.55; P < 0.0001) negative relationship (Fig. 1c), despite their contrasting lithology (Fig. 1d). The negative correlation is qualitatively supported by our new data from glacial streams in the NE Pamirs (Supplementary Fig. 4): all 10 sampling sites from glacier margins to downstream exhibit systematically higher δ7Li values during the dry season (spring, low ice melt) compared to the wet season (summer, high ice melt) (Supplementary Fig. 5a).

Fig. 1: High-resolution river water δ7Li, 87Sr/86Sr, and hydrometeorological data from the NE Tibetan Plateau.
figure 1

Weekly variations of δ7Li and 87Sr/86Sr in the carbonate-dominated BH (a) and silicate-dominated SL (b) catchments (Supplementary Fig. 2) along with daily Qw and precipitation, showing inverse trends between δ7Li and Qw in each river. When plotting up weekly data from the two rivers together (c), there is still an overall negative relationship, highlighting a strong hydrology control on riverine δ7Li. (d) 87Sr/86Sr versus Qw, showing large differences between the two rivers, reflecting their distinct lithology (Supplementary Fig. 3). The dashed lines in a and b represent ice-melting times. Errors for δ7Li are <0.9‰. The shaded regions in c show 95% confidence intervals. Symbols with black borders in c and d represent wet seasons, and others are dry seasons.

Full size image

We have extended our investigation into other rivers, globally, including both the spatially seasonal dataset (Supplementary Fig. 5) and the time-series δ7Li dataset from the very small Strengbach to the large tropical Congo (Supplementary Fig. 6). A common pattern is observed: each river exhibits higher δ7Li in dry seasons compared to wet seasons (Fig. 2), with a caveat that a few data show small seasonal δ7Li variations less than the analytical error (<±1‰, Supplementary Fig. 7) either due to a low degree of precipitation seasonality (e.g., the Columbia River draining the east of the Cascades38) or very similar Qw of sampling seasons (Supplementary Fig. 8). Overall, the difference in riverine δ7Li between dry and wet seasons ranges from −0.3‰ to +16.4‰ (Fig. 2). The Earth’s two largest river systems, the Congo and the Amazon Rivers, which together contribute ~20% of the freshwater supply to the oceans, show 7.5‰ differences in δ7Li values between dry and wet seasons in the Congo (Fig. 2), and >10‰ at the Amazon mouth46, respectively. The seasonal data discussed here (6 sets of time-series and 72 seasonal datasets) includes the mainstems and tributaries of several world’s largest rivers, i.e., the Amazon, Congo, Ganges, Brahmaputra, Yenisei, Yellow Rivers. The calculated total annual Li flux of these large rivers is 1.44 × 109 mol/yr (Supplementary Data 1), accounting for ~52% of the Li flux for major world rivers estimated by Huh et al.14. As with any river sample set, spatial gaps in continental coverage are inevitable. However, the rivers investigated here are globally representative: they form a dataset that covers a large range of vastly contrasting climates and vegetation (from high-latitudes to the equator), basin sizes (from small catchments to Earth’s largest rivers, Supplementary Fig. 9), and geomorphic settings (Arctic permafrost, Rocky and Andean mountains, Loess Plateau, Pamir-Tibetan Plateaus, Himalayan floodplains, and tropical rainforests). Such consistent seasonal δ7Li pattern differs completely from the temporal variability of Sr isotopes in global rivers, i.e., all seasonal and time-series δ7Li across latitudes investigated here show lower values in wet relative to dry seasons. By contrast, time-series 87Sr/86Sr values display both increasing and decreasing trends from dry to wet seasons (Supplementary Fig. 10 and Data 3). The difference between Li and Sr isotopes suggests distinct control mechanisms, with the 87Sr/86Sr probably reflecting lithological variability as shown by Sr data from the BH and SL catchments (Fig. 1d; Supplementary Fig. 3).

Fig. 2: Seasonal differences in river water δ7Li across latitudes.
figure 2

Mainstreams and tributaries in river basins from ~1 to >106 km2 in drainage area show systematically higher δ7Li values in dry seasons (δ7Lidry, blue crosses) than those in wet seasons (δ7Liwet, grey crosses). For time-series data (weekly or monthly samples at one sampling site) in the BH and SL (Fig. 1, this study), Yenisei48, Strengbach10, Yellow47, and Congo37 (Supplementary Fig. 6), the highest values in dry seasons and lowest in wet seasons are presented to explore their maximum differences (Red bars, defined as: δ7Lidry − δ7Liwet). For other seasonal data, the Columbia38, Gaizi (this study), and Ganges-Brahmaputra-Meghna River (G-B-M R.) systems8,39 show spatial spot samples from upstream to downstream in each basin, with samples collected both in dry and wet seasons for each sample site (Supplementary Fig. 5). Red bars represent their seasonal differences by sample sites. In the Gaizi River, seasonal data are distributed from glacier margins to downstream. In the G-B-M R. system, seasonal data are presented from the small headwater of the Ganges8 to large main tributaries (Ganges, Brahmaputra, Meghna) and further to downstream mainstem (G-B-M)39. “n” is the number of the data collected for each basin. See Supplementary Note 1 and Data 1 for additional details and data sources. Errors for δ7Li are similar to the symbol size.

Full size image

Previous interpretations suggest that seasonal δ7Li variation in individual rivers could reflect tributary mixing37, seasonal temperature shifts47, or influence of fluid residence times38,39,48. If mixing of different water bodies with distinct δ7Li is the main control, δ7Li shifts from dry to wet seasons should either increase or decrease, depending on isotopic compositions and Qw of tributaries. The consistency of δ7Li across rivers of vastly different catchment areas and network structures indicates that tributary mixing cannot explain all observations (Supplementary Note 1). A significant temperature control can also be excluded for some river basins characterized by the relatively small seasonal air temperature variation, in particular in the Congo (Supplementary Fig. 11) and Amazon46 rainforests, and no correlation between temperature and δ7Li in the Yangtze River headwaters33. In addition, groundwater contribution cannot explain the consistently elevated δ7Li of river waters during dry seasons (Figs. 1 and 2). Because groundwaters have both low and high δ7Li values49, ranging from +6‰ to +29‰, and thus can either raise or lower river water δ7Li. Moreover, several rivers show higher δ7Li than that of groundwaters13,34,35,38,39. Direct input of groundwaters with lower δ7Li (e.g., in the Ganges-Brahmaputra basins39) would decrease river δ7Li, which is at odds with the consistent increase in δ7Li in dry seasons across latitudes (Figs. 1 and 2). By the same token, the observed consistent decreases in river δ7Li with increasing Qw during wet seasons argue against a first order control of groundwaters.

While human activities may significantly increase river Li level and decrease δ7Li in the Han River50, our assessment suggests that anthropogenic activities do not seem to have a widespread impact on large river basins with high population densities, although anthropogenic Li influences deserve further attention (see details in Supplementary Note 3). A recent study51 in relatively dry Loess Plateau proposes that evaporation can increase river water δ7Li via enhanced secondary mineral precipitation in soil waters. However, within the BH and SL catchments, stronger evaporation occurs during wet seasons (Supplementary Fig. 12), but river waters bear lower δ7Li values, with a negative relationship between δ7Li and Qw (Fig. 1c). We also noted that in some humid regions such as Congo tropical rainforest that are characterized by roughly stable temperature and evaporation, riverine δ7Li variations are very sensitive to seasonal Qw variations (Supplementary Fig. 11b). Overall, we believe evaporation plays an insignificant role in affecting riverine δ7Li.

Here we argue that the water residence time control on Li isotopes can be generalized globally to explain the observed seasonal variability. As dry and wet season river water tend to have long and short water residence times, respectively42,48, we suggest that higher δ7Li in dry seasons can be attributed to a larger amount of secondary mineral formation from fluids with longer water-rock interactions (which could promote mineral saturation and thus formation; Supplementary Fig. 13), leading to a higher proportion of 6Li incorporated in clays and higher δ7Li values in river waters. The role of residence time is strongly supported by laboratory experiments where increases in water-rock interaction time resulted in continuous and large (12‰) increases in δ7Li in solutions in 12 days, and more than 16‰ in a month52. Furthermore, substantial increases (~8‰) in solution δ7Li driven by longer residence time is confirmed by dissolution experiments with loess samples over 10 days (Supplementary Fig. 14). These experiments indicate that fluid residence time can effectively cause large solution δ7Li changes over time scales that are comparable to seasonal variations observed in natural river systems. Additionally, cave drip water data also support an important role of fluid residence time in δ7Li variations on monthly to seasonal timescales40.

Spatial river δ7Li and annual runoff

The hydrological control on riverine δ7Li is not only observed in seasonal variations, but also when comparing different river systems. We compiled all published riverine δ7Li data from specific geomorphic settings (i.e., rivers draining lowlands or mountains) and for which runoff data was available, to isolate the influence of hydrology (see Methods). Our compiled δ7Li from medium to large rivers (103 km2 to >106 km2 in size) draining only flat lowlands (i.e. having similar geomorphic setting) show different δ7Li values: dry, middle-to-high latitude rivers have higher values than wet, tropical rivers (Fig. 3a). The negative relationship between the seasonally averaged δ7Li and annual runoff for lowland rivers (Fig. 3b) can be interpreted in the same way as seasonal hydrological shifts in each river: cold, drier conditions with lower precipitation lead to a longer fluid residence time, improved or greater mineral saturation, more secondary mineral formation, and thereby higher riverine δ7Li values.

Fig. 3: Riverine δ7Li and Li yield from various geological settings.
figure 3

(a) δ7Li versus Li/Na. Tropical lowlands (blue crosses) and mountain areas (grey squares) show similar δ7Li values, but have lower δ7Li than middle-to-high latitude (MHL) lowlands (yellow dots). (b) δ7Li versus runoff, showing an overall negative correlation between MHL and tropical lowlands. Red symbols are averages of δ7Li and runoff in the MHL (red dot), tropical lowlands (red cross) and mountain areas (red square). Red errors are standard deviations. Two dashed ellipses cover the majority of the MHL and tropical lowlands, respectively. (c) Dissolved Li yield versus runoff, showing higher runoff with higher Li yield. There are fewer data points of MHL lowlands in b and c because no runoff data is available for some rivers. The MHL lowland dataset includes 27 rivers draining the Greenland Shield, Canadian Shield, Siberian Shield, and Baikal Rivers. The tropical lowland includes 12 rivers draining the Amazon Shield, Orinoco Shield, and the Congo River. The active mountain dataset includes 25 rivers draining the Andes, New Zealand Alps, Himalaya, Rocky and the Mackenzie Mountains, and upstream of the Yangtze, Mekong, and Salween Rivers (see methods).

Full size image

In comparison, rivers draining only mountain ranges, characterized by high rates of weathering and erosion (Supplementary Fig. 15), exhibit low δ7Li, and importantly, data from these settings plot on a similar trend defined by lowland rivers with the exception of some New Zealand rivers (Fig. 3b). One plausible explanation for the similar δ7Li between mountain rivers and tropical lowland rivers, despite having very different topography and erosion rates (Supplementary Fig. 15), could be their similarly higher runoff (Fig. 3b) and therefore shorter water residence times than dry, middle-to-high latitude rivers. We note that the river basins compiled here differ largely in other aspects such as sediment concentration, mineralogy, weathering rate/intensity, vegetation, and geomorphic settings, which all have been proposed to affect Li isotopic fractionation7,9,12,14,16,19,26,33,36,38, yet there is still a consistent hydrological control, suggesting that a common mechanism governs temporal and spatial δ7Li on the continents. The relatively small deviation of δ7Li within each geomorphic setting imply that superimposed upon the major and common hydrology imprint, other factors (e.g., topography, vegetation, soil thickness) may, to some extent, also contribute to the riverine δ7Li variability (Supplementary Fig. 16).

δ7Li evolution over geological timescales

In addition to the modern river dataset (Figs. 2 and 3), there is evidence for an important hydrological control on δ7Li from geological archives across a range of timescales. First, speleothems from two Israeli caves53 record high δ7Li values (~23‰) during drier, glacial periods, and low values (~10‰) during wetter inter-glacials (Fig. 4c; Supplementary Fig. 17). Second, over million-year timescales, several climatic events are characterized by changes of seawater δ7Li during the Cenozoic. The lowest seawater δ7Li value (~22‰) is recorded during the Paleocene-Eocene Thermal Maximum (PETM)6 (Fig. 4d), with rapid ~3‰ negative excursion of δ7Li over only ~100 kyr54, when precipitation and continental runoff was dramatically strengthened in a much warmer world than today54,55,56,57,58. Similarly, the Early Eocene Climatic Optimum (EECO) had much lower seawater δ7Li (~23‰) than the present-day (Fig. 4d). Proxy records suggest the period was broadly coincident with a shift to wetter climate56,59, characterized by enhanced erosion and weathering compared to the non-glacial Quaternary60. The Mid Miocene Climatic Optimum (MMCO) also shows a negative Li isotopic excursion6. Evidence from plant leaf wax δD and detrital sediment records indicate an intensified hydrological cycle in the Antarctic and the NE Tibetan Plateau at this time61,62. Third, extending to older times beyond the Cenozoic, the Cretaceous Ocean Anoxic Events OAE1a and OAE2, characterized by rapid increase in pCO2 and global warming, show abrupt declines of marine carbonate δ7Li (Fig. 4e, f), consistent with accelerated hydrological cycles5,63,64,65,66.

Fig. 4: Temporal evolution of δ7Li on various timescales ranging from days to months, millennial, and million years.
figure 4

(a) Storm events in tropical Guadeloupe showing a decrease of stream δ7Li from 9.3‰ to 7.8‰ within 1–2 days92, similar to the onset of monsoon at Tibetan rivers (Fig. 1). (b) Seasonal (time-series) variations of rivers from the Arctic to the equator, showing systematically elevated δ7Li in dry seasons. (c) Speleothems δ7Li from two Israeli caves during the last glacial cycle53. (d) Seawater δ7Li evolution during the Cenozoic6. Pink bars mark global climate events, including Paleocene-Eocene Thermal Maximum (PETM, ~56 Ma), Early Eocene Climatic Optimum (EECO, 50–52 Ma) and Mid Miocene Climatic Optimum (MMCO, 14–17 Ma). (e) and (f) Marine carbonate δ7Li of two major Ocean Anoxic Events (OAE1a, ~120 Ma)63 and (OAE2, ~93.5Ma)5. (g) Marine carbonate δ7Li during the end-Ordovician Hirnantian glaciation (~445Ma)68. The events of OAE2, OAE1a and Hirnantian glaciation lasted for ~440ka5, ~1.1Ma63 and ~1–2Ma68, respectively. Blue bars for dry events in c and g exhibit positive δ7Li excursions. Pink bars for wet events in d, e and f show lower δ7Li values (EECO and PETM), and negative δ7Li excursions (MMCO, OAE2 and 1a). The red bars in c–g show the excursion amplitudes of δ7Li (defined as δ7Lidry - δ7Liwet) during the events, with blue and red short arrows showing positive and negative excursion, respectively. See Fig. 2 for the legends in a and b. Different symbols in e and f represent marine carbonate sections at different locations, and in g represents bulk carbonates and brachiopods at same location. Errors for δ7Li are similar to the symbol size.

Full size image

The Hirnantian glaciation (~445 Ma) recorded a global temperature drop of 8–10 °C, culminating in an ice-sheet over Gondwana and subsequent global sea-level fall67,68. Marine carbonate δ7Li values show a positive excursion during this event (Fig. 4g). Sedimentary evidence indicate a climate shift from warm, humid to overall cold, arid67,69, consistent with the positive δ7Li excursion suggesting a 4-fold reduction in global weathering flux68.

To understand the past variability in the δ7Li of seawater, it is also important to consider any related changes in the dissolved Li flux to the ocean (Fig. 3 and Supplementary Fig. 18), as this can impact the residence time of Li in the ocean (Supplementary Note 2). The above-mentioned geological events support that the modern/Neogene/OAEs Li ocean mass balance were probably fundamentally different and residence times were much shorter in the past5,63,65. This was implied by modeling the OAE2 event (i.e., an increase in river Li flux results in a decrease in seawater Li residence time and δ7Li value5), and is also indicated at the timescales shown in Fig. 4. Together, all the events described herein display a response of δ7Li to climate change that is consistent with present-day riverine observations, i.e., δ7Li values decrease when climate becomes wetter which ensues shorter average fluid residence time in the continental weathering zone, and vice versa.

The role of hydrology in Cenozoic seawater δ7Li evolution

We have shown that hydrology exerts a primary control on riverine δ7Li values over seasonal and annual timescales, across latitudes and basin sizes (Figs. 2 and 3). These patterns can be explained by changes in the mean fluid residence time in a river basin, which influences the degree of Li isotope fractionation between primary minerals and the fluid phase via secondary mineral formation. The results from modern rivers are broadly coherent with Li isotope ratio shifts across pronounced changes in climate over tens of thousands (glacial cycles) to millions of years (e.g., PETM, OAEs) (Supplementary Fig. 17; Fig. 5). The associated large and rapid δ7Li shifts (e.g., ~13‰ in OAE2 with a duration of only ~440 kyr 5) appear too rapid to be linked to tectonic processes, but instead are consistent with a common hydrological change (Fig. 4e–g). Altogether, we propose that climate-driven hydrological changes alone can produce large δ7Li shifts over various timescales investigated here.

Fig. 5: Geological δ7Li records, climate and weathering flux.
figure 5

(a) Silicate weathering flux (grey solid line)4. (b) Relative strength of weathering feedback, with dark and light blue shadings indicating low and high atmospheric CO2 scenarios4, respectively. (c) Deep ocean temperature (dark green)73. (d) Seawater δ7Li (yellow) evolution during the Cenozoic6. (e) Continental mean annual precipitation from the Pacific (green) and Atlantic (grey and light purple) sides of Eurasia75. During the Cenozoic cooling, increasing δ7Li coincides with decreasing precipitation recorded at the two sides of Eurasia (e) and are likely coupled with a stable or decreased silicate weathering rate (a). Shadings in a and b indicate 1σ uncertainty, and in e show mean annual range of precipitation.

Full size image

These findings may map onto the long-term Cenozoic shift in δ7Li values of seawater. Previous studies have proposed that increasing Cenozoic seawater δ7Li, 87Sr/86Sr and 187Os/188Os values towards the present could reflect increases in global weathering fluxes linked to mountain uplift, but this remains intensively debated2,3,4,16,32. More recently, marine beryllium (Be) isotope records and marine calcification indicators were interpreted as reflecting either increased70, or near constant2, or a decreased3 continental weathering flux during cooling of the Neogene, or even the Cenozoic4 (Fig. 5). Combined with 10Be/9Be, new interpretation of seawater δ7Li records infers an increase in the feedback strength of silicate weathering (or increase in land surface reactivity) along with a stable weathering flux driven by rock uplift32. This means that even though global climate cooled, weathering fluxes did not decline, because silicate weathering became more sensitive to climate over this period. Model results suggest this came about because of increased erosion in tectonically active mountains, making more of the terrestrial land surface locations where silicate weathering was no longer limited by supply of minerals, and instead primarily controlled by runoff and temperature4,32,71,72.

Based on the data herein (Figs. 1–4), we propose to add a complementary piece to the puzzle of Cenozoic evolution: overall drying of climate during cooling since 50 Ma has left an imprint on marine δ7Li records (Fig. 5). A number of explanations6,24,26,27 for the 9‰ positive shift in seawater δ7Li over the last ~50 Ma have converged on the need for a significant increase in the riverine δ7Li values27. This could be achieved by a less intense continental hydroclimate that resulted in higher δ7Li of continental runoff (Fig. 3 and Supplementary Fig. 18). Any associated Li flux reduction would increase the Li residence time in the ocean and thereby increase seawater δ7Li (Supplementary Note 2), similar to inferences from global climatic events, e.g., the rapid decline of δ7Li during OAE2 that suggests a 2–4 times increase in river fluxes and 25%–50% decrease in seawater Li residence times relative to the present day, coupled with wetter continental conditions5.

The proposed drying climate over the Cenozoic is consistent with a >120 m drop of sea level73 and step-wise aridification recorded in many regions74,75,76,77 (e.g., central Asia, North America, Europe, Africa, and Australia). In addition to million-year records76, modern meteorological data confirm the positive relationship between global-mean temperature and precipitation78,79,80, with a 4% decrease in global runoff per °C cooling78. Given a cooling Cenozoic, these observations suggest lower continental runoff. Critically, our proposed hydrological control on the Cenozoic seawater δ7Li evolution is strongly supported by a clear negative trend between δ7Li and continental mean annual precipitation from the Pacific and Atlantic sides of Eurasia during the last 50 Ma (Fig. 5d–e).

This proposed mechanism, the reduction of continental runoff (and increased time for water-rock interaction) controlling long-term seawater δ7Li, raises questions on how silicate mineral weathering linked to hydrology mediates global climate over the Cenozoic. There are two possibilities: (i) that this led to decreased weathering fluxes; (ii) it led to no change in weathering rate if compensated by an increase in the strength of the silicate-weathering climate feedback32 and/or by changes in CO2 release from volcanism81,82, metamorphism in continental arcs83 or sedimentary rock weathering31,84,85,86. In terms of the first scenario, based on our dataset from flat lowlands and active mountains, a lower runoff would result in a lower silicate weathering rate (Supplementary Fig. 15c). This would also be consistent with other observations71,72 that suggest a 1% decrease in global runoff is accompanied by a 0.4–0.7% decrease in solute fluxes, while being in line with new data from the Himalayan-Tibetan areas that show no increase in weathering87 and erosion88 over these timescales. Together, this would challenge the long-standing uplift/weathering hypothesis28.

Recent studies have suggested global stability of the chemical weathering flux during the Cenozoic2, which supports the second scenario. This could come about if declining atmospheric CO2 was primarily driven by decreasing solid earth degassing rates, as proposed by recent studies81,82,83, which would require stable silicate weathering rates while atmospheric CO2 declined. The release of CO2 through weathering of sedimentary rocks31,84,85,86 could also play a role in net carbon cycle balance. Finally, a change in the strength of the feedback between climate and silicate weathering rate could have occurred32,72,89. A strengthening of the weathering-climate feedback32 should be expected due to increased land surface reactivity related to mountain building and higher physical erosion28,90. Our dataset and this weathering-feedback scenario would compensate each other, i.e., the Cenozoic δ7Li increase is driven by the lengthening of the average residence time of water on the continents as the climate cooled and runoff declined. This weakened hydrology and weathering (Supplementary Fig. 15c), combined with increased feedback strength caused by late Cenozoic uplift (Fig. 5b), could have sustained weathering fluxes even as cooling proceeded91. The extent to which this increased feedback strength can offset the decreased weathering flux caused by weakened hydrology is beyond the scope of our current study and needs more quantitative constraint from future research.

In summary, there are multiple lines of evidence across a range of temporal and spatial scales that support a strong control of continental hydrology (via fluid residence time) on Li isotopes. By recognizing a hydrological control on the δ7Li of continental runoff, our findings call for a renewed focus on how changing hydrological regimes affect Earth’s weathering and the carbon cycle over tens of millions of years, and during more rapid changes in warming global climate. The link between hydrology and δ7Li, from seasonal to geological timescales, makes it as a useful tool to reconstruct past hydrological changes, for which long-term, continuous terrestrial records remain extremely limited58,60.

Methods

Hydrological data

The daily river water discharge (Qw) of the Buha (BH) and Shaliu (SL) Rivers from 2007 to 2009 were monitored at the Buha and Gangcha hydrological stations, respectively (Supplementary Fig. 2).

Sample collection

A total of 103 river water samples were collected weekly from the BH and SL Rivers at the Buha and Gangcha hydrological stations in 2007 and 2009, respectively (Supplementary Fig. 2). Twenty river water samples were collected at 10 sampling sites from glacial margins to downstream during two field campaigns in summer 2014 and spring 2016 at the Gaizi River, NE Pamir Plateau (Supplementary Fig. 4). All water samples were filtered on site through 0.2 μm Whatman® nylon filters. The samples were collected into a 60 mL polyethylene bottle pre-acidified with 6 M quartz-distilled HNO3 and acidified to pH < 2. All samples were kept chilled until analysis.

Analysis

Lithium concentrations of all river water samples were analyzed by PerkinElmer NexION 300D ICP-MS at the State Key Laboratory of Loess and Quaternary Geology (SKLLQG) with rhodium as an internal standard. The analytical precision is better than 5%. A total of 90 samples were selected for the measurements of lithium isotope ratios. Detailed pretreatment and measurement procedures were conducted following Gou et al.47. Each water sample containing 300 ng Li was dried and purified by single-step cation exchange chromatography filled up with 8 mL resin (Bio-rad® AG50W X-12, 100–200 mesh), with 0.5 M HNO3 as an eluent. Analyses were performed on a Thermo Neptune plus multi-collector inductively coupled plasma mass spectrometer (MC-ICP-MS) at the SKLLQG. A seawater reference material (NASS-6) was analysed as an unknown and repeated measurement over a one-year period yielded a δ7Li value of +31.1 ± 0.7‰ (2σ, n = 15), in agreement with the global average seawater value of +31.0 ± 0.5‰ (ref. 6). Our long-term external reproducibility is better than ±  0.9‰ (2σ) for δ7Li measurements47. Li isotopes are all reported relative to the standard L-SVEC. For 87Sr/86Sr analysis, the pretreatment and measurement procedures were conducted following Jin et al.45. Each water sample containing 100 ng Sr was dried and purified by Eichrom SrSPEC exchange column (0.5 mL bed volume each column) preconditioned with 3 M HNO3 and eluted with 4 mL UHQ (ultra high quality) deionised water. All 53 weekly SL River water samples were measured on a MC-ICP-MS in the Isotope Geochemistry lab at the Taiwan Cheng Kung University. Standard reference material NBS 987 (recommended value = 0.710245) was periodically measured to check accuracy. Replicate analyses of NBS 987 yielded an average 87Sr/86Sr ratio of 0.710255 ± 0.000022 (2σ, n = 42). The reported uncertainties were much less than the large 87Sr/86Sr ranges of weekly samples from 0.711700 to 0.712500. Raw 87Sr/86Sr ratios for all samples and standards were corrected for mass bias by normalizing to 86Sr/88Sr = 0.1194 and corrected for 87Rb and 86Kr isobaric interferences. The blank Sr (<1 ng) was less than 1% of the processed water samples. Results for Li and Sr isotope analyses were compiled in Supplementary Data 1.

Spatial comparison of lithium isotopes at lowlands and mountain rivers

We compiled a set of river data corresponding to rivers draining either only active mountain ranges or flat lowland areas (Supplementary Data 2). The goal of this compilation is to compare the largest river catchments (because they integrate large areas) having: (1) similar geomorphic settings (flat lowland shields) but different climatic conditions from high latitudes to the equator; (2) contrasting geomorphology between flat lowland and active mountain ranges. Herein, we used the most downstream sample data corresponding to the same geomorphic setting. For mountain ranges, this strictly corresponds to rivers that were sampled upstream floodplain areas since it has been suggested that additional weathering reactions may take place in floodplains7,19. Our compilation includes 25 rivers from major orogenic belts (the Andes, the Himalayas-Tibetan Plateau, the New Zealand Alps, the Rocky and Mackenzie Mountains, and upstream of the Yangtze, Mekong, and Salween Rivers) for which runoff data were available. The sediment fluxes for these rivers are higher than 100 t km−2 yr−1 (high erosion rate) and the runoff range between 212 and 9526 mm yr−1.

For lowland rivers, our compilation includes 12 rivers in the tropical lowlands across the equator (the Amazon, Congo and Orinoco Rivers) and 27 rivers in the lowlands at middle-to-high latitudes (the Mackenzie, Lena, Yenisei, and Baikal Rivers). The middle-to-high latitudes and tropical lowland rivers are chosen from rivers draining similar relatively flat topography with low erosion rates (erosion rate < 100 t km−2 yr−1) to isolate the influence of climate. For each river, we used a single or an average of all the δ7Li measurements at the same sampling and the annual runoff. For some rivers, the runoff and sediment flux are not available and therefore not shown on Fig. 3. However, the difference in the average and range of δ7Li is minor between rivers with and without runoff data.

Lithium isotopes of geological records

We further compiled extensive δ7Li data from events across the last glacial cycle to end-Ordovician Hirnantian glaciation (~445 Ma). The dataset includes δ7Li of speleothem records from two Israeli caves during the last glacial cycle (from the present to 200 ka)53, δ7Li of foraminifera samples from 8 ocean drill sites during the Cenozoic (60 Ma)6, δ7Li of 3 marine carbonate sections at the Eastbourne and South Ferriby, UK, and Raia del Pedale, southern Italy recording the Ocean Anoxic Event 2 (OAE2, ~93.5 Ma)5, δ7Li of 4 marine carbonate sections recording the Early Aptian OAE1a (~120 Ma)63, and the δ7Li of bulk carbonates and brachiopods from Anticosti Island, Canada (Pointe Laframboise) recording the Late Ordovician Hirnantian glaciation (~445 Ma)68. Carbonate-based Li isotope records spanning the past 3 billion years indicate that bulk carbonates can be used to faithfully reconstruct seawater Li isotope values in the deeper past25.