Introduction

The climatic significance of δ¹⁸Op variations in the EASM proxy archives, exemplified by iconic EASM speleothem δ¹⁸O records, particularly from central and eastern China, remains contested (see review in ref. 1). Early interpretations framed these variations in terms of precipitation seasonality2 or the cumulative degree of rainout along the moisture transport pathways3, but subsequent studies have shown that they reflect a more complex interplay of hydrological processes operating across the EASM region1. These studies emphasize upstream deep convection and rainout as primary drivers, which preferentially remove heavy isotopologues through Rayleigh distillation, leaving downstream vapor isotopically depleted4,5,6,7,8. In this context, ENSO-modulated changes in the strength and location of convection, both within the moisture source regions and along the moisture transport pathways, are seen as key mechanisms linking upstream convection and rainout to downstream δ¹⁸Op variability over East Asia9,10,11,12,13,14,15. An alternate variant of the upstream framework emphasizes an Indian Summer Monsoon (ISM)–centric interpretation, where stronger ISM rainfall enhances rainout along the southwesterly monsoon flow from subcontinental India and the Indian Ocean into China, leaving the downstream vapor more depleted and vice versa16,17,18,19.

Another line of reasoning argues that δ¹⁸Op variability reflects the large-scale changes in EASM “intensity”, generally defined as stronger southerly monsoon winds transporting more negative δ¹⁸O oceanic moisture over China and vice versa20,21,22,23,24. Additionally, seasonal shifts in the location of moisture sources25,26,27,28, or in the relative contribution of remote ¹⁸O-depleted Indian Ocean versus local ¹⁸O-enriched western Pacific and South China Sea moisture29,30, have also been proposed as important controls on EASM δ¹⁸Op. Beyond these perspectives, an alternative mechanism frames the interannual EASM δ¹⁸Op variability arising from the interaction of the upper tropospheric westerlies with the monsoon circulation31. In this view, a weakened and southward-displaced westerly jet delays the onset of the EASM, allowing relatively heavy springtime δ¹⁸Op values to persist into summer while limiting the advection of oceanic moisture due to weaker low-level southerlies32. These diverse interpretations highlight that no single framework can fully explain δ¹⁸Op variability in East Asia, nor is one likely to exist, given that multiple processes can operate across a range of temporal and spatial scales, with different mechanisms assuming greater importance at different timescales. For instance, EASM intensity may dominate at orbital timescales25,26, whereas upstream convection may exert a stronger influence on interannual variability11,12. The challenge, therefore, lies in identifying which mechanism exerts the strongest influence across the widest spatial domain at a given temporal scale of interest.

The underlying mechanisms of EASM interannual δ¹⁸Op variability remain particularly difficult to resolve. A key limitation has been the restricted spatial and temporal coverage of observational δ¹⁸Op data, which cannot resolve large-scale variability or clarify the relative importance of different mechanisms1. Isotope-enabled general circulation models (GCMs) have advanced understanding, but most models still struggle with substantial biases in convection, precipitation, and moisture transport, and even the better-performing ones often fail to capture the spatial and temporal structure of observed precipitation and δ¹⁸Op variability32,33,34,35,36. Here, we address these shortcomings by using a 74-year (1950–2023) simulation from IsoGSM3, the latest version37 of an isotope-enabled atmospheric general circulation model36 nudged with ERA5 reanalysis. IsoGSM3 and its previous variants have been extensively validated with observation and consistently shown to reproduce observed precipitation and δ¹⁸Op variability across the EASM and the wider Asian Summer Monsoon (ASM) domains with high fidelity14,32,33,34,35,36,37,38,39,40 (see “Methods”). In addition, its water-vapor tagging framework allows quantification of moisture contributions from multiple oceanic and terrestrial sources14,38,39. Together, these features provide the most comprehensive δ¹⁸Op dataset yet available for systematically investigating the relative influence of upstream processes, moisture sources, and monsoon circulation on EASM interannual δ¹⁸Op variability. Our analysis covers the EASM domain, with a particular focus on the central-eastern China (CEC) region (27°–37°N, 104°–116°E), which hosts the majority of high-resolution speleothem records and provides a critical testbed for evaluating competing interpretations of δ¹⁸Op variability1.

Results and discussion

Interannual δ18Op variability

We applied an empirical orthogonal function (EOF) analysis to the May through October (hereafter, the summer-half) amount-weighted IsoGSM3 δ¹⁸Op standardized field to identify dominant patterns of variability over China and surrounding regions (see “Methods”). We focus on this temporal window to capture both the pre-monsoon and the retreat phase of the monsoon, as δ¹⁸Op anomalies in these transitional months can substantially influence the seasonal mean. The leading EOF mode explains 21% of the total variance (Fig. 1a) and is distinctly separated from all other modes (Fig. S1a), and exhibits a same-sign pattern over China, extending into the Bay of Bengal, the Indochinese Peninsula, and across the Maritime Continent and South China Sea. Its principal component (PC1) contains significant power in the 2–6-year band (Fig. S1b) and is well correlated with the CEC-averaged δ¹⁸Op index (r = 0.71, p < 0.01) as well as the summertime Nino3.4 index ((r = 0.71, p < 0.01), indicating that interannual δ¹⁸Op variability in this region largely follows the broader EOF1 pattern (Fig. 1b). Notably, nearly all maxima in PC1 (tracking peaks in the CEC-averaged δ¹⁸Op timeseries) coincide with years of the growing phase DJF(0) of El Niño (see “Methods”).

Fig. 1: Leading mode of δ¹⁸Op variability and ENSO.
Fig. 1: Leading mode of δ¹⁸Op variability and ENSO.
Full size image

a First EOF mode of the detrended and normalized amount-weighted summer-half (MJJASO) δ¹⁸Op from IsoGSM3, explaining 21% of the total variance. The map shows correlations between the principal component and δ¹⁸Op anomalies (color bar: correlation coefficient r, dimensionless); the rectangle marks central eastern China region (27°–37°N, 104°–116°E). The thin dotted 2500 m elevation contour line (brown) approximates the boundary of the Tibetan Plateau. For clarity, EOF1 loadings with absolute values smaller than 0.5 are masked, highlighting the regions of strongest and most coherent spatial expression. b Standardized time series of the Nino 3.4 SST index for the summer-half (green), first principal component (PC1, blue), and the CEC-averaged δ¹⁸Op (black) sharing the left Y-axis. Note that the scale is reversed with PC1, SST, and δ¹⁸Op increasing down. Red vertical lines with asterisks denote the developing El Niño years (DJF 0). The El Niño years not aligned with the maxima in PC1 are shown with the red asterisks. Pairwise correlation coefficients and their significance between Niño3.4, PC1, and CEC δ¹⁸Op are shown in the panel. c First mode of the maximum covariance analysis (MCA) between summer-half δ¹⁸Op (left field) and SST (right field). The SST data is from ref. 68. This mode explains 76% of the squared covariance (SCF), with fractions of variance (FOV) of 68% and 62% for the left and right fields, respectively. The left map shows the heterogeneous regression of δ¹⁸Op anomalies (‰) onto the SST expansion coefficient, and the right map shows the homogeneous regression of SST anomalies (°C) onto the same coefficient. Only statistically significant regression fields (α = 0.1, two-tailed) are shown.

The ENSO imprint on the leading mode is further indicated by strong correlations of the boreal summer-half Niño3.4 index with both PC1 (r = 0.81, p < 0.01) and CEC-δ¹⁸Op (r = 0.67, p < 0.01) (Fig. 1b) linking enriched δ¹⁸Op and high PC1 values to warm central-eastern equatorial Pacific SSTs. Maximum covariance analysis (MCA) confirms this association, showing higher δ¹⁸Op over much of the EASM domain concurrent with canonical ENSO-related SST anomalies (Fig. 1c). Moreover, the left expansion coefficient of the MCA δ¹⁸Op field is essentially identical to PC1 (r = 0.97; Fig. S1d), underscoring that both variance- and covariance-based approaches extract the same physically robust ENSO-modulated mode of δ¹⁸Op variability. The temporal evolution of this relationship is highlighted by the leading MCA expansion coefficients from the summer-half δ18Op and SST fields, which are strongly correlated (r = 0.81; Fig. S1d). Parallel EOF analyses of three other isotope-enabled models (see “Methods”) broadly reproduce the leading EOF pattern from IsoGSM3, supporting its robustness and arguing against it being a model artifact (Fig. S2). Strikingly, the ENSO-driven spatial structure of δ¹⁸Op also dominates the leading EOF from the IsoGSM last-millennium (851–2000 CE) simulation41 (Fig. S3a). The associated PC1 retains pronounced power in the 2–8 year ENSO band (Fig. S3b) and correlates strongly with the CEC δ¹⁸Op time series (r = 0.9; Fig. S3c). Its wavelet spectrum reveals sustained ENSO-band energy across the last millennium (Fig. S3d), highlighting the ENSO’s persistent influence on interannual δ¹⁸Op variability in CEC and more broadly across East Asia, consistent with the global-scale imprint of Pacific Walker Circulation dynamics on δ¹⁸Op30.

Upstream convection and downstream δ18Op link

The link between upstream convection and downstream δ¹⁸Op over China has been documented in a number of previous studies using station-level daily and monthly observations from the Global Network of Isotopes in Precipitation (GNIP)9,10,11,12. Here, we use the IsoGSM3 simulated data to extend this perspective over a much longer period (1950–2023) and a broader spatial domain. The spatial correlation between the CEC-δ¹⁸Op with the National Oceanic and Atmospheric Administration (NOAA) outgoing longwave radiation (OLR)42 for the summer-half indicates positive correlation (i.e., signifying a negative correlation between δ¹⁸Op and convection strength) over key upstream regions, including subcontinental India, the Bay of Bengal (R1), the Indochina Peninsula, the South China Sea and adjacent western Pacific (R2), and the Maritime Continent (R3) (Fig. 2a). The annual cycles of CEC-δ¹⁸Op and OLR averaged across all upstream regions (5°S–25°N; 65°E–130°E) show a broadly similar progression (Fig. 2b), demonstrating that large-scale upstream convective activity and downstream CEC-δ¹⁸Op co-vary through the year, essentially mirroring the seasonal migration of the Intertropical Convergence Zone (ITCZ) in this longitudinal band (Fig. S4a). A consistent one-month offset is, however, evident, with the OLR decreasing abruptly in May as upstream convection intensifies, the ITCZ crosses the equator and migrates north in May (Fig. S4a) while the δ¹⁸Op over CEC declines about a month later in June (Fig. 2b). This lag indicates that the isotope signal over CEC is realized only after oceanic moisture transport into CEC is fully established with the onset of monsoon circulation in June. The offset also explains why climatological δ¹⁸Op minima over CEC occur in August–September, about a month later than the upstream OLR minima in July–August (Fig. 2b).

Fig. 2: Relationship between upstream convection and downstream δ¹⁸Op variability.
Fig. 2: Relationship between upstream convection and downstream δ¹⁸Op variability.
Full size image

a Correlation map of mean summer-half δ¹⁸Op over CEC with OLR, with three upstream subregions (R1–R3) outlined. The larger solid rectangle marks the full upstream domain used in panel b. Negative correlations are masked for clarity, and only correlations significant at the 95% confidence level (α = 0.05) are shown. b Climatological annual cycles of δ¹⁸Op over CEC and OLR, the latter averaged over the entire upstream domain (dimensions shown on the panel). Arrows highlight the May decrease in OLR and the June decrease in δ¹⁸Op, indicating a ~ 1-month lag between the onset of enhanced upstream convection and the isotopic response. Error bars denote ±1 standard deviation for both variables. c Vertical velocity at 500 hPa (ω500) anomalies composited for high minus low CEC-δ18Op index years (N = 18 for each group). d Same as (c) but excluding ENSO years (N = 8 for each group). Stippling marks grid points where the difference is significant at the 95% confidence level, based on a two-tailed Student’s t-test of the difference in means.

At the monthly scale, spatial correlation patterns reveal a seasonal migration of convection across the primary upstream regions linked to δ¹⁸Op variability over CEC (Fig. S4b). In May, correlations are broad but weak, consistent with pre-monsoon conditions before the large-scale summer monsoon circulation is established. During June and July, the strongest correlations are concentrated in the “western sector” (west of 100°E), where deep convection peaks during the ISM onset and mature phase. In August and September, correlations weaken across most upstream regions, coinciding with the late phase of the ISM/EASM. By October, correlations strengthen again and shift to the “eastern sector”—the upstream regions spanning the Maritime Continent, Indochina Peninsula, South China Sea, and western Pacific (Fig. S4b).

Sensitivity experiments with IsoGSM provide further insight into how upstream convective processes shape the annual δ¹⁸Op cycle over CEC (see “Methods”). In these experiments, atmospheric isotopic fractionation during condensation and raindrop–vapor exchange was selectively disabled over different source regions Suppressing all fractionation immediately south of China, across the corridor from the western Indian Ocean through the Bay of Bengal and Indochina to the western Pacific (7°N–25°N; 30°E–150°E; SR3 in Fig. S5a), removes the seasonal cycle evident in the control simulation (Fig. S5b). In contrast, turning off fractionation in the midlatitude continental areas west of China along the prevailing westerlies (26°N–50°N; 30°E–110°E; SR4) or within continental eastern China (26°N–50°N; 110°E–120°E; SR5) has no major effect, indicating that midlatitude or local continental processes make little contribution to the seasonal δ¹⁸Op cycle over CEC. Similarly, turning off fractionation over the Maritime Continent (10°S–10°N; 90°E–150°E; SR1) or the tropical Indian Ocean (10°S–10°N; 30°E–90°E; SR2) also leaves the cycle essentially unchanged, because in this configuration convection within SR1 and SR2 does not impart an isotopic signal, and the air masses acquire their depletion signature only as they pass through SR3 where convection and rainout result in depleted δ¹⁸O of vapor before it enters CEC. When fractionation is disabled across the entire upstream domain (10°S–25°N; 30°E–150°E; SR), the CEC seasonal cycle is again eliminated, highlighting the importance of upstream convection and rainout, both at the source and along southerly moisture pathways before air masses enter CEC.

ENSO-modulated convective mode of δ18Op variability

To assess the spatial-temporal patterns of the upstream convection-downstream δ¹⁸Op relationship and the extent to which it is modulated by ENSO, we used mid-tropospheric vertical velocity (ω) outputs from IsoGSM3. ω is preferred over NOAA’s OLR because the latter is only available from 1979 onward, which reduces the ensemble of ENSO events available for evaluating its role in influencing δ¹⁸Op. Moreover, ω provides a direct measure of large-scale vertical motion in the mid-troposphere, making it a robust diagnostic of convective circulation. A height–longitude section of the summer-half ω anomalies (10°S–10°N) composited as the difference between high-minus–low CEC-δ¹⁸Op years, defined as those above the 75th and below the 25th percentile of the CEC-δ¹⁸Op index distribution, highlights eastward displacement of the ascending limb of the Walker circulation that is typical of El Niño events (Fig. S6a; see “Methods”).

At the monthly level, vertical velocity at 500 hPa (ω500; Pa s−1 with negative values denoting ascent) was composited, as above, to examine seasonal changes. A time-longitude cross-section of ω500 anomalies, averaged over the broader upstream region (5°S to 25°N), shows prominent positive anomalies (i.e., descent) over the Maritime Continent sector, the Indochina Peninsula, and the South China Sea, which is particularly expressed from August to October, as El Niño events transition to maturity (Fig. S6b). When ENSO years are excluded (i.e., removing El Niño and La Niña years from the high- and low-δ¹⁸Op CEC-index years, respectively), the anomalies over the Maritime Continent sector weaken (Fig. S6c). These results indicate that during the developing phase of the El Niño years, as deep convection shifts eastward, 18O-enriched vapor over the Indo-Pacific Warm Pool and Maritime Continent is transported downstream, ultimately manifesting as 18O-enriched precipitation over CEC. This far-field El Niño impact on downstream δ¹⁸Op is further amplified by reduced convection and rainout along the moisture transport pathways (Fig. S6b).

Given the robustness of the ENSO-modulated convective signal in IsoGSM3, it is relevant to consider whether this relationship remains stable through time, particularly in light of suggestions of non-stationarity in the ENSO–δ¹⁸Op relationship on annual timescales across two late-20th-century climate transitions centered around 1976–1977 and 1988–198912. However, this apparent non-stationarity arises from a weaker ENSO–δ¹⁸Op relationship over the EASM domain in dry-season precipitation prior to 1977, whereas the wet-season relationship remains relatively stable across both transitions12. Our analysis, focusing on summer-half precipitation, likewise identifies a stable ENSO–δ¹⁸Op relationship within this seasonal window. The persistence of the wet-season ENSO–δ¹⁸Op relationship across the instrumental period and in the last-millennium IsoGSM simulation41 (Fig. S3) indicates a stable relationship over a wide range of timescales.

Circulation intensity mode of δ18Op variability

The preceding analyses highlight the role of ENSO as a primary modulator of δ¹⁸Op anomalies over CEC. A key question thus emerges as to what drives δ¹⁸Op variability in the absence of ENSO forcing. A number of studies have previously implicated variations in “monsoon intensity” as a dominant control on δ¹⁸Op in East Asia20,21,22,23,24. Although the term is not always defined uniformly, it generally refers to the northward advance of monsoon southerlies that transport oceanic moisture into China. To evaluate its role, we use the v-component of the JJAS 850 hPa wind averaged over 20°–45°N, 110°–125°E43, hereafter referred to as the v-wind index, as a metric for assessing interannual changes in monsoon intensity. Because monsoon circulation intensity can be influenced by ENSO, we conducted a parallel set of analyses to isolate its role in modulating δ¹⁸Op over CEC, both in the presence and absence of ENSO forcing.

The summer-half δ¹⁸Op anomalies, expressed as the difference between high and low v-wind index years without including ENSO events, show that depletion is concentrated over CEC (~ −1.04‰) and adjacent northern China (Fig. 3a). The corresponding vertically integrated moisture-flux (VIMF) anomalies exhibit enhanced meridional transport with a broad clockwise circulation pattern that draws moisture from the western Pacific and South China Sea into southern China and then northward into central and northern China (Fig. 3a). The seasonal cycle of δ¹⁸Op further shows that values are consistently depleted in high v-wind years from June through October (Fig. 3b). Furthermore, there is a modest increase in precipitation amount over CEC, indicating a weak negative relationship between δ¹⁸Op and rainfall amount (r = –0.21 Fig. 3c). When ENSO years are included, key differences emerge. The δ¹⁸Op anomalies are more depleted and spatially extensive, extending in a southwest–northeast band from southern Tibet to northeast China, with mean depletion over CEC reaching ~ −1.4‰ (Fig. 3d). Negative anomalies also emerge over the upstream source regions, with a pronounced center over the Maritime Continent resembling the leading EOF pattern (Fig. 1a). Monthly δ¹⁸Op values are consistently and relatively more depleted from May through October, with the largest differences during the late and post-monsoon months (Fig. 3e). Moisture transport along the tropical easterlies strengthens and broadens equatorward, channeling air masses through convective centers over the Maritime Continent and South China Sea, where they acquire additional ¹⁸O–depleted vapor, amplifying anomalies across East Asia (Fig. 3d). The summer-half precipitation increases over CEC and northern China relative to non-ENSO years (Fig. 3f), and the negative relationship with δ¹⁸Op is relatively stronger (r = –0.48 over CEC). Collectively, these results show that changes in monsoon circulation intensity can influence δ¹⁸Op over CEC and the broader EASM domain, while ENSO further strengthens and broadens this influence through additional upstream depletion.

Fig. 3: δ¹⁸Op and precipitation composites for high and low meridional wind index years.
Fig. 3: δ¹⁸Op and precipitation composites for high and low meridional wind index years.
Full size image

a Summer-half δ¹⁸Op anomalies (‰; shaded) composited as high minus low meridional (v) wind index years, with ENSO years included. Arrows show VIMF anomalies significant at the 95% level (two-tailed Student’s t-test). b Annual cycles of δ¹⁸Op averaged over CEC (27°–37°N, 104°–116°E) for high-index (blue) and low-index (red) years with filled circles marking months where the mean difference is significant at the 95% level (two-tailed Student’s t-test), open circles indicate insignificant difference. c Precipitation anomalies (mm/day; shaded) for the same composites as in (a). Precipitation data is from ref. 69. Stippling marks grid points where the difference is significant at the 95% confidence level, based on a two-tailed Student’s t-test of the difference in means. d–f Same as (a–c), but with ENSO years excluded. The dotted contour (pink) shows the 2500 m elevation, approximating the boundary of the Tibetan Plateau.

ENSO’s imprint on westerly jet and δ¹⁸Op variability

Beyond ENSO’s influence on low-level southerlies, it can also affect the upper-tropospheric circulation, particularly the position of the westerly jet relative to the Tibetan Plateau, with El Niño years typically inducing a southward bias44,45. Previous studies have shown that the seasonal northward migration of the jet plays an important role in shaping the EASM by influencing rainfall timing and distribution31. Analyses based on an earlier variant of IsoGSM3 suggested that a southward-displaced jet delays the onset of the EASM, allowing relatively heavier springtime δ¹⁸Op values to persist into summer while limiting the advection of ¹⁸O-depleted oceanic moisture due to weaker low-level southerlies and vice versa32. Here, we leverage the longer and improved IsoGSM3, which extends back to 1950 and thus provides an expanded ensemble of ENSO events, to test how such shifts in the jet position shape the δ¹⁸Op response over the EASM domain and whether this response is modulated by ENSO.

We first examined differences in the mean position of the westerly jet between El Niño (N = 17, DJF(0) years) and La Niña events (N = 17), at both monthly (Fig. S7) and seasonal (summer-half) scales (Fig. 4a). The jet position was defined as the latitude of maximum zonal wind speed at 200 hPa and computed as the monthly mean within 25°–55°N and 40°–120°E (see “Methods”). Over the East Asian sector of the westerly jet (75°–120°E), the jet position shows no significant or systematic difference between El Niño and La Niña years from May through July (Fig. S7). However, the jet is displaced significantly southward during El Niño years in September through October, with the largest shift in October (Δ = −3.51°, 95% CI −4.46 to −2.46, p < 0.001). When averaged across the summer-half, the jet lies ~1° farther south in El Niño years, but this bias is sourced almost entirely from September and October.

Fig. 4: Summer-half westerly jet positions and multi-field MCA analysis.
Fig. 4: Summer-half westerly jet positions and multi-field MCA analysis.
Full size image

a Mean jet latitude composited for El Niño (red) and La Niña (blue) years. The dotted contour (blue) shows the 2500 m elevation, approximating the boundary of the Tibetan Plateau. The average latitudinal position of the jet, the mean difference (Δ) between the two composites, and the associated 95% confidence intervals and p-values are reported on the figure. Significance was assessed using both a two-tailed Student’s t-test and bootstrap resampling (see “Methods”). b Same as (a), but for high (red) and low (blue) CEC δ¹⁸Op index years after regressing out the Niño 3.4 influence (see “Methods”). c, d Multi-field maximum covariance analysis (MCA) between tropical Pacific SST (shaded, left panel, c) and East Asian δ¹⁸Op (shaded) and 200 hPa zonal winds (contours) (right panel, d) for September–October. The SST data is from ref. 68. Shown are homogeneous regression maps for SST (units: °C per normalized SST time series) and heterogeneous regression maps for δ¹⁸Op (‰ per normalized SST time series) and zonal winds (m s−1 per normalized SST time series). Contours are plotted at 0.5 intervals, with negative values dashed, positive values in solid blue, and the zero contour in thick black. The Tibetan Plateau (≥2500 m) is outlined with a blue dotted line. The leading MCA mode explains 83% of the squared covariance fraction.

We repeated the same analysis to estimate the differences in the jet position between high- and low-CEC δ¹⁸Op years. The seasonal evolution of the jet position in this case (Fig. S8a) is broadly similar to the El Niño and La Niña years, with a statistically significant southward-displaced jet position only in September and October, albeit with a relatively smaller amplitude (for example, October Δ = −1.91°, 95% CI −3.23 to −0.55, p < 0.01) (Fig. S8a). This overall similarity with the monthly jet positions in El Niño and La Niña years is not unexpected, given that high CEC-δ¹⁸Op years (N = 18) include several El Niño years (N = 11) and vice versa. Removing El Niño and La Niña years in a parallel analysis (N = 7) illustrates the key differences. The southward bias in the jet position during September and October disappears, while other months show broadly similar patterns, though small sample sizes reduce statistical confidence (Fig. S8b). To overcome this constraint, we regressed out the Niño-3.4 signal from the CEC-δ¹⁸Op index and reclassified the residuals into high (75th percentile, N = 18) and low (25th percentile, N = 18) years. The composites of monthly jet positions for high- and low-δ¹⁸O residual years show no systematic differences, except during May and August, when the jet lies farther north during high δ¹⁸O residual years (Fig. S9). When averaged over the summer-half, the mean jet position lies slightly northward during positive δ¹⁸O residual years (Fig. 4b). Overall, while these results partly echo the δ¹⁸O–westerly jet relationship identified in an earlier study32, our analysis demonstrates that the linkage arises primarily from ENSO’s influence on the subtropical jet, expressed most strongly during September and October.

A multi-field MCA of tropical Pacific SST (left field) with concatenated δ¹⁸Op and 200 hPa zonal winds (right field) further identifies September–October as the seasonal window when ENSO’s imprint on East Asian δ¹⁸Op is also most pronounced (Fig. 4c, d). The leading mode explains 83% of the squared covariance fraction and links canonical El Niño SST anomalies with positive δ¹⁸Op over East Asia, accompanied by upper-tropospheric zonal wind anomalies characterized by weakened westerlies north of the Tibetan Plateau and a narrow band of strengthened flow to the south. This ENSO-related covariance pattern suggests an equatorward displacement of the subtropical westerly jet as it migrates south of the Plateau during the withdrawal phase of the EASM. In contrast, a parallel MCA analysis for the core monsoon season (JJA) shows relatively weaker El Niño SST anomalies, associated with spatially diffuse and muted positive δ¹⁸Op anomalies over East Asia, and a circulation pattern consistent with the climatological summertime state, with the westerly jet positioned well to the north of the Plateau (Fig. S10).

An enhanced equatorward shifted position of the subtropical jet in September–October during the developing and mature phase of El Niño44,45 is dynamically consistent with the extratropical response to anomalous tropical diabatic heating31, in which Rossby wave propagation modifies upper-tropospheric flow and alters meridional temperature gradients over the East Asian sector (Fig. 4a). As the EASM weakens in September–October, an equatorward-displaced subtropical jet (Fig. S7) suppresses late-season southerly monsoonal flow into East Asia resulting in reduced contribution of ¹⁸O-depleted oceanic moisture, while there is a corresponding increase in contribution of ¹⁸O-enriched moisture from terrestrial recycling, which together favor higher δ¹⁸Op over the EASM domain (Fig. 4d). This physically plausible pathway linking the ENSO to late-season δ¹⁸Op variability over East Asia is supported by a strong negative correlation (r = −0.51) between late-season CEC δ¹⁸Op anomalies and the latitude of the 200 hPa jet maximum (Fig. S7c).

Moisture dynamics and CEC δ¹⁸Op variability

The preceding analyses show a larger role of ENSO than previously recognized in shaping interannual δ¹⁸Op variability over the EASM through its impact on upstream convection and the position of the westerly jet. Changes in moisture dynamics, such as shifts in moisture transport pathways or contributions25,26,27,28,29,30 have also been proposed as important controls on interannual variability. To evaluate this perspective, we use the IsoGSM3 water-tagging framework, in which the model assigns evaporated vapor to regional tags and tracks it through until precipitation occurs. For this study, eleven source regions were defined across the Indian, Atlantic, and Pacific Oceans, and major continental areas (Fig. 5a). We aggregated all individual terrestrial sources into a single (LND), combined the South China Sea with the Pacific Ocean (PO), and merged the tropical and southern Indian Ocean into one Indian Ocean source (IO) (see “Methods”). Together, contributions from these sources account for ~99% of column-integrated total precipitable water (TPW) and total precipitation (TP) over both the summer half-year (May–October) and winter half-year (November–April) within CEC (Fig. S11). Other sources, such as the North Atlantic (NATL), Bay of Bengal (BoB), and Arabian Sea (ARAB), were retained but not emphasized here in the subsequent discussion, as their contributions to the CEC summer-half moisture budget are minor (Fig. S11).

Fig. 5: Tagged moisture source regions in IsoGSM3 and their climatological characteristics.
Fig. 5: Tagged moisture source regions in IsoGSM3 and their climatological characteristics.
Full size image

a Schematic representation of moisture source regions considered in this study (see “Methods”). The oceanic sources include Arabian Sea (ARAB); Bay of Bengal (BoB); tropical Indian Ocean (TIO); southern Indian Ocean (SIO); South China Sea (SCS); Pacific Ocean (PAC); South Atlantic (SATL); North Atlantic (NATL). The terrestrial sources include Indo-China Peninsula (ICP); the Indian subcontinent (IND), and continental China (CHN). For analysis, the 11 sources are grouped into three components: Indian Ocean (IO), comprising TIO and SIO; Pacific Ocean (PO), comprising PAC and SCS. All terrestrial sources (CHN, IND, ICP, and all remaining land surfaces) are lumped together to form a single land source (LND). b Climatological summer-half contributions of total precipitable water (TWP) from the three aggregated moisture source regions: LND (shaded), IO (solid blue contours), and PO (dashed contours). The CEC domain is indicated by the rectangle. The dotted contour (white) shows the 2500 m elevation, approximating the boundary of the Tibetan Plateau. c Annual cycles of fractional moisture contributions from the three aggregated source regions (IO, PCO, and LND) and the individual ARAB and BoB regions. Shading around each curve indicates ±1 standard deviation. IO, PCO, BoB, and ARAB are plotted against the left y-axis; LND is plotted against the right y-axis.

During the summer-half, within CEC and across China, LND dominates the moisture budget, contributing 40–70% of both TPW and TP, with values decreasing southward (Fig. 5b; see also Fig. S12a, c). The IO contributes 10–20% of TPW and TP, with contributions tapering northward, while PO contributes 10–30% and decreases westward within the CEC (Fig. 5b). During the winter-half year, LND remains the largest contributor (50–70%) to TPW and TP over CEC (Fig. S12b, d) consistent with previous estimates46,47,48. The seasonal evolution of moisture contributions over CEC (Fig. 5c) shows a marked increase in IO moisture, reaching 15–20% during the core monsoon season (JJAS). In contrast, PO contributions decline through winter and spring to a minimum in June and then increase through the summer and fall, peaking at ~25% by October. Ag Against this climatological backdrop, we next assess whether interannual variability in δ¹⁸Op over CEC is accompanied by shifts in the relative contributions of oceanic and terrestrial moisture sources under contrasting low-level monsoon circulation, and whether these shifts are modified by ENSO.

We address this question using a single diagnostic that isolates the role of ENSO in circulation-related moisture-source partitioning by contrasting strong and weak circulation states with and without ENSO-years included. Specifically, we compare high-minus-low v-wind composites constructed using the full v-wind index with equivalent composites from which ENSO years are excluded. In the ENSO-inclusive case, strong (weak) circulation years are preferentially associated with La Niña (El Niño) conditions, such that the moisture-source anomalies reflect the ENSO phase inherent in the circulation contrast. The difference between these two composites therefore, isolates the contribution of ENSO phase embedded within the circulation contrast to changes in moisture-source contributions over CEC (Fig. 6). This comparison shows that the inclusion of ENSO years produces only a small modification to the moisture-source budget over CEC with combined contribution from the IO and PO increasing by only ~2–3% during strong circulation years, accompanied by a compensating reduction in LND fraction.

Fig. 6: Difference between ENSO-inclusive and ENSO-excluded strong–weak circulation composites of summer-half moisture sources.
Fig. 6: Difference between ENSO-inclusive and ENSO-excluded strong–weak circulation composites of summer-half moisture sources.
Full size image

Panels show the difference between high-minus-low v-wind composites constructed using the full record with ENSO-years included and equivalent composites with ENSO years removed, thereby isolating the ENSO contribution to circulation-related changes in moisture-source partitioning. Shown are percent changes in a total precipitable water (TPW), b Indian Ocean (IO) moisture, c Pacific Ocean (PO) moisture, and d land recycling (LND) moisture during the summer half-year (May–October). Positive values indicate an increased contribution in ENSO-inclusive strong circulation states relative to ENSO-excluded states. Stippling denotes regions where differences are statistically significant at the 95% confidence level, and contours highlight the magnitude of the changes. The black box outlines the CEC region used for area-averaged diagnostics. The blue contour shows the 2500 m elevation, approximating the boundary of the Tibetan Plateau.

The limited importance of these ENSO-related changes in moisture-source partitioning becomes clear when they are evaluated against the magnitude of the CEC-averaged δ¹⁸Op anomalies diagnosed from the strong-minus-weak circulation composites shown in Fig. 3. Strong-minus-weak circulation composites excluding ENSO yield a CEC-averaged depletion of ~ −1.0‰, while ENSO-inclusive composites exhibit a larger depletion of ~ −1.4‰ (Fig. 3). A simple mass-balance consideration demonstrates that a 2–3% redistribution among moisture fractions is far too small to generate δ¹⁸Op anomalies of this magnitude, amounting to only a few tenths of a per mil even under generous assumptions. These results indicate that interannual δ¹⁸Op variability over CEC is not controlled by ENSO- driven changes in moisture-source contribution but instead reflects upstream convective and rainout processes that modify the δ¹⁸Op of vapor prior to its transport into CEC.

The magnitude of these shifts can be evaluated against the corresponding δ¹⁸Op response illustrated in Fig. 3, where strong-minus-weak circulation composite excluding ENSO yield a CEC-averaged depletion of ~ −1.0‰, and ENSO-inclusive composite shows a larger δ¹⁸Op depletion ~ −1.4‰. A simple mass-balance consideration shows that a 3–4% redistribution among moisture sources is far too small to generate δ¹⁸Op anomalies of this magnitude, amounting to only a few tenths of a per mil even under generous assumptions. This indicates that interannual δ¹⁸Op variability over CEC cannot be explained by ENSO- or circulation-related changes in the relative contributions of oceanic and terrestrial moisture sources. Instead, the dominant control must lie in upstream processes such as convective rainout and associated isotopic depletion that modify the isotopic composition of vapor prior to its transport into CEC.

Implications for interannual δ¹⁸Op variability and proxy records

Our analyses show that ENSO is the leading mode of δ¹⁸Op interannual variability in CEC and much of East Asia. ENSO imprint is expressed both through its modulation of upstream processes as well as shifts in the subtropical westerly jet position during the late monsoon season. Previous studies have linked interannual δ¹⁸Op variability in the EASM domain to north-south shifts in the subtropical westerly jet position, with positive δ¹⁸Op anomalies associated with a southward-displaced jet and weakened low-level southerlies32. Our results broadly echo this relationship but reveal two important distinctions. First, that the jet–δ¹⁸Op linkage is expressed primarily during the late or post-monsoon season, and second, that this relationship is largely mediated by ENSO forcing of the jet. Other studies have attributed interannual δ¹⁸Op variability to changes in the relative contributions of distant Indian Ocean versus more proximal Pacific and South China Sea sources29, or have framed the problem in an Indian Ocean-centric paradigm16,18. Our results suggest that the relative contributions of Indian and Pacific Ocean moisture are stable at interannual scales over CEC and do not vary enough to account for isotopic anomalies. Instead, their influence arises from a varying degree of depletion during transit through the upstream convective regions. This places the role of the Indian Ocean within a broader Indo-Pacific upstream convective framework, in which ENSO-driven modulation of convection exerts the dominant control on interannual δ¹⁸Op variability.

Although ENSO emerges as the dominant mode of interannual δ¹⁸Op variability across the EASM domain, the leading EOF explains only ~21% of the total variance, with remaining modes statistically indistinguishable (Fig. S1a). The ENSO signal in δ¹⁸Op therefore represents a modest fraction of total interannual variance, complicating its detection in proxy archives, particularly speleothem δ¹⁸O records. This limitation is amplified by karst mixing and by temporal averaging when speleothems are sampled at multi-year resolution, which attenuate interannual variability and redistribute variance toward lower frequencies. To heuristically illustrate this limitation, we conducted a simple sensitivity test pairing the last-millennium IsoGSM amount-weighted δ¹⁸Op output from CEC (Fig. 3C) with a speleothem proxy system model (PSM) (see “Methods”), which assumes a single, well-mixed reservoir of meteoric water within the karst with a specified mean transit time49. In the frequency domain, the input IsoGSM δ¹⁸Op shows a prominent power in the ENSO band (2–8 years), but when passed through the PSM, the variance in this band declines. At a one-year transit time, the modeled drip water δ¹⁸O retains ~75% of the original variance, indicating that the ENSO signal can still be reliably identified in high-resolution (<1 year) records. Beyond a two-year transit time, however, the signal can no longer be distinguished from background noise (Fig. S13). These results help explain why speleothem δ¹⁸O records from sites in central and eastern China, despite experiencing broadly similar climate regimes and being separated by only a few hundred kilometers, show poor coherence at interannual to decadal scales1,50,51,52, whereas speleothem records from the same region exhibit strong coherence on centennial to orbital timescales. In also help explain why oxygen isotopes in tree cellulose (δ¹⁸Otr) records from the EASM domain often retain a clear ENSO imprint in their power spectra53,54, as trees integrate δ¹⁸Op during the summer growing season.

Although our analysis focuses on interannual variability, the results carry broader implications for interpreting East Asian speleothem δ¹⁸O records on millennial to orbital timescales. At these timescales, variability in Chinese speleothem δ¹⁸O has traditionally been interpreted in terms of precessional insolation forcing and millennial-scale North Atlantic variability, which together produce a spatially coherent signal across the EASM domain1,2,3,20,23. Our findings suggest that part of this long-term coherence may also reflect large-scale reorganizations of Indo-Pacific convection. Such reorganizations need not mirror ENSO in frequency or structure but may represent low-frequency analogues of the same upstream convective processes that dominate δ¹⁸Op variability at interannual timescales. Under this framework, speleothem δ¹⁸O records may integrate changes in the mean state and variability of tropical convection over long timescales, providing a complementary perspective to interpretations based solely on monsoon intensity or insolation forcing.

Methods

CNTR Run

IsoGSM3 is the most recent version37 of the isotope-enabled model from the Scripps Experimental Climate Prediction Center version of the Global Spectral Model36. The control simulation (CTRL) spans from January 1950 to December 2023 and was conducted at T62L28 resolution, providing an approximate horizontal resolution of 1.875° (about 200 km) and 28 vertical sigma levels. The model uses the Noah Land Surface Model55, and atmospheric moisture is convected in the horizontal spectral grid and vertical sigma coordinate using the Relaxed Arakawa-Shubert scheme56. The new version incorporates the Non-iteration Dimensional-split Semi-Lagrangian tracer transport scheme57, which reduces spectral artifacts, prevents negative humidity, and improves vapor transport in dry and cold regions. It also replaces NCEP-R2 with ERA5 reanalysis58 for nudging, which is applied to temperature and wind speed every 6 h at all sigma levels, using a nudging coefficient of 0.9 for scales larger than ~1000 km, so that simulated circulation remains consistent with large-scale atmospheric states. These two updates mark the principal distinction from the previous version37. In a recent global evaluation against GNIP monthly δ¹⁸O, IsoGSM3 nudged to ERA5 achieved a higher mean correlation (~0.61) than the NCEP-R2 configuration (~0.57)37. Most isotopic fractionation in the model is treated as thermodynamic equilibrium, except for open-water evaporation, condensation under supersaturation (vapor–ice), raindrop re-evaporation, and air–rain isotopic exchange, which are treated with kinetic fractionation. All precipitation is fully mixed into a simple single bucket-type model for treatment of soil water isotopes, which provides storage of all water species, and consequently, the isotope ratio of evapotranspiration is assumed to be the same as stored values.

IsoGSM3 validation

The current and previous variants of the IsoGSM have been benchmarked extensively against observations across the globe36, and particularly over the South and East Asian monsoon domains14,32,34,38,39. The model’s performance is also comprehensively evaluated in coordinated intercomparisons with other isotope-enabled models32,33,34. For brevity, we do not reproduce the results from these studies here and instead refer readers to the detailed assessments presented in the original references, which are briefly summarized as follows. In a comprehensive inter-model comparison over East Asia, IsoGSM outperformed all other isotope-enabled models in reproducing the seasonal amplitude and phase of δ¹⁸Op, achieving a warm-season (JJASO) δ¹⁸Op correlation of r = 0.71 against more than 20 GNIP stations in eastern China (105°–121°E, 23°–35°N) (p < 0.01), while alternative models showed markedly weaker skill32. Similar cross-model comparison analysis for the Indian monsoon domain shows that IsoGSM ranked among the best performers in reproducing the annual cycle, spatial pattern, and interannual variability of monsoon rainfall with δ¹⁸Op correlations with GNIP and gridded precipitation products exceeding 0.85 in the core monsoon region34. In the broader Maritime Continent and adjacent tropics, the model reproduces seasonal and intraseasonal δ¹⁸O variations associated with convective migration, with significant site-level correlations to observations38. Over the Tibetan Plateau, a recent assessment shows that the simulated δ¹⁸Op performs well against multiple GNIP stations across the region, providing a stringent, region-wide validation of model skill14. A more recent process-based intercomparison further demonstrates that the latest IsoGSM version used in this study, when nudged to ERA5 reanalysis, delivers the highest level of observational fidelity37. In this configuration, IsoGSM3 achieved the strongest global agreement with observations, with mean correlations of ~0.67 for precipitation δ¹⁸O across more than 1000 GNIP stations and up to ~0.9 for vapor isotopes against satellite products37.

Sensitivity experiments on upstream isotopic modification

We conducted five IsoGSM3 sensitivity runs (SR1–SR5) in which isotopic fractionation was disabled only within a predefined region (Fig. S5a). Inside the selected box for a given run, we set α(¹⁸O) = α(D) = 1.0 for condensation and evaporation so that isotopic ratios remained unchanged during phase changes, and we set the falling exchange coefficient to zero to suppress raindrop–vapor isotopic exchange; outside the box, all model physics remained identical to the control. The defined regions are SR1 (90°–150°E, 10°S–10°N), spanning the equatorial eastern Indian Ocean through the Maritime Continent into the western Pacific; SR2 (30°–90°E, 10°S–10°N), covering the equatorial western and central Indian Ocean; SR3 (30°–150°E, 7°–26°N), encompassing the tropical Northern Hemisphere monsoon corridor from the western Indian Ocean across the Bay of Bengal, Indochina, the South China Sea, and the western Pacific; SR4 (30°–110°E, 26°–50°N), extending across midlatitude continental regions upstream of China in the prevailing westerlies; and SR5 (110°–120°E, 26°–50°N), located within continental eastern China and partially overlapping the CEC domain (Fig. 5a). Each sensitivity run is otherwise identical to the control simulation. We then compared the simulated CEC annual cycle of δ¹⁸O in precipitation between the control and each sensitivity experiment to diagnose whether suppressing isotopic modification within a given region altered the seasonal cycle in CEC (Fig. S5b).

Vapor tagging simulations

In tracer mode, IsoGSM tracked moisture back to its last contact with the surface by setting the surface evaporative fractionation factor to 1 for the source region of interest (and 0 elsewhere) and disabling all other atmospheric isotopic fractionation38. Evaporated water is marked with a tag identifying its origin and then follows the same atmospheric hydrological cycle as untagged vapor until it is removed as precipitation. We tracked water vapor originating from 11 source regions (Fig. 5): the Bay of Bengal (BoB), Arabian Sea (including the Red Sea and Gulf of Oman; ARAB), Tropical Indian Ocean (TIO), Southern Indian Ocean (SIO), Pacific Ocean (PO), South China Sea (SCS), North Atlantic (NATL), South Atlantic (SATL), Indian Peninsula (IND), continental China (CHN), and the Indochina Peninsula (ICP). Residual land surfaces (RLND) represented all remaining land areas. For some analyses, regions were combined: TIO and SIO as IO (Indian Ocean), SCS and PO as PO (Pacific Ocean), and all land sources (CHN, IND, ICP, RLND) as LND (land). In the case of terrestrial tags, IsoGSM3 assumed that all evapo-transpired vapor originated from a single, well-mixed soil-water reservoir, with no direct open-water evaporation, so the tagged land-sourced vapor represented recycled moisture from the land surface59. While several other studies have implemented an alternative approach in which δ¹⁸O of precipitation and vapor is diagnosed separately for each tagged source25,26,27, we did not adopt this method because the δ¹⁸O signal in IsoGSM3 is a nonlinear emergent property of the full hydrological cycle, shaped by Rayleigh-type condensation, sub-cloud exchange, re-evaporation, and mixing among air parcels. These processes cannot be reliably linearly partitioned back to individual source tags, and assigning per-source δ¹⁸O values would therefore be physically inconsistent.

EOF-analysis

An EOF analysis of IsoGSM3 amount-weighted summer-half (MJJASO) δ¹⁸Op was performed over a broad domain (0°–50° N, 70°–130° E) for 1950–2020. Prior to analysis, δ¹⁸Op anomalies at each grid point were detrended and standardized by their standard deviation to normalize variance across space and prevent regions of high variability from dominating the EOF solution. The leading mode (EOF1) explains 21% of the total variance and exhibits a spatially coherent pattern across the analysis domain (Fig. 1a). Its associated principal component (PC1) was standardized for comparability with other indices and shows a strong correlation with CEC (27°–37° N, 104°–116° E) δ¹⁸Op (Fig. 1b). The distinctness of EOF1 was confirmed using North’s rule of thumb60, which tests whether adjacent eigenvalues are statistically separable (Fig. S1a). To evaluate the robustness of this large-scale mode beyond IsoGSM3, we conducted parallel EOF analyses on several other isotope-enabled models ranging from a stand-alone atmospheric model to atmospheric components of fully coupled GCMs. These included ECHAM6-wiso nudged to ERA5 reanalysis61, iCAM5, the atmospheric component of the Community Earth System Model62, an isotope-enabled version of the GISS ModelE2-R63, and IsoGSM2, the earlier version of IsoGSM3, which is nudged to NCEP R2 reanalysis36. Model outputs are available for varying periods beginning in 1979, and further details can be found in the original references (Fig. S2). Finally, to assess whether the dominant δ¹⁸Op mode persists over longer timescales and under different boundary conditions, we extended the analysis to the last-millennium IsoGSM simulation41 (851–2000 CE), which was forced with the Community Climate System Model version 4 (CCSM4) SST and sea-ice distributions.

Maximum covariance analysis

MCA extracts coupled modes of variability between two fields by singular value decomposition of their cross-covariance matrix64. The first mode of MCA maximizes the covariance between the two data sets, while successive pairs describe a maximum fraction of quadratic covariance not explained by the previous pairs. The squared covariance fraction quantifies the fraction of squared covariance between two fields represented by a given mode of variability (i.e., the proportion of the total variance jointly explained). Additionally, the algorithm computes values for the fraction of variance separately, which quantifies how much of the total variance in the individual field is accounted for by a given mode. MCA was applied to summer-half δ¹⁸Op and SST fields (Fig. 1c). Anomalies were calculated by removing the mean and linear trend at each grid point, and fields were area-weighted by the square root of cosine latitude. Regression maps were constructed by regressing δ¹⁸Op anomalies (‰) (left field) and SST anomalies (°C) (right field) onto the SST expansion coefficient, yielding heterogeneous and homogeneous maps, respectively. Values represent anomalies per one standard deviation of the SST expansion coefficient. Statistical significance was assessed using a two-tailed Student’s t-test at the 10% level (α = 0.1). To identify coupled modes between SST and both δ¹⁸Op and U200 (Figs. 4c and S10), we also conducted a multifield MCA in which the left field was SST and the right field was formed by concatenating δ¹⁸Op and 200-hPa zonal winds, with anomalies preprocessed in the same way as for the two-field case (detrended, area-weighted, and normalized). This approach allows SST patterns to be extracted that maximize shared covariance with both isotopic and circulation fields simultaneously.

Multitaper spectral analysis

Spectral analysis was conducted using the Multitaper Method (MTM)65, which reduces spectral leakage and variance by employing a set of orthogonal tapers rather than a single spectral window. We used a time–bandwidth product (nw) of 2, offering a balance between frequency resolution and statistical stability. To assess the significance of spectral peaks, we applied a robust red-noise test based on a least-squares fit of a lag-1 autoregressive (AR(1)) model to a median-smoothed version of the raw spectrum66. This approach mitigates the inflation of false positives associated with conventional AR(1) models, particularly in the low-frequency range where strong periodic or quasi-periodic signals may distort background estimates. To further minimize the risk of false positives, we utilized a Monte Carlo approach, generating 10,000 random AR1 null distributions with the same lag-1 and variance as the original input data. Peaks exceeding the AR1 null model distribution are considered robust.

Definition of ENSO years

We used the NOAA definition of ENSO years based on the Oceanic Niño Index (ONI) https://origin.cpc.ncep.noaa.gov/products/analysis_monitoring/ accessed on (09/16/2025). The ONI is calculated as a 3-month running mean of SST anomalies in the Niño-3.4 region (5°S–5°N, 170°–120°W). El Niño (La Niña) events are defined when these anomalies exceed +0.5 °C ( − 0.5 °C) for at least five consecutive overlapping 3-month seasons. The developing, or growing, phase corresponds to the boreal summer (JJA/SON) preceding a mature ENSO event in DJF(0/1). Developing years are thus defined as JJA(0) preceding mature El Niño or La Niña conditions in DJF(0/1), while decaying years are defined as MAM(1) and JJA(1), when ENSO conditions subside. Numbers in parentheses indicate year relative to the peak DJF season (e.g., JJA(0) = summer preceding the mature event, JJA(1) = summer following it). Based on NOAA’s definition, 17 growing-phase El Niño years (1953; 1957; 1963; 1965; 1972; 1976; 1982; 1987; 1991; 1993; 1997; 2002; 2004; 2006; 2009; 2015; 2018) and 15 La Niña years (1954; 1955; 1956; 1964; 1967; 1970; 1971; 1973; 1975; 1984; 1988; 1999; 2007; 2010; 2011) are identified during 1950–2020. We note that different studies sometimes report slightly different ENSO year lists despite citing the same ONI definition, due to updates in the underlying SST dataset, treatment of borderline cases near the ±0.5 °C threshold, or differences in how multi-year or overlapping events are segmented. Such minor discrepancies, however, do not materially affect the conclusions of our analysis.

Composite analyses

We conducted a series of composite analyses based on multiple indices derived from IsoGSM3 output. These indices included CEC (27°–37°N, 104°–116°E) δ¹⁸Op, the leading principal component (PC1) of the EOF analysis, a meridional (v-wind) circulation index, and the monthly and seasonally averaged latitude of the upper-tropospheric westerly jet. Composites were constructed using high and low years, defined as those exceeding the 75th percentile or falling below the 25th percentile of each index distribution, respectively. Applying this criterion to the full 74-year (1950–2023) simulation yields 18 years in each group, which provides balanced sample sizes and avoids undue influence from isolated extreme years. To remove the role of ENSO, we repeated the analysis after excluding ENSO years as defined by NOAA’s ONI (see ENSO definition). This generally entailed eliminating both El Niño and La Niña years from the high and low groups of each index, but in practice, the relative impact differed by index. For example, in the high v-wind index, excluding ENSO years primarily meant the removal of La Niña years, while in high PC1 years (signifying lower upstream convection), it reflected the removal of El Niño years. After this adjustment, each group contained about eight or nine years, which was sufficient for generating stable composites and identifying robust isotopic and circulation anomalies when the signals are spatially coherent. The main exception was assessing the significance of the westerly jet position in the absence of ENSO years, where the reduced number of years limited statistical confidence. In this case, the summer-half Niño-3.4 index was first regressed out of the CEC δ¹⁸Op series (the two series are correlated at r = 0.7, p < 0.01), ensuring that the resulting composites isolated jet anomalies independent of ENSO forcing.

Westerly jet latitude positions

We computed the latitude of the upper-tropospheric westerly jet (200 hPa) for each year of the IsoGSM simulation. At each longitude within the analysis domain (25°–55°N, 40°–140°E), the jet core was identified as the latitude of maximum zonal wind speed. This procedure was repeated for all months from May through October, producing a monthly jet latitude time series across the domain. The jet positions were composited separately for El Niño and La Niña years, and the differences between the two groups were evaluated using two complementary approaches. First, we applied bootstrap resampling, in which the years belonging to each group were randomly resampled with replacement 5000 times. This procedure produced an empirical distribution of the mean difference in jet latitude, from which the mean estimate (Δ) and its 95% confidence interval were derived. Second, we conducted a permutation test by randomly reassigning the ENSO group labels (El Niño or La Niña) 10,000 times, recalculating the difference each time. The proportion of permutations yielding a difference equal to or greater than the observed value provided a two-sided p-value for the null hypothesis of no real group difference. Statistical testing was focused on the East Asian sector (75°–120°E), where the jet exerts a primary control on monsoonal circulation and moisture transport into China. We also repeated the analysis using the CEC δ¹⁸Op index (after removing the linear influence of concurrent May–October Niño-3.4 from the CEC δ¹⁸Op series (r = 0.7, p < 0.01), and the resulting residuals were then used to define high and low years (upper and lower 25% of the residual distribution, N ~ 17 per group over 1951–2020). The same bootstrap and permutation procedures were applied to assess the significance of differences in jet latitude between these residual-based composites. The monthly jet latitude results were also averaged across the summer-half (MJJASO) to obtain the seasonally averaged jet position, with statistical testing performed in the same manner.

Proxy system model

To assess how karst residence time affects the preservation of ENSO-related variability in precipitation δ¹⁸O, we employed the speleothem sensor model in its well-mixed reservoir configuration67. Input was the summer-half amount-weighted δ¹⁸O time series extracted from the IsoGSM last-millennium simulation over the CEC domain, which exhibits strong spectral power in the 2–8-year ENSO band. The model was iteratively run with mean residence times (τ₀) set to 1, 2, 3, and 4 years, a range chosen to approximate short residence times inferred for many thin-overburden cave systems. Our objective was to isolate the effect of karst mixing on signal transmission, so we focused on drip water δ¹⁸O (δ¹⁸Ok) and omitted the calcite fractionation step, which would introduce only a systematic offset without altering variability. Both input and modeled output series, each spanning 1000 years, were analyzed using the MTM spectral estimator. To quantify the impact of mixing on the ENSO signal, we computed ENSO-band retention as the ratio of integrated spectral power between 2 and 8 years in the modeled drip water series relative to the precipitation input. This experiment is intended as an illustrative test of signal preservation rather than a site-specific calibration.

Statistical tests

All statistical tests used in this study are described in the “Methods” and figure captions. Unless otherwise noted, tests are two-sided, with significance evaluated at the 95% confidence level using a Student’s t-test. Composites and regression analyses are based on independently sampled years that meet the criteria specified in the “Methods”. For the westerly jet position, statistical significance of group differences was additionally evaluated by bootstrap resampling as detailed in the “Methods”.