Introduction

Atmospheric rivers (ARs), “the rivers in the sky”, are defined as long, narrow, transient corridors of intense horizontal moisture transport in the atmosphere1. ARs have drawn increasing scientific attention because of their crucial role in the global hydrological cycle and extreme weather events2,3,4. In western North America and Europe, where many megacities, large populations, and complex ecosystems are located, ARs are particularly vital. Extreme AR landfalls can bring life-threatening conditions, including heavy precipitation, strong winds, flooding, and landslides, causing substantial socioeconomic damage5,6,7,8,9,10,11. Conversely, ARs provide beneficial water vapor transport from ocean to land, sustaining ecosystems, water resources, and agriculture12,13,14. For instance, much of western North America is arid, with most annual precipitation occurring during winter months. Mountain snowpack serves as a natural reservoir, storing winter precipitation and releasing it during the warm and dry season. Dry winters with fewer AR landfalls result in reduced snowpack15,16, potentially leading to drier springs or early summers, with a corresponding rise in wildfire risks17,18. Additionally, a recent study highlighted the role of ARs in atmospheric energy transport, contributing to anomalously warm winters and extreme winter heat events19.

Advancing scientific understanding of AR dynamics under ongoing global warming is critical for water resource management, climate risk adaptation, and societal resilience. Many modeling studies have illustrated that ARs are likely to become more frequent, intense, and widespread under future warming20,21,22,23,24, consistent with increased atmospheric moisture content in a warmer atmosphere (see review25). However, whether ARs have already increased during the historical period has received relatively less attention.

Several recent studies have aimed to address this question; however, all of them focused on AR trends only after 197926,27,28,29. During this period, two main interdecadal climate oscillations each underwent phase transition (i.e., the positive-to-negative Interdecadal Pacific Oscillation (IPO) and the negative-to-positive Atlantic Multidecadal Oscillation (AMO)), dominating the linear trend of AR frequency. During 1980–2021, the increasing AR frequency over the Arctic was contributed by atmospheric moistening in the warming Arctic, with spatial patterns shaped by the IPO and AMO phase transitions30. From 1980 to 2020, AR frequency has increased over the eastern United States, primarily due to the negative trend of the Pacific North America (PNA) pattern31. Globally, AR activities have exhibited poleward shifts during 1979–2021, also associated with the positive-to-negative IPO phase shift32. When we extend the analysis back to 1950–2022, during which the AMO and IPO each underwent two phase transitions and showed no significant linear trends, we find that these interdecadal oscillations exert limited influence on the long-term AR trends (as described below). During this longer period, the AR trends exhibit distinct spatial patterns and are governed by different mechanisms.

In this study, we focus on wintertime AR activities (December to March) from 1950 to 2022 over the mid-latitude Northern Hemisphere. We identify “wetter north–drier south” meridional dipolar patterns of AR trends in both western North America and Europe. Using station-based observations, we further verify the essential role of ARs in driving corresponding north-south contrasts in both total and extreme precipitation trends over land. We find that these meridional dipolar AR trends are not linked to meridional displacements of the westerly jet stream. Instead, they coincide with positive trends of the PNA and North Atlantic Oscillation (NAO). We also show that the atmospheric moistening in the warming climate has contributed to AR’s increase rate by ~0.6–0.8% per decade, highlighting the thermodynamic contribution to AR trends. Through the large-ensemble atmospheric model analysis, we show that observed sea surface temperature (SST) changes dominate Atlantic AR trends while having minor effects in the Pacific. Last but not least, we make our AR database publicly available, which is based on the PanLu AR algorithm33,34 with \(1^\circ \times 1^\circ\) spatial and 6-hourly resolution during 1950–2022.

Results

Observed historical trends of atmospheric rivers and precipitation

During 1950 to 2022, winter AR activities have become more frequent across the mid-latitude Northern Hemisphere, with distinct inter-basin contrasts (increasing more intensive in the Atlantic than the Pacific) and meridional dipole patterns (increase in the north and decrease in the south) (Fig. 1a). Here, wintertime AR frequency for each grid cell is defined as the percentage of December–March (DJFM) days with AR occurrence (Methods). Slightly shifts in the study period do not change the spatial pattern of AR trends (Supplementary Fig. 1). Consistent patterns are also present in the historical trends of extreme precipitation (EP) frequency (Fig. 2a) and seasonal total precipitation (Supplementary Fig. 2a) based on the ERA5 reanalysis data. We also verify the precipitation trends using two station-based gridded precipitation datasets E-OBS over Europe35 and CPC over the contiguous U.S. (CONUS)36 (Fig. 2b, c). EP days are defined as daily precipitation exceeding both the 85th percentile threshold and 5 mm, with the 5 mm threshold ensuring a focus on hydrologically significant precipitation. When categorizing precipitation as AR-related (occurring within 2° of ARs) or AR-unrelated types, large proportions of both extreme and total precipitation are associated with ARs, as shown by the climatological spatial distribution of AR-related fraction in Supplementary Fig. 3. Only AR-related EPs show similar spatial patterns with total EP trends, while AR-unrelated EPs show modest changes (Supplementary Fig. 4). These results highlight the essential role of ARs in driving long-term changes in extreme and total precipitation. Notably, changing the EP threshold to the 80th or 90th percentile does not change the result (Supplementary Figs. 5 & 6).

Fig. 1: Historical trends of wintertime AR frequency from 1950 to 2022 based on ERA5 reanalysis.
figure 1

a Linear trend of AR frequency (shading), with AR frequency climatology (contours). Grid cells with significant trends at the 99% confidence level (p-value < 0.01) are marked by black scatters. b–g The 73-year time series of winter AR days (black dashed lines) in each target region (polygons in panel (a)), with seven-year running averages (black solid lines) and its linear trend (blue lines). Solid blue lines denote trends significant at the 95% level (p-value < 0.05), while dashed blue lines indicate insignificant trends. The increasing rates (slope) of winter AR days are provided in the panel legends. For each region, winter AR days are defined as days when > 10 grid cells (\(1^\circ \times 1^\circ )\) within the region are covered by ARs during DJFM (122 days in total).

Full size image
Fig. 2: Consistent spatial patterns in historical extreme precipitation trends.
figure 2

a Linear trends of winter extreme precipitation (EP) frequency from 1950 to 2022 based on ERA5 reanalysis data (shading), alongside AR frequency trends (contours). b and c the same as a but based on CPC precipitation dataset over the contiguous U.S. (CONUS) and E-OBS precipitation dataset over Europe, respectively. Grid cells with significant trends at the 99% confidence level (p-value < 0.01) are marked by black scatters. d–i 73-year time series of winter EP days (black dashed lines) in each target region (polygons in panel (a)-(c)), along with the seven-year running averages (black solid lines) and its linear trend (black straight lines). Red and blue lines denote AR-related (EP days coinciding with ARs) and AR-unrelated EP days, respectively. Regional EP days are defined as days when regional mean precipitation exceeds the 85th percentile threshold. Observational precipitation data are used to define regional EP days for CONUS and Europe, whereas ERA5 data are used for the Hawaiian Islands and British Columbia. Solid lines denote trends significant at the 95% confidence level (p-value < 0.05), while dashed lines indicate insignificant trends. The increasing/decreasing rates (slopes) are provided in the panel legends.

Full size image

The most pronounced increase in AR frequency is observed over the northern North Atlantic, where AR frequency has risen by nearly 50% over 73 years (i.e., ~7% increase relative to the ~15% AR frequency climatology) (Fig. 1a). Over the British Isles, for example, the number of winter AR days has increased by ~2.7 days per decade (statistically significant at the 95% confidence level; p-value < 0.05), from ~55 days in the 1950s to ~75 days in the 2020 s (Fig. 1c). Such intense AR increase has driven more frequent EP days in the British Isles, while AR-unrelated EP days show no significant trend (Fig. 2e). In eastern Europe and western Russia, where climatological AR frequency is much lower (~5%), winter AR days have surged from ~25 days in the 1950s to ~40 days in the 2020 s (Fig. 1d). Driven by enhanced inland AR penetration, the total precipitation has increased by nearly 50%, from ~110 mm (1950s) to ~170 mm (2020 s) (Supplementary Fig. 2f). Such winter precipitation increases, often falling as snow, could alter the winter snowpack and snowmelt in the coming spring, with potential implications for agriculture and hydrology in the warm seasons. On the other hand, AR activities have not increased everywhere along with the increasing global mean temperature. Slight decreases in AR frequency are evident over the Iberian Peninsula and Morocco (Fig. 1a, g), alongside corresponding declines in precipitation (Fig. 2c). This meridional dipolar pattern of AR trends is closely linked to the positive trend in the North Atlantic Oscillation (NAO) in recent decades (will be discussed later).

In the Pacific sector, the AR activities exhibit weaker increasing trends, while the meridional dipolar pattern is still present along the west coast of North America, with pronounced increases over the Gulf of Alaska (Fig. 1a). Over northern British Columbia, winter AR days have risen from ~45 days in the 1950s to ~60 days in the 2020 s (Fig. 1b), driving the more extreme precipitation in that region (Fig. 2b). In contrast, an “AR increasing hole” (a region of little to no increase, or slight decrease, in AR frequency) is observed in the Pacific Northwest (PNW) of the U.S. (Fig. 1f). The PNW is a climatological AR landfall hotspot (see Fig. 1a contours). Its mountain snowpack, primarily sustained by winter precipitation and AR landfalls, serves as the key water source for hydrology and agriculture during the snowmelt season37. The total winter precipitation over the PNW has declined from ~800 mm (1950s) to ~600 mm (2020 s), a significant decrease of 22.3 mm per decade (p-value < 0.05; Supplementary Fig. 2h). This long-term winter drying could threaten snowpack resilience, potentially prolonging spring/summer aridity, extending snow-free periods, and exacerbating wildfire risk17,18,38. Notably, AR-related precipitation accounts for ~88% of the total decline (19.7 of 22.3 mm per decade), despite ARs contributing only ~50% of total winter precipitation climatologically (Supplementary Fig. 2h). In other words, the drying trend in the PNW has been predominantly an AR-driven decline. We noticed the inconsistency between CPC and ERA5 precipitation trend over inland regions of western U.S. In the east of the Cascade Range, the complex topography makes the interpolation of station data (in CPC) and the simulation of precipitation (in ERA5) more challenging. This limitation means that the long-term precipitation trends in these interior regions are indeed uncertain. Thus, we do not place confidence in, nor draw conclusions from, the localized precipitation trends over the inland PNW in our study.

Another “AR increasing hole” in the Pacific sector appears near the Hawaiian Islands (Fig. 1e), implying a suppression of the so-called “Pineapple Express” (PE) ARs in recent decades. The PE ARs are a special category of ARs that originate near Hawaii and travel northeast, make landfall and bring heavy precipitation to the west coast of North America39,40. Interestingly, if we select the PE ARs only, we can observe their overall decreasing trend along their typical pathway from Hawaii to the PNW (Fig. 3a). In contrast, all other Pacific ARs (those not originating in the Hawaiian Islands box) exhibit no significant decreasing trend (Fig. 3b). In all the landfalling AR over PNW, PE ARs account for the majority of the declining AR trend, despite their lower frequency compared to other Pacific ARs (Fig. 3c). This comparison suggests that reduced AR activity in the PNW is primarily driven remotely by suppressed PE ARs originating near Hawaii. When further categorizing the AR-related precipitation into PE-related and other ARs-related, we can observe the pronounced decline of PE-driven precipitation along its pathway (Fig. 3d). Remarkably, the PE ARs contribute ~26% (5.9 mm/decade) of the PNW’s total precipitation decline (22.3 mm/decade), although they only account for ~6% (~45 mm) of PNW’s climatological precipitation (~711 mm) (Fig. 3f). Other Pacific AR types account for an additional decline of 13.8 mm/decade, implicating secondary drivers of PNW aridification to be discussed later.

Fig. 3: Suppressed “Pineapple Express” (PE) ARs reduce AR landfalling and precipitation in the Pacific Northwest (PNW).
figure 3

a, b Linear trend of winter PE ARs and all other Pacific ARs (shading) and their climatology (contours). PE ARs are defined as AR events originating from the box near the Hawaiian Islands (19°N–25°N, 160°W–150°W). Grid cells with trends significant at the 99% confidence level (p-value < 0.01) are marked by black scatter. c The 73-year time series of winter AR days over PNW driven by PE ARs (red lines) and other Pacific ARs (blue lines), along with seven-year running averages (solid lines) and linear trends (straight lines). d, e Linear trend of PE AR-related and other AR-related precipitation. f The 73-year time series of PNW precipitation caused by PE ARs (red lines) and other Pacific ARs (blue lines), with the seven-year running averages (solid lines) and linear trends (straight lines). Solid straight lines denote trends significant at the 95% confidence level (p-value < 0.05).

Full size image

Different roles of dynamic and thermodynamic effect

To further disentangle the drivers of historical AR trends, we decompose the AR changes into thermodynamic contribution (changes in atmospheric moisture content) and dynamic contribution (changes in large-scale atmospheric circulation) (Methods)30,41,42. By suppressing interannual variation of specific humidity, we first constructed hypothetical vertically integrated vapor transport (IVT) fields and built a “dynamic” AR database. The linear trend of “dynamic” AR frequency could represent the dynamic contribution to historical AR trends (Fig. 4a). The thermodynamic contribution is then estimated as the difference between the observed AR trends and the dynamic contribution (Fig. 4b). Here, we would like to clarify that this decomposition method is a qualitative diagnostic tool rather than a strict cause-and-effect attribution.

Fig. 4: Different roles of dynamic and thermodynamic effects in historical AR trends.
figure 4

The decomposed (a) dynamic and (b) thermodynamic contributions to AR trends. c Historical trends in 500 hPa zonal wind (U500) (shading), 500 hPa geopotential height (Z500) (contours), and 500 hPa wind vectors. d Historical trends in the IWV field. e Historical trends in U200 (shading), climatological U200 (contours), and 200 hPa wind vectors. Grid cells with trends significant at the 99% confidence level (p-value < 0.01) are marked by black scatters. f Regional comparison of total, dynamic, and thermodynamic effects of AR trends across six target regions.

Full size image

The greater availability of water vapor in the atmosphere alone has increased AR frequency by about 0.6–0.8% per decade since 1950 in many regions, including the Gulf of Alaska, northwest Europe, eastern US, Kuroshio current, and Gulf of Stream region (Fig. 4b). The ubiquitous positive thermodynamic contribution of AR trends is broadly consistent with the increasing trend of integrated water vapor content (IWV) field (Fig. 4d). Their pattern correlation (r) is 0.74 over the grids where the climatological AR frequency exceeds 5%, validating the robustness of the scaling method applied in this study. Nevertheless, the spatial pattern of thermodynamical contribution of AR trends looks substantially different from the observed AR trends (compare Fig. 4b and Fig. 1a). Instead, the observed AR trends closely resemble the dynamically driven AR trends (Fig. 4a) (pattern correlation \(r=0.89\)), underscoring circulation changes as the primary driver of the spatial contrasts of AR trends. Consistently, the dynamic contribution of AR trends highly resembles the seasonal mean 500 hPa wind speed trends (\(r=0.77\); Fig. 4c). A summary of the dynamic and thermodynamic contributions in several key regions is presented in Fig. 4f. In short, while atmospheric moistening raises the AR frequency overall, the changes in large-scale circulation shape the spatial contrasts, either offsetting or enhancing the thermodynamic effect across different regions.

Over the North Atlantic, the meridional dipolar AR trend and associated precipitation changes is dominated by dynamical effects (Fig. 4a). More specifically, it is primarily driven by positive trend of the North Atlantic Oscillation (NAO)43,44,45. In the long-term trend of 500 hPa geopotential height (Z500; Fig. 4c), the positive NAO-like pattern is present, with anomalous westerly/easterly winds over the northern/southern North Atlantic, leading to more AR landfalls in northern Europe and fewer to the south. Over the Pacific, the dynamic effects similarly drive a meridional dipolar AR trend along the west coast of North America while through different mechanisms. A positive Pacific-North America (PNA)-like pattern is observed in the Z500 trends (Fig. 4c). In this pattern, the anomalous anticyclone over the northwest of Hawaiian Islands produces anomalous southwestward winds that suppress AR genesis near Hawaii, reducing northeastward-propagating PE ARs that typically make landfall in the PNW. Simultaneously, an anomalous anticyclone over northern North America further prevents the AR landfalls over PNW by anomalous easterlies and redirects ARs toward northern British Columbia and Alaska through anomalous northerly steering. Previous studies have linked the northward intensification of mid-latitude precipitation to the poleward shift of the westerly jet stream under global warming32,46,47,48. However, during 1950–2022, the winter Asia-Pacific jet stream only intensified over Asia and the western North Pacific without a discernible poleward shift or intensification (Fig. 4e). Instead, the “wetter north-drier south” meridional dipole patterns of AR and precipitation trends over western North America are modulated by the positive PNA trend.

We acknowledge that reanalysis products become progressively more confident after the advent of routine satellite data in 1979. In the pre‑1979 period, ERA5’s moisture and wind fields over the open ocean depend heavily on the forecast model because of sparse radiosonde coverage. As a sensitivity test, we confirmed that the positive NAO and PNA-like patterns were also present in the historical trend of the Z500 field during 1950–2015 on Twentieth Century Reanalysis version 3 (20CRv3)49 (Supplementary Fig. 7), which is thought to be a trend-preserving reanalysis as it only assimilates observations that are consistently available throughout the analysis period50,51.

Response of boundary forcing versus atmospheric internal variability

Are the observed changes in ARs and precipitation primarily a forced response to historical boundary-condition changes? To answer this question, we first employ a 100-member large ensemble of atmospheric general circulation model (AGCM) simulations from Policy Decision Making for Future Climate Change (d4PDF)52,53, which is the AMIP-type simulation with the largest number of ensembles we can find. Each ensemble member prescribes the same observed evolving boundary conditions (e.g., SST, sea ice concentration, greenhouse gases, and aerosols) but has different initial conditions. The ensemble mean reflects the forced response to historical boundary condition changes and the spread across members implies the influence of atmospheric internal variability.

The ensemble mean reproduces key observed spatial patterns, albeit with weaker magnitudes. For Z500 trends, the ensemble mean captures the positive NAO-like pattern over the Atlantic and the first two lobes of the positive PNA-like pattern over the Pacific (Fig. 5a, b). Similarly, for precipitation, the ensemble mean replicates the observed “wetter north–drier south” meridional dipole over western North America and Europe, as well as precipitation declines near Hawaii (Supplementary Fig. 8). These results imply that observed large-scale circulation changes linked to Northern Hemisphere AR trends are not solely due to atmospheric internal variability. Instead, they are at least partially attributable to historical changes in boundary forcings, especially in shaping the spatial contrast.

Fig. 5: Response of boundary forcing versus atmospheric internal variability.
figure 5

a Historical trend in 500 hPa geopotential height (Z500; shading) and wind vectors (1951–2011; ERA5). b The same as a but for the 100-member d4PDF ensemble mean. c Scatter plot of PNA and NAO trends in d4PDF members and AMIP-hist runs in CMIP6, with observations (red and green stars). Kernel density estimates (grey contours) illustrate scatter distributions. d Histogram of NAO trends in PiControl (orange) and d4pdf and AMIP-hist simulations (blue), with observations (red and green line). e The same as (d) but for PNA trends. f The 7-year-running mean and linear trends of NAO indices (ERA5 EOF-based (red), Hurrell station-based (orange), d4PDF ensemble mean (green)) and AMO index (blue). Solid lines denote trends significant at p-value < 0.001. g The 7-year-running mean and linear trend of PNA indices (EOF-based (red), ERA5 pointwise (orange), d4PDF ensemble mean pointwise (green)) and IPO index (blue). Climate index definitions are detailed in Methods.

Full size image

To further diagnose the relative roles of boundary forcing and atmospheric internal variability, we analyze the ensemble spread of long-term PNA/NAO trends across the 100 d4PDF members. In addition, we assess PNA/NAO trends from other AMIP-type simulations, including 20 members from CESM2 global and tropical AMIP runs and 31 members from six CMIP6 “amip-hist” models (see Methods). The resulting scatter plot (Fig. 5c) shows that both the d4PDF and other AMIP-type simulations cluster consistently toward the “positive PNA–positive NAO” quadrant, suggesting a coherent circulation response to external boundary forcings. Moveover, we generated 100,000 synthetic 61-year segments from a 700-year preindustrial control (PiControl) simulations to model the unforced climate internal variability. Using the Kolmogorov-Smirnov (KS) test, we statistically confirmed that trends in the boundary-forced AMIP-type simulations differ significantly from unforced climate internal variability, especially for NAO trends. In addition, long-term trend of precipitation over four target regions is also examined in d4PDF and their scatter plots also reveal a clustering toward “wetter north–drier south” quadrant over both western North America and Europe (Supplementary Fig. 9a, b). This further implies that the observed trends are very unlikely to have arisen from internal variability alone, underscoring a substantial influence of historical boundary forcing.

The majority of NAO trends in the historical ensemble lie in the positive sector, with distribution significantly distinct from PiControl in the KS test (p-value < 1e−10; Fig. 5d). Over Europe, precipitation trends in d4PDF historical simulations show a significant shift toward wetter conditions in the British Isles and drier conditions in the Iberian Peninsula compared to PiControl (p-value < 1e−5; Supplementary Fig. 9c, d). These results highlight the dominant role of boundary forcing in driving the observed positive NAO trend and the “wetter north–drier south” precipitation pattern over Europe. During the shorter period 1981–2021, the negative-to-positive phase shift of the Atlantic Multidecadal Oscillation (AMO) played the dominant role in the AR frequency increase over the Atlantic sector of Arctic30. However, during 1950–2022, the AMO experienced the positive-negative-positive transition and showed no significant linear trend (Fig. 5f), indicating its limited linkage to the positive NAO trend and AR trends in the mid-latitude North Atlantic. Instead, several studies pointed out that the positive NAO trend in recent decades can be attributed to enhanced warming over the tropical Indian Ocean (TIO), via global atmospheric teleconnections fueled by increased latent heating released during enhanced precipitation over TIO43,44,45,54. Moreover, the reduction of biomass-burning aerosols over the Indian subcontinent since the 1950s plays the primary role in the enhanced TIO warming, remotely contributing to the positive trend of NAO55.

Over the Pacific, the historical ensemble members show a much weaker shift toward positive PNA trends (Fig. 5c). Although these trends still differ statistically from climate internal variability (PiControl), the signal is far less robust in the KS test (p-value \(\approx\) 0.0015; Fig. 5e). Moreover, the positive trend of ensemble mean is insignificant (green line in Fig. 5g). Together, these results suggest that historical changes in boundary conditions have played a minor role in driving PNA trends and associated Pacific AR changes. During 1979–2022, the PNA had a negative trend (opposite of our 1950–2022 result) and ARs were present poleward shifts, which is mainly attributed to the IPO’s positive-to-negative phase transition, with the meridionally broad wedge-shaped SST cooling trend over the eastern tropical Pacific32. However, during 1950–2022, the IPO experienced two phase transitions without significant linear trend, implying its limited role in positive PNA trend and Pacific AR trends (Fig. 5g). Moreover, the SST trend in the Pacific during 1950–2022 differed from a classic IPO pattern, instead showing overall warming over tropical Pacific with a narrow equatorial cooling band (Supplementary Fig. 10). This pattern is more analogous to the Pacific climate change (PCC) pattern, which is determined as a climate change signal emerged over eastern tropical Pacific since mid-1950s by a recent study56. The concurrence of the positive PNA trend and PCC pattern might imply their linkage. However, the extratropical circulation response of the PCC pattern remains unclear so far. The weak PNA positive trend in the ensemble mean and large ensemble spread in d4PDF cannot support the linkage of PCC and PNA, either. It could be due to the influence of the PCC pattern on the extratropical Pacific circulation is marked by the atmospheric internal variation, or model limitations in capturing the response of boundary forcing over the Pacific. Notably, before 1990 the 7-year running-mean PNA index closely tracked the IPO, but after ~1990 the PNA remained positive despite the IPO turning negative, creating the discrepancy in their 1950–2022 trends (Fig. 5g). Whereas the driver of the sustained positive PNA since 1990s and its positive trend during 1950–2022, as well as its related AR changes, remain an open question and requires further investigations.

Discussion

In this study, we investigate the historical trend of winter AR frequency over the Northern Hemisphere mid-latitudes and unveil its underlying mechanisms. We focus on 1950–2022, a period during which dominant interdecadal oscillations (IPO and AMO) underwent two phase changes and thus exerted limited net influence. By decomposing the AR trends into dynamic and thermodynamic effects, we demonstrate that the observed atmospheric moistening associated with global warming may have increased the winter AR frequency by approximately 0.6–0.8% per decade. On the other hand, the changing large-scale circulation could either suppress or enhance the positive thermodynamic effect, shaping the spatial contrast of AR trends. From 1950 to 2022, the meridional dipolar patterns were present in AR trends over both western North America and Europe (increased in the north and decreased in the south). Consequently, northwest Canada and Northern Europe have experienced more AR-induced extreme precipitation, while the PNW and Southern Europe have become drier because of fewer AR landfalls in recent decades. We point out that the “wetter north–drier south” trends are not linked to the poleward shift of the westerly jet stream but instead align with positive trends in the PNA and NAO patterns. Furthermore, the “Pineapple Express” ARs are suppressed by the first anticyclonic lobe of the PNA pattern, exacerbating declines of AR landfalls and precipitation over PNW, underscoring the role of remote dynamical influences. In the future projections of AR and extreme precipitation changes under global warming, both thermodynamic and dynamic drivers should be carefully considered, as their interplay will critically determine regional hydroclimatic risks. The dynamic effect would be the key source of uncertainty. As the climate continues to warm, the relative contributions of atmospheric warming and circulation change to AR trends and precipitation may evolve, and it warrants continued study.

Through the large ensemble AMIP-type simulation analysis, we demonstrate that observed SST changes dominate Atlantic AR trends, possibly via atmospheric teleconnections tied to enhanced Indian Ocean warming. In contrast, Pacific AR trends exhibit weaker boundary-forced signals, indicating the prominence of internal variability or model limitations in capturing Pacific responses. Moreover, the post-1990 decoupling of PNA and IPO (Fig. 5g) raises questions about non-stationary teleconnections under anthropogenic climate change. If the PNA remains phase-locked independent of IPO, as our results suggest, future AR trajectories may diverge from projections relying on IPO–PNA coupling. We did not explicitly attribute AR trends to anthropogenic forcing in this study, because the extent to which the observed SST trend pattern is itself a result of human influence remains an open question56. Notably, the observed multi-decade SST trend exhibits a La Niña-like pattern in the Pacific, whereas the multi-model mean projection shows an El Niño-like pattern. This persistent discrepancy between observed and modeled Pacific warming patterns awaits to be fully understood57,58,59,60.

Methods

Reanalysis and observational dataset

Vertically integrated water vapor transport (IVT) was calculated using specific humidity (\(q\)) and horizontal wind components (\(u\) and \(v\)) at 20 pressure levels (1000–300 hPa) from the European Centre for Medium-Range Weather Forecasts Reanalysis version 5 (ERA5), with 1° × 1° spatial resolution and 6-hourly temporal resolution during 1950–202261. ERA5 is not designed to be trend-reserving, so we validated our findings using the Twentieth Century Reanalysis version 3 (20CRv3)49, which assimilates only observations consistently available throughout the analysis period and is therefore thought to be a trend-preserving reanalysis50,51. Two observational daily precipitation datasets were used. Over Europe, the 0.25° × 0.25° gridded daily precipitation dataset, interpolated from station data, was retrieved from the European Climate Assessment & Dataset (E-OBS v27.0e)35. Over Contiguous U.S. (CONUS), daily precipitation data from the CPC Global Unified Gauge-Based Analysis with 0.5° × 0.5° spatial resolution were employed36. To our knowledge, these are the only large-area gridded observational daily precipitation datasets spanning from 1950 onward.

Atmospheric Model Intercomparison Project (AMIP) simulations

AMIP models driven by observed changes in boundary forcing are useful tools to study how boundary forcing drives the observed trend of the atmosphere. First, we utilized the d4PDF (Policy Decision-Making for Future Climate Change) historical 100-member large ensemble simulations from 1951–201152,53. The simulations were generated by the Meteorological Research Institute atmospheric general circulation model version 3.2 (MRI-AGCM3.2) at 60-km resolution, forced by observed monthly sea surface temperature (SST), sea ice concentration and climatological monthly sea ice thickness. External forcings included global mean greenhouse gas concentrations and three-dimensional ozone/aerosol distributions. The Centennial Observation-Based Estimates of SST version 2 (COBE-SST2) was applied as the SST boundary conditions. The ensemble included 100 members with slightly different initial conditions and the SST perturbations. We also employ the 700-year preindustrial control (PiControl) simulations from MRI to estimate the unforced internal variability of large-scale circulation and precipitation changes.

To further verify the result obtained from d4PDF, we also employed the large ensemble of AMIP-type models that participated in CMIP6 (7 models with 31 members in total), as well as the global AMIP and tropical AMIP Ensemble from CESM2 (GOGA and TOGA; 10 members each). Details on the models’ name, institution, and the number of ensemble members used for each model are described in Table S1.

AR detection algorithm and Global AR database

An AR database over the Northern Hemisphere during 1950–2022 was constructed using the PanLu AR algorithm33,34 with some modifications. The input data is IVT from ERA5 reanalysis data with \(1^\circ \times 1^\circ\) spatial and 6-hourly temporal resolution. We detected preliminary AR pathways using the 85th percentile IVT threshold field with 12-degree Gaussian kernel smoothing and 150 \({kg}\,\cdot\,{m}^{-1}\,\cdot\,{s}^{-1}\) as the lower boundary of the global IVT threshold. Then, we remove (1) AR pathways with length < 1000 km, area < 40,000 \({{km}}^{2}\) or mean IVT intensity < 250 \({kg}\,\cdot\,{m}^{-1}\,\cdot\,{s}^{-1}\) (to eliminate small-scale weak moisture transport features); (2) the tropical cyclone-related high IVT fields with curvatures > 360° (definition of AR curvature is provided in refs. 33,34); (3) tropical moisture transport filaments with >95% of grid cells within the tropical region (20°S–20°N); and (4) trade wind-related moisture transport filaments with >50% of grid cells within the tropical region and >50% of grid cells with absolute values of IVT direction <15° poleward. Two AR pathways were grouped into one AR event if their spatial overlap exceeded 30% (relative to the smaller AR pathway) and their time interval was less than 18 h. If an AR pathway overlaps with more than one AR pathway in the next timestep, the one with the largest overlap area is selected. All AR events identified in this study lasted at least 24 h (including four 6-hourly timesteps). At each grid cell, winter AR frequency was calculated as the percentage of 6-hourly timesteps (December to March) with detected ARs.

The PanLu AR detection algorithm has participated in the atmospheric river tracking method intercomparison project (ARTMIP). The algorithm’s robustness in global AR detection has been demonstrated across multiple datasets in ARTMIP intercomparison studies, including Merra-2 reanalysis62, the tier 2 high-resolution global warming experiment22, and tier 2 paleoclimate experiments63. In addition, we make our AR database publicly available at https://github.com/panmengxin/AR_database, covering 1950–2022 over the Northern Hemisphere.

Decomposition of dynamic and thermodynamic effects on AR trends

Intense moisture transport, the main feature of ARs, is characterized by strong horizontal winds and high atmospheric moisture content. This means that AR variability or long-term trends are primarily governed by thermodynamic effects (moisture content) and dynamic effects (horizontal wind fields). To isolate these contributions to long-term AR trends, a scaling method is employed in this study30,41,42. First, a hypothetical IVT field representing dynamic effects solely was constructed by multiplying the 6-hourly specific humidity field by scaling factors (\(\alpha\)) before vertical integration. Scaling factors are defined as \({q}_{c}/{q}_{s}\), where \({q}_{c}\) is the climatological seasonal mean specific humidity (1950–2022) and \({q}_{s}\) is the seasonal mean specific humidity (varying by year). Scaling factor (\(\alpha\)) is defined for each grid cell, pressure level (300–1000 hPa), and season (3-month moving window). By suppressing year-to-year variability of atmospheric moisture content, the hypothetical IVT field could represent wind-driven (dynamic) moisture transport only. A “dynamic” AR database was constructed using the hypothetical “dynamic” IVT field and thresholds derived from the original IVT field. Then, the dynamic effect on AR trends was estimated as the long-term trend of “dynamic” AR database. A similar scaling method could be used to estimate thermodynamic effects. However, scaling the wind vectors would bring external bias, as the wind would blow in various directions during different years. Additionally, a previous study has found that the two components are largely linearly additive41. Thus, the thermodynamic contribution was estimated as the difference between AR trends in the original and “dynamic” databases.

Definition of Climate indices

We applied two definitions of North Atlantic Oscillation (NAO) in this study. One is the first Empirical Orthogonal Function (EOF) mode of DJFM Z500 anomalies over 20°N–80°N, 90°W–40°E, based on ERA5 reanalysis data. The other one is the station-based Hurrell NAO Index based on 20CRv3 reanalysis data. Based on the Climate Prediction Center definition, we also applied two definitions for the Pacific-North America (PNA) pattern. The first definition is the second mode of a Rotated EOF analysis on Z500 anomalies over 0–90°N64. The second definition is the linear combination of area-averaged Z500 anomaly over four centers65,66, as follows:

$${PNA}={Z}^{* }\left(15^\circ -25^\circ N,\,180^\circ -140^\circ W\right)-\,{Z}^{* }\left(40^\circ -50^\circ N,\,180^\circ -140^\circ W\right)+\,{Z}^{* }\left(45^\circ -60^\circ N,\,125^\circ -105^\circ W\right)-\,{Z}^{* }\left(25^\circ -35^\circ N,\,90^\circ -70^\circ W\right)$$

where Z* denotes the standardized 500-hPa geopotential height anomaly. Tripole Index for the Interdecadal Pacific Oscillation (IPO) is applied, defined as difference between the 10-year low-pass filtered SST anomalies averaged over central equatorial Pacific (10°S–10°N, 170°E–90°W) and the SST anomalies averaged over the northwest (25°N–45°N, 140°E–145°W) and southwest Pacific (50°S–15°S, 150°E–160°W)67. The Atlantic multidecadal oscillation (AMO) is defined as the detrended 10-year low-pass filtered annual mean area-averaged SST anomalies over the North Atlantic basin (0°N–65°N, 80°W–0°)68.

In d4PDF ensemble, EOF analysis produced divergent PNA spatial patterns across members, complicating direct comparison of PNA trends, so we calculated the PNA index using the area-averaged definition. For consistency, the NAO in the d4PDF was similarly defined as the difference of standardized area-averaged Z500 anomalies over southern (30°N–45°N, 60°W–0°) and northern (55°N–70°N, 60°E–0°) North Atlantic. Using the same definitions, we derived 700-year PNA and NAO indices in DJFM for the PiControl simulations. To model climate internal variability, we generated 100,000 synthetic ensemble members by randomly sampling continuous 61-year periods from the PiControl simulations. The long-term trends of PNA, NAO, and precipitation over four target regions were calculated in each synthetic ensemble member. The distribution of trends across these synthetic members reflects climate internal variability. We applied the Kolmogorov-Smirnov (KS) test to determine whether trends in the boundary-forced d4PDF simulations differed significantly from unforced PiControl variability. A significant difference indicates that observed trends are attributable to historical boundary forcing (e.g., SST, greenhouse gases) rather than internal variability.