Introduction

Atmospheric Rivers (ARs) are narrow, filament-like structures that transport large amounts of water vapor across the globe1,2. They often form ahead of the cold fronts of an extratropical cyclonic circulation accompanied by a low-level jet on the equatorward side and strong convection along the front, constituting large-scale cloud systems with persistent precipitation. ARs undergo vigorous upward motion along topographic barriers like mountains or coastlines where they make landfall, leading to torrential downpours. At any given instant, about four or five ARs per hemisphere provide much of the poleward water vapor transport across the midlatitudes3. When time-averaged, over 90% of the total annual poleward water vapor transport in the midlatitudes is controlled by ARs that account for only 10% of the total global surface area4. AR genesis and activity also modulate the global mean atmospheric energy budget via sensible and latent heat transports5.

Understanding ARs and determining their response to greenhouse warming is extremely important as they play a fundamental role in global and regional hydroclimate6. From a global perspective, ARs account for up to 35% of the annual mean precipitation7 and around 50% of the extreme precipitation and surface wind events in the midlatitudes8. Regionally, ARs substantially modulate mean and extreme precipitation along land-falling coastal regions, such as the U.S. west and east coasts9,10,11,12, European west coast13,14,15, South American coastline16, and Australasia summer monsoon regions17,18,19,20. ARs also influence the intensity of snowfall and snowpack variability over many land-falling regions21,22,23,24. The mean hydroclimate over these areas are shaped by ARs in the present as well as past climates25,26,27. More recently, it has been found that ARs influence the ice melting in the polar regions and glaciers, owing to consistently strong precipitation and long-wave feedbacks associated with AR cloud cover. The complex interplay between ARs, clouds, and ice sheets leads to different outcomes in different regions, such as anomalous snow accumulation over East Antarctica28, surface melting over West Antarctica29, reduction of sea-ice in both the Antarctic and the Arctic30,31, and extreme melting of Greenland ice sheet32. Thus, ARs play an important and intricate role in global and regional climate feedbacks.

Recent modeling efforts have advanced our understanding of the dynamics, forced responses, and impacts of ARs. These studies suggest that the frequency of ARs will increase, and their landfall locations will shift poleward in response to anthropogenic greenhouse warming33,34. Additionally, due to increases in atmospheric moisture, ARs are expected to carry more moisture and increase the associated likelihood of extreme precipitation, particularly over regions with elevated terrain6,35,36,37,38. While considerable progress has been made in the simulation, prediction, and projection of ARs using climate models, coarse-resolution models still have large biases in representing the spatio-temporal characteristics of ARs as well as their interaction with topographic features39. These model deficiencies are closely related to biases in the representation of midlatitude storm tracks and jet streams40,41. Although more work needs to be done to definitively conclude whether ARs are better represented in higher resolution models, AR frequencies are suggested to be highly influenced by the horizontal grid resolutions used in climate models42. For example, model simulations with higher horizontal resolutions are expected to represent more realistic precipitation patterns43,44,45 and frontal-scale midlatitude air-sea interactions46,47, thus better capturing the interactions between ARs and complex topography6. Moving forward in understanding AR dynamics and reliable projections, then, requires reconciling findings from both coarse and high-resolution modeling studies.

While developments in high-resolution modeling have much potential to advance our understanding of ARs, current scientific consensus on new findings is elusive as non-negligible differences in present and projected AR features have been attributed to fundamental differences between AR detection algorithms (ARDTs)34,48,49. The majority of pre-existing ARDTs use absolute or relative thresholds based on climate variables, such as integrated water vapor content (IWV) or integrated water vapor transport (IVT) (hereafter thresholds), along with few shape, size and location constraints to identify ARs in the present climate50,51. Studies on reanalysis data sets that use detection methods with absolute thresholds show a high concentration of ARs along the warm western boundary current region in the global oceans and an asymmetric distribution of ARs between the Northern and Southern Hemispheres52,53,54. By contrast, reanalysis studies that use thresholds relative to the mean state, as a function of space and time, show a more symmetric distribution of ARs between the two hemispheres and detect considerably more ARs over land7,26,55,56. Differences in results obtained between these two methods are likely attributable to the uneven spatial distribution of mean water vapor between hemispheres, which is inherently captured differently by absolute and relative thresholds. Even given the relative consistency between reanalysis data sets in representing present climate, as compared to historical model simulations, building consensus on AR characteristics is difficult as the analysis is extremely sensitive to ARDTs49.

The sensitivity of results from AR analysis to ARDTs is even greater when estimating the changes in AR characteristics between various mean climate states, such as between different models and different climate change scenarios. Compared to interpreting AR characteristics in the present climate, analyzing projected AR changes in future climate becomes even more complicated as methods can be based on historical thresholds (i.e., absolute or fixed relative schemes) or based on mean state thresholds (i.e., relative schemes)34. For example, model studies that use methods based on historical thresholds tend to project ARs as much longer and wider, with small changes in mean strength and precipitation in response to greenhouse gas (GHG) warming33,34. Conversely, model studies using relative mean state thresholds tend to project ARs as greatly strengthened in both IVT and precipitation, with smaller changes in their length and width57. The proliferation of the number and type of ARDTs has led to multiplicitous and conflicting interpretations of projected changes, obscuring scientific consensus on AR characteristics in present and future climate.

To simplify the study of present and future AR changes, alternative methods without dependence on climate variable thresholds are being explored by the community, with some studies highlighting the potential benefits of machine learning or image processing techniques58,59,60. To develop a robust ARDT in mind, we use a methodology that reframes the basis of thresholds away from climate variables like IVT or IWV towards shape-based metrics that are not empirically derived from climate data, and thus may be less sensitive to underlying changes in climate mean states (see “Scalable Feature Extraction and Tracking (SCAFET)” in “Methods” section).

Considering the potential of high-resolution modeling to improve our understanding of ARs, and the conflicting AR analysis that appears to arise from IVT and IWV threshold-dependence, this study addresses externally forced changes in ARs and the associated precipitation in response to anthropogenic GHG warming using (1) ultra-high-resolution (UHR) fully coupled simulations, and (2) a novel detection method based on a local geometric shape extraction, namely Scalable Feature Extracting and Tracking (SCAFET)61. We analyze present-day (PD), CO2 doubling (2CO2), and quadrupling (4CO2) simulations from the UHR Community Earth System Model (CESM) version 1.2.2 with about 0.25∘ and 0.1∘ horizontal resolution for atmosphere and ocean, respectively62. See  the “Methods” section for description of the data, model and methods used in this study. ARs within CESM UHR simulations and reanalysis are detected using SCAFET, and are validated against the observed features and the projected changes in AR characteristics from other studies (sections “ARs in the present climate simulation” and “AR responses to greenhouse warming”). Their role in inducing extreme precipitation is then discussed in the following section (see “Extreme precipitation associated with ARs”).

Results

ARs in the present climate simulation

Using SCAFET, we compare the 20-year mean spatial distribution of daily AR frequency in the CESM UHR PD simulation with the observed high-resolution counterpart obtained from the European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis Version 5 (ERA5). The CESM UHR PD simulation well captures the observed global distribution of the climatological AR frequency (Fig. 1a, b). The maximum values of the AR frequency are found along the storm track regions in the midlatitudes (contours in Fig. 1a, b). AR frequencies are also high over the warm western boundary currents and the currents’ extensions throughout the year. These AR structures are associated with the frequent moving-extratropical cyclones (ECs) in that region. The integral role of ARs in regulating midlatitude water cycle is likewise observed in the CESM UHR PD and ERA5 datasets, where ARs account for a considerable portion of the precipitation (1d, e). ARs exhibit an equator to poleward migration during the respective summers for both hemispheres, and a pole to equatorward migration during the respective winters (Supplementary Fig. 1). Not only does the PD simulation analyzed with SCAFET captures the climatological AR frequency and the fraction of AR precipitation as in ERA5, but it is also comparable to the results derived from various other databases using a variety of ARDTs (see Fig. 97, Fig. 149, Supplementary Figs. 2 and 3).

Fig. 1: The mean AR statistics in CESM ultra-high-resolution (UHR) PD simulation compared with ERA5 reanalysis.
figure 1

Fraction of AR days per year (nondimensional) in the CESM UHR PD simulation (a), ERA5 (b), and their difference (c), and the ratio of AR precipitation to annual mean precipitation (nondimensional) in the CESM UHR PD simulation (d), ERA5 (e), and their difference (f). The model climatology is calculated from the daily mean data of the last 20 years of the PD simulation, and the observed climatology is obtained for the 20 years of 2000–2019. The cyan contours in a and b show the column integrated eddy kinetic energy in Mega Joules (MJ).

Full size image

As observed by recent studies, we find that the spatial pattern of mean AR frequency is largely insensitive to the spatial resolution of the simulations used63,64 (see Supplimentary Fig. 4). As the parameters used for AR detection in low resolution (LR) datasets are same as that with the high resolution (HR), this also demonstrates the ability of SCAFET to detect ARs from datasets with varying resolution. However, the HR simulations show a more realistic representation of the precipitation and convection characteristics compared to LR counterparts65. Consistent with these studies, the CESM UHR PD simulation presented here shows more accurate representation of the AR characteristics as compared to analogous LR simulations (Supplementary Fig. 5). The differences between HR and LR AR characteristics are considerable in the extreme attributes. Since extreme precipitation is an important facet of ARs, HR simulations and resolution-insensitive detection algorithms should be used in future studies into the mechanism and the impacts on extreme precipitation.

Whereas the overall patterns in AR frequency between SCAFET and other ARDTs are comparable, SCAFET tends to detect weaker ARs more than most canonical ARDTs and thus detects a higher overall AR frequency (see also Supplementary Figs. 2 and 3). The detection of a large number of weak ARs is expected, as the algorithm does not set a minimum threshold on IVT. AR-like objects are identified from the IVT fields by extracting ridge, and dome-shaped objects within the data. A scale-insensitive, bounded varibale called Shape Index (SI)66 is used to segment the IVT field into various shapes. In essence, regions with SI > 0.375 will be labeled as ARs, given that they are precipitating, large, elongated and have a coherent vapor transport (see "Methods" for further details). Although the algorithm sets an absolute (independent of both space and time) threshold, unlike other ARDTs that sets absolute thresholds (Supplementary Figs. 2a–c), we get a hemispherically uniform distribution of ARs similar to ARDTs based on relative thresholds (Supplementary Figs. 2e–g). Thus, although the algorithm cannot theoretically define an AR, it puts forward a unique perspective in its detection. Excitingly, such a framework can be extended to the detection of other climate and weather features61.

The most notable difference between ERA5 and the CESM UHR PD simulation is the location of the maximum AR frequency and associated precipitation (Fig. 1c, f). This discrepancy could be due to model biases in the location and extent of the jet streams and associated storm tracks in the simulation (Supplementary Fig. 6). Despite minor inter-data and inter-detection-method differences, there is good agreement in AR characteristics between ERA5 and the CESM UHR PD simulation using SCAFET, and little divergence in this analysis from the other ARDTs (see Supplementary Figs. 2 and 3). The choice of the reanalysis datasets will not affect observed biases as detection of ARs is not sensitive to the absolute values of the IVT and is similar to relative threshold based ARDTs63 (Supplementary Fig. 7).

The mean climatology of AR frequency exhibits considerable similarity to the means of other synoptic-scale weather systems such as storm tracks, warm conveyor belts (WCBs), ECs, and atmospheric fronts (AFs). Maximum AR frequencies are manifested on the equatorward side of active storm track regions (Fig. 1a, b contours). Storm track activity is directly proportional to the track density of ECs, and the starting and ending points of the WCBs match well with the genesis and lysis locations of the cyclones67,68,69. Also, the distribution of AFs tends to strongly correlate with the AR activity (see Fig. 270). From a cyclone-centered view, ARs can be viewed as vestiges of a poleward moving extratropical storm71,72, with robust co-occurrences and moderate correlations between AR and storm strength73,74. Connecting all these features, ARs are generally formed along the equatorward side of strong cyclonic circulations, extending from WCB genesis locations along with the warm sector of AFs. Importantly, the SCAFET framework well-captures this dynamic link between ARs and synoptic-scale features.

Fig. 2: Projected changes in AR characteristics, thermodynamic and dynamic components of integrated water vapor transport (IVT).
figure 2

Future changes in the annual fraction of AR days (a, b), AR precipitation to annual mean precipitation ratio (c, d) from 2CO2 (a, c) and 4CO2 (b, d) experiments in comparison to PD. Stippling in a–d denotes significant changes at α = 0.05 level in a two-tailed t-test (n = 20). The vector projection of thermodynamic (e, f), and dynamic (g, h) components onto the mean IVT from PD are shown as shading.

Full size image

AR responses to greenhouse warming

By combining the novelty of both the CESM UHR simulations and the local geometric shape-based detection method, SCAFET, we aim to more reliably quantify the AR responses to GHG forcing. Our results indicate that changes in the 20-year mean of AR frequency under 2CO2 and 4CO2 relative to the PD simulation are highly regional (Fig. 2a, b). Overall, mean AR frequency manifests a robust linear response to increased GHG forcing. Over the jet exit regions in North Pacific, North Atlantic, and South Pacific, there is a statistically significant reduction in the mean AR frequencies, though only for the 4CO2 scenario. The spatial patterns in the Southern Ocean, especially over the Southern Indian and Atlantic basins resemble a poleward shift in the mean AR frequencies. The Indian and East Asian monsoon regions display a slight poleward shift in the mean AR distributions. Large increases in AR frequency are also observed along the near-equatorial regions in North Pacific and North Atlantic, the southwestern part of North America, the northwestern part of Europe, and western Africa. A homogenous upsurge in AR frequencies can also be seen across the Arctic region, with pervasive influence on icesheet and sea ice dynamics across the domain. Other regions like Australia and Central Asia (east of the Caspian Sea) see large-scale reductions in AR activity. The global mean of AR frequency is projected to increase by about 6% and 12% in the 2CO2 and 4CO2 simulations, respectively, relative to the PD simulation. In line with previous studies, our findings indicate a general increase in the average frequency of ARs in response to GHG-induced warming, although notable regional reductions are observed.

Estimated global mean changes in AR frequency are considerably less than those cited in previous studies. For example, one study using historical thresholds projected an increased global mean AR frequency of about 50% by the end of the 21st century in the RCP 8.5 scenario33. Application of the same detection method using historical thresholds for the CESM simulations quantifies a 42% and 61% increase in global mean AR frequency in the 2CO2 and 4CO2 simulations, respectively, exhibiting projections that are more than 5 times larger than the SCAFET-derived estimation (Supplementary Fig. 8a, b). If the same detection method is applied with the mean state thresholds (not shown), the increase in global mean AR frequency is estimated to be 11% and 19% in the 2CO2 and 4CO2 simulations, respectively, which is still considerably larger than the SCAFET-derived estimation of 6% and 12%. The large discrepancy in results between SCAFET and canonical ARDTs may in part be because SCAFET is not affected by the homogenous increase in IVT from global warming, while threshold-based detection methods are fundamentally affected. Utilizing TempestExtremes54 to calculate the responses of ARs to GHG warming shows results that are closer to the changes obtained from SCAFET (Supplementary Fig. 8c, d). This is because, to detect ARs, TempestExtremes applies thresholds on the Laplacian of the IVT field, which by definition is the divergence of the spatial gradient and is thus similarly insensitive to homogenous mean state increases in IVT. By contrast, ARDTs that use historical or present day IVT or IWV thresholds identifies larger AR areas as they do not account for a future mean state that has ubiquitously higher specific humidity34. By increasing the minimum thresholds for IVT or precipitation in the SCAFET, similar patterns of AR frequency changes can be observed compared to conventional ARDTs (Supplementary Fig. 9). Ultimately, the choice of AR detection method depends on the specific research question being addressed.

Reconciling the patterns of projected AR frequency changes between modeling studies is difficult not only because of different threshold-dependent ARDTs but also because AR projections depend on variable thermodynamic and dynamic physical quantities, like specific humidity and wind speed. Modeling studies have shown that a 1 °C increase in global mean temperature corresponds to about a 7% increase in water vapor content and CESM UHR simulations also show that the lower tropospheric mean specific humidity (1000 hPa to 700 hPa) globally increases by about 7.3% and 8% globally per 1 °C global warming in 2CO2 and 4CO2 scenarios, respectively75. The changes in wind magnitude and direction are less certain and are dependent on the regional and model characteristics. A combination of these climatological changes modulates the the spatial pattern of IVT, and by extension the spatial pattern of AR frequency. To separate out the influence of humidity increases and wind changes, we split total IVT changes (ΔIVT) into the thermodynamic (\(\int\overline{{{{{{{{\bf{V}}}}}}}}}{{\Delta }}qdp\)) and dynamic (\(\int\,q{{\Delta }}\overline{{{{{{{{\bf{V}}}}}}}}}dp\)) components, hereafter TIVT and DIVT respectively (see the “IVT decomposition” in “Methods” for complete decomposition of IVT). By separating the thermodynamic and dynamic components, we explicitly show how SCAFET delineates projected AR characteristics from the underlying dynamics and mean state trends that modulate projected changes in AR frequency.

To demonstrate how TIVT and DIVT affect the mean IVT and hence the distribution of AR frequency, we take the vector projection of TIVT and DIVT onto mean IVT from PD (Fig. 2e–h). In linear algebra the vector projection of a vector \(\overrightarrow{a}\) onto another nonzero vector \(\overrightarrow{b}\) is defined as \(a\cdot b/\left\vert b\right\vert\). Positive values of the scalar product indicate that the changes act to reinforce the mean circulation in PD while negative values imply that the changes are in the opposite direction of the mean circulation and result in the weakening of the total transport. ARDTs that are based on historical or relative thresholds estimate AR frequency changes that are sensitive to both components, and thus AR frequency is largely increased following TIVT (Supplementary Fig. 8a, b). However, SCAFET is more responsive to changes in DIVT, similar to other ARDTs that either use a relative or second-derivative threshold (Supplementary Fig. 8c, d). Regions where both TIVT and DIVT act to strengthen the mean circulation show the largest increase in AR frequency, as in the Southern Ocean (Fig. 2e–h). In most regions where the DIVT changes are opposing the TIVT changes, AR frequency often seems to follow the mean DIVT, particularly over the midlatitudes. As the wind response to GHG warming is highly model dependent, mean AR frequencies obtained using SCAFET may vary between models (Supplementary Fig. 10). Examining the zonal and meridional components separately also tells the same story in that AR frequency changes are modulated by DIVT, though the meridional components are inessential in explaining the overall pattern (Supplementary Figs. 11–14). Compared to other ARDTs, AR frequency is more modulated by DIVT in our scheme due to the fact that SCAFET identifies regions with local maxima in IVT regardless of its absolute value. In other words, SCAFET by construct not only identifies regions with high water vapor content but also the low level jet structure associated with it, which is represented by DIVT.

The precipitation changes associated with ARs are consistent with the changes in AR frequency (Fig. 2c, d). Globally, the contribution of AR precipitation to mean annual precipitation increases by 6%(2.4%/°C) and 15%(3%/°C) in 2CO2 and 4CO2 scenarios respectively, which was also seen in previous studies76. The greatest increases in AR precipitation occur over high and midlatitudes (Fig. 2c, d), closely following the projected changes in total precipitation62 (Supplementary Fig. 15). Additionally, we find that ARs are projected to become stronger by 15%(6%/°C) and 34%(7%/°C) in 2CO2 and 4CO2 simulations respectively, consistent with the Clausius-Clapeyron scaling75 (Fig. 3a). The mean precipitation within the ARs increases by 4%(1.6%/°C) and 10%(2%/°C) congruent with the hydrological sensitivity estimated from other coupled climate simulations77 (Fig. 3b) while the maximum precipitation within ARs shows amplification by 12%(5%/°C) and 30%(6%/°C) for 2CO2 and 4CO2 scenarios, respectively (Fig. 3c). These changes in maximum precipitation within ARs are similar to responses observed in other saturated atmospheric environments like tropical cyclones78. Although the changes in mean AR precipitation is dependent on the AR boundary which in turn is highly sensitive to the choice of ARDT, maximum AR IVT and precipitation are largely insensitive to the detection methods used. AR duration shows an extension of 1.6% and 2.3% in 2CO2 and 4CO2 scenarios (Fig. 3d). Moreover, ARs tend to become 1.5% and 3% longer, bigger, and more elongated in 2CO2 and 4CO2 simulations (Fig. 3e–g). However, these geometric parameters are highly sensitive to the ARDT used. For instance, results obtained using historical thresholds show larger increases in length, width, and area with only moderate changes in mean AR IVT and precipitation34. Results obtained using mean state thresholds show moderate changes in geometric parameters and a generous increase in IVT, similar to our results57. Local shape-based methods like SCAFET have the potential to reduce the ambiguity in projected changes of ARs arising from IWV or IVT threshold-based ARDTs while also successfully capturing the changes in important AR characteristics such as mean IVT and the associated precipitation.

Fig. 3: Changes in the distribution of AR properties.
figure 3

Probability density function of AR (a) mean integrated water vapor transport (IVT), (b) mean precipitation (PRECT), (c) maximum precipitation, (d) duration, (e) area, (f) length, and (g) the ratio of length to width for ERA5 (black), PD (blue), 2CO2 (orange), and 4CO2 (red). The vertical lines represent the median value of the distribution. Y-axis for all plots are represented in log scale.

Full size image

Extreme precipitation associated with ARs

Extreme precipitation (XP) is defined as daily precipitation greater than the 95th percentile of daily precipitation within the CESM UHR PD simulation and ERA5 reanalysis, respectively. In other words, the definition of extreme precipitation does not change between present and future scenarios. To examine the link between ARs and extreme precipitation events, we calculate the conditional probability of extreme precipitation given the presence of an AR, or P(XP∣AR). As expected, the presence of an AR leads to an increased probability of extreme precipitation everywhere (Fig. 4a, b). Notably, ERA5 and CESM UHR PD identify the same regions for the risk of AR-induced extreme precipitation, although the CESM UHR PD simulation slightly overestimates the probability of AR-induced extreme precipitation. The increased probability is most visible along the coastlines where ARs make landfall and the winds hit perpendicular to regions with a strong topographic gradient (Supplementary Fig. 16a). For example, there are some regions in which the presence of an AR makes precipitation 10 times more likely, such as over the southwestern coast of the U.S., western Greenland, and around the Antarctic coastline (Fig. 4a, b). Other hotspots for AR-induced extreme precipitation are the Iberian Peninsula, the Scandinavian mountains in Norway, the Andes mountains, the Ural mountains, the Hindukush mountains, western Australia, and western South Africa. Expansion of the subtropical dry belt in response to GHG warming leads to a regional reduction of mean precipitation and of AR-induced extreme precipitation in those regions (Fig. 4c–e, Supplementary Figs. 15 and 16). By contrast, the probability of extreme AR precipitation increases substantially on the poleward side of the subsidence regions. The largest increases are seen along the windward side of mountain ranges perpendicular to the large-scale circulation (Fig. 4c–e). To reiterate, ARs and associated mean and extreme precipitation play an important role in ice, precipitation, and cloud feedbacks, and better quantifying the impacts of AR-induced extreme precipitation events on sea ice, ice sheets and permafrost regions will be of paramount importance for future studies in understanding Earth’s hydroclimatic response to GHG warming.

Fig. 4: Probability of extreme precipitation given an AR.
figure 4

Probability of extreme precipitation (95th percentile of daily mean precipitation) given the presence of an AR, P(XP∣AR), for ERA5 (a) and CESM UHR PD simulation (b), respectively. Changes in the probability of AR-induced extreme precipitation for 2CO2 (c) and 4CO2 (d) simulations. The zonal mean of the probability with lines representing ERA5 reanalysis (blue), present day (orange), 2CO2 (green), and 4CO2 (red) simulations are plotted as e.

Full size image

Conclusions

We combined the novelty of CESM UHR simulations and application of a novel, local geometric shape-based ARDT that is independent of IVT or IWV thresholds to estimate and analyze projected changes in AR characteristics. The results are consistent with previous studies indicating an increase in the mean frequency (by 6% and 12%) and annual mean precipitation associated with ARs (by 6% and 15%) in response to increased atmospheric CO2 concentrations (for 2CO2 and 4CO2 simulations). The mean precipitation increases within ARs matche the mean hydrological sensitivity77 while the maximum precipitation intensity shows similar changes as in other hydrometeorologically saturated conditions78. Owing to the increased atmospheric water vapor content under global warming, we also observe a strengthening of ARs (as measured by mean IVT) by 15%(6%/°C) and 34%(7%/°C) in 2CO2 and 4CO2 simulations respectively, approximately following the Clausius-Clapeyron scaling (7%/°C). This leads to a slight enlargement of ARs in warmer conditions, but, this enlargement is much smaller than what is reported in other studies using ARDTs based on historical thresholds33,34.

Defining ARs from historical thresholds in a warmer and wetter world results in the identification of a much larger area of conglomerated AR events, with lower mean AR precipitation and strength compared to the threshold independent or relative or second-derivative threshold based methods. Furthermore, the application of historical thresholds to future warming scenarios corresponds to a reduction in the number of AR-like structures per time step, as multiple AR-like structures are clumped together in a single conglomerate against the backdrop of scopious, basin-wide increases in mean IVT. By comparison, detection methods based on relative thresholds or local-shape indicate ARs slightly increase in area but with a relatively large enhancement in AR strength and mean precipitation. Additionally, they also project a rise in the number of AR events per instant as AR structures are isolated from the elevated background mean state34. According to our analysis, changes in AR frequency identified using SCAFET are tightly linked to the mean circulation changes in response to increased GHG forcing. Moreover, the UHR simulations are expected to better resolve the synoptic scale circulation patterns compared to coarser resolution simulations. Our results also appear to capture the interaction of AR circulation systems with the landfalling topography. These aspects of UHR simulations help us to more reliably project and explain AR-associated precipitation characteristics. The accumulated, mean, and extreme precipitation associated with ARs increase in response to increased GHG concentrations. Combining estimated AR frequency and precipitation changes may inform future water management policies as consensus builds on projected changes under global warming.

Importantly, the lack of coherence between IVT and IWV threshold-dependent detection methods fundamentally stems from the lack of a quantitative definition for ARs, though this is difficult to systematically apply globally and in perpetuity. To address the ambiguity in defining thresholds for ARs, we develop SCAFET and use the local geometric shape of IVT magnitude to detect strong local maxima and coherence within water vapor transport to identify AR-like conditions. Detection algorithms for cyclones, AFs, and other phenomena may encounter comparable challenges when attempting to estimate future changes in their frequency and the associated precipitation79,80,81. This is because they are also expected to be responsive to variations in the mean state, which presents exciting opportunities for further investigation and improvement. When analyzing synoptic-scale dynamics that overlie different mean states, especially between present and increased CO2 scenarios, using a local shape-based feature extraction algorithm could be useful for building scientific consensus going forward. The ARDT introduced in this study is now expanded into a general feature extraction and tracking method framework allowing us to more objectively quantify (background-state independent) future changes in synoptic-scale phenomena61.

Methods

ERA5 reanalysis

European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis Version 5 (ERA5) data82 is used to validate the capability of the CESM UHR PD simulation in capturing AR characteristics and associated precipitation in the present climate. We use daily mean data of zonal velocity (U), meridional velocity (V), and specific humidity (Q) from 1000 hPa to 300 hPa for calculating integrated water vapor transport (IVT) and daily mean precipitation (PRECT) for 20 years (2000–2019) with a horizontal resolution of 31 km and 37 vertical levels. IVT is a vector quantity given by,

$${{{{{{{\rm{IVTx}}}}}}}}=-\frac{1}{g}\int\nolimits_{1000hPa}^{300hPa}q{{{{{{{\bf{U}}}}}}}}dp$$
(1)
$${{{{{{{\rm{IVTy}}}}}}}}=-\frac{1}{g}\int\nolimits_{1000hPa}^{300hPa}q{{{{{{{\bf{V}}}}}}}}dp$$
(2)
$$\left\vert {{{{{{{\bf{IVT}}}}}}}}\right\vert =\sqrt{{{{{{{{{\rm{IVTx}}}}}}}}}^{2}+{{{{{{{{\rm{IVTy}}}}}}}}}^{2}}$$
(3)

Ultra-high-resolution CESM1.2.2 simulations

This study analyzes fully coupled ultra-high-resolution simulations of the Community Earth System Model83 version 1.2.2 (CESM UHR) with 3 different CO2 concentrations: (1) the present-day value (PD, CO2 = 367 ppm), (2) double CO2 (2CO2, CO2 = 734 ppm) and (3) quadruple CO2 (4CO2, CO2 = 1468 ppm) to represent present climate and enhanced GHG conditions. All three simulations use the Community Atmosphere Model (CAM5)84 with spectral element dynamic core at a horizontal resolution of around 0.25∘ and 30 vertical layers, and the Parallel Ocean Program version 2 (POP2)85 at a horizontal resolution of 0.1∘ and 62 vertical levels as the atmosphere and ocean components, respectively. The land component is the Community Land Model version 4 (CLM4)86, and the ice component is the Community Ice Code version 4 (CICE4)87. The configuration has been successfully used in many previous studies44,62,65,88,89. The PD simulation was run for 140 years, while 2CO2 and 4CO2 simulations were run for 100 years after branching out from PD at year 71. For AR detection, we use U, V, Q, and PRECT (same as ERA5) for the last 20 years of each simulation.

Scalable Feature Extraction and Tracking (SCAFET)

Although there is a wide variety of peer-reviewed ARDTs50, this study uses a novel local geometric shape-based method to identify ARs, with broader applications for extracting synoptic-scale features for different background states61. As discussed in the main text, the aim of the new method is to rethink how the thresholds are used for feature extraction. In contrast with many existing ARDTs, we use local geometric shape of a field to extract AR structures instead of applying IVT or IWV thresholds directly. The method uses Shape Index (SI), a measure of the local shape of a field irrespective of the magnitude or the total curvature to identify features of interest. We expect this method to reduce the sensitivity of the estimated AR response to global warming with regard to the types and values of the thresholds used, as unlike climate variable thresholds, shape-based thresholds are not empirically derived for each unique data set and mean state. The representative vector field used for AR detection is IVT. The major steps in the algorithm are explained below. The result after the execution of each step is illustrated in Supplementary Fig. 16. The detected ARs compare very well with results obtained using existing ARDTs in representing not only the mean distribution (Supplementary Fig. 2) but also in event detection (Supplementary Fig. 18).

  1. 1.

    Scale-space Selection: The first step in the algorithm is to suppress variability in scales much smaller than the feature to be extracted. For atmospheric rivers, a simple, grid-aware Gaussian smoothing (see https://unidata.github.io/MetPy/latest/api/generated/metpy.calc.smooth_gaussian.html) suppresses variability smaller than 1000 kms.

  2. 2.

    Shape Index Extraction: In classical differential geometry, eigenvalues of the Hessian of a field are used to measure local curvature of the surface. Here, we use eigenvalues of the Hessian of IVT magnitude to construct a Shape Index (SI) depicting the local shape of the surface as follows66.

    $$SI=\frac{2}{\pi }ta{n}^{-1}\left[\frac{{k}_{2}+{k}_{2}}{{k}_{2}-{k}_{1}}\right]$$
    (4)

    Where k1 and k2 are the two eigenvalues satisfying k1 ≥ k2 for the Hessian matrix of \(\left\vert {{{{{{{\bf{IVT}}}}}}}}\right\vert\) given by,

    $${{{{{{{\mathcal{H}}}}}}}}\, \left(\left\vert {{{{{{{\bf{IVT}}}}}}}}\right\vert \right)=\left[\begin{array}{ll}\frac{{\partial }^{2}\left\vert {{{{{{{\bf{IVT}}}}}}}}\right\vert }{\partial {x}^{2}}&\frac{{\partial }^{2}\left\vert {{{{{{{\bf{IVT}}}}}}}}\right\vert }{\partial x\partial y}\\ \frac{{\partial }^{2}\left\vert {{{{{{{\bf{IVT}}}}}}}}\right\vert }{\partial y\partial x}&\frac{{\partial }^{2}\left\vert {{{{{{{\bf{IVT}}}}}}}}\right\vert }{\partial {y}^{2}}\end{array}\right]$$
    (5)

    The SI is a unitless, bounded quantity between -1 to 1, independent of the absolute magnitude or resolution of the scalar field. This invariance is crucial as it ensures that the SI focuses solely on the nature of curvature, independent of its magnitude or size. A traditional ARDT may use the absolute magnitude7 of the field or its curvature58 to identify AR like structures while SI would characterize all ridge structures on the \(\left\vert {{{{{{{\bf{IVT}}}}}}}}\right\vert\) as ARs. The basic philosophy of SI can be summarized as “all circles are circles regardless of its radius".The SI provides a single-value representation that captures different shapes: values close to -1 represent concave surfaces, values near +1 indicate convex shapes, and the intermediate range corresponds to transitional shapes like flat or saddle-shaped regions. This intuitive interpretation makes the SI an effective tool for characterizing and analyzing local shapes across a wide range of applications. To obtain AR-like structures, the SI threshold of 0.375 is chosen to extract local shapes such as ridges, caps, and domes. Additionally, points where the IVT deviates by more than 45∘ from the local ridge are also removed to ensure coherence in the transport direction. The local ridge is aligned in the direction of the eigenvector corresponding to k2.

  3. 3.

    Filtering: A variety of geometric and physical characteristics of the objects obtained from the previous step are estimated. Later, weak structures are removed by fixing a minimum mean precipitation of 1mm/day for each object. Other filters are applied to eliminate small (area < 2 × 106km2), short (length < 2000km) and circular (eccentricity < 0.75) structures. AR-like structures associated with the Inter-Tropical Convergence Zone (ITCZ) are removed by masking out objects centered between 20∘S and 20∘N.

  4. 4.

    Tracking: The distance between filtered objects for each adjacent timestep is calculated based on their centroids. The closest objects among them are identified as the progression of one AR. If the distance traveled is greater than 4000km per day, the object is considered to be separate. Objects with a duration of less than two days are removed from the analysis.

IVT decomposition

IVTx is decomposed into mean and anomaly as follows

$${{{{{{{\rm{IVTx}}}}}}}} =-\frac{1}{g}\int\nolimits_{{p}_{s}}^{{p}_{t}}(\bar{q}+{q}^{{\prime} })(\bar{u}+{u}^{{\prime} })dp\\ =-\frac{1}{g}\int\nolimits_{{p}_{s}}^{{p}_{t}}\bar{q}\bar{u}+{q}^{{\prime} }{u}^{{\prime} }+\bar{q}{u}^{{\prime} }+{q}^{{\prime} }\bar{u}dp$$
(6)

Where \(\bar{u}\), and \(\bar{q}\) are the monthly mean zonal velocity and specific humidity respectively and \({u}^{{\prime} }\), and \({q}^{{\prime} }\) are anomalous zonal velocity and specific humidity with respect to monthly means. The pressure levels are integrated from ps = 1000 hPa to pt = 300 hPa. The same pressure bounds are also chosen for the IVT calculation in AR detection.

For a particular mean state, the time averaged IVTx is calculated as,

$$\overline{{{{{{{{\rm{IVTx}}}}}}}}} =-\frac{1}{g}\int\nolimits_{{p}_{s}}^{{p}_{t}}\overline{\bar{q}\bar{u}}+\overline{{q}^{{\prime} }{u}^{{\prime} }}+\overline{\bar{q}{u}^{{\prime} }}+\overline{{q}^{{\prime} }\bar{u}}\ dp\\ =-\frac{1}{g}\int\nolimits_{{p}_{s}}^{{p}_{t}}\bar{q}\bar{u}+\overline{{q}^{{\prime} }{u}^{{\prime} }}$$
(7)

The change in \(\overline{{{{{{{{\rm{IVTx}}}}}}}}}\) between two mean states is calculated as,

$${{\Delta }}\overline{{{{{{{{\rm{IVTx}}}}}}}}}\ ={\overline{{{{{{{{\rm{IVTx}}}}}}}}}}_{2}-{\overline{{{{{{{{\rm{IVTx}}}}}}}}}}_{1}\\ =-\frac{1}{g}\int\nolimits_{{p}_{s}}^{{p}_{t}}\left[{\bar{q}}_{2}{\bar{u}}_{2}+\overline{{q}_{2}^{{\prime} }{u}_{2}^{{\prime} }}-{\bar{q}}_{1}{\bar{u}}_{1}-\overline{{q}_{1}^{{\prime} }{u}_{1}^{{\prime} }}\right]dp$$
(8)

Substituting \({\bar{u}}_{2}={\bar{u}}_{1}+{{\Delta }}\bar{u}\), and \({\bar{q}}_{2}={\bar{q}}_{1}+{{\Delta }}\bar{q}\) in Eq. (8)

We identify the dynamic, thermodynamic, and eddy terms as follows.

$$\begin{array}{ll} {{{{\Delta}}}} \overline{{{{{\mathrm{IVTx}}}}}} &= \overline{{{{{\mathrm{IVTx}}}}}}_{2}-\overline{{{{{\mathrm{IVTx}}}}}}_{1}\\ & =-\frac{1}{g}{\int^{p_t}_{p_s}} \left[({{\bar{q}}}_1+{{{\Delta}}}{{\bar{q}}})({{\bar{u}}}_1+{{{\Delta}}} {{\bar{u}}}) + {{\overline{q^{'}_2u^{'}_2}}} - {{\bar{q}}}_1{{\bar{u}}}_1 + {{\overline{q^{'}_1u_1}}} \right] dp\\ & =-\overbrace{\frac{1}{g}{\int^{p_t}_{p_s}} {{\bar{q}}}_1{{{\Delta}}}{{\bar{u}}}dp}^{\substack{{{{{\mbox{Zonal}}}\, {component}}}\atop {{{{\mbox{of}}}\, {dynamic}\, {term}}}}} - \underbrace{\frac{1}{g}{\int^{p_t}_{p_s}} {{\bar{u}}}_1 {{{\Delta}}} {{\bar{q}}} dp}_{\substack{{{{{\mbox{Zonal}}}\, {component}}}\atop {{{{\mbox{of}}}\, {thermodynamic}\, {term}}}}} -\overbrace{\frac{1}{g}{\int^{p_t}_{p_s}} {{\overline{q^{'}_2u^{'}_2}}}- {{\overline{q^{'}_1u_1}}}}^{\substack{{{{{\mbox{Zonal}}}\, {component}}}\atop {{{{\mbox{of}}}\, {eddy}\, {term}}}}} dp \end{array}$$
(9)

Similarly, IVTy is decomposed into dynamic, thermodynamic, and eddy terms as

$$\begin{array}{ll} {{{\Delta}}} \overline{{{{{\mathrm{IVTy}}}}}} &= \overline{{{{{\mathrm{IVTy}}}}}}_{2}-\overline{{{{{\mathrm{IVTy}}}}}}_{1}\\ & =-\frac{1}{g}{\int}^{p_t}_{p_s} \left[({{\bar{q}}}_1+{{{\Delta}}}{{\bar{q}}})({{\bar{v}}}_1+{{{\Delta}}} {{\bar{v}}}) + {{\overline{q^{'}_2v^{'}_2}}} - {{\bar{q}}}_1{{\bar{v}}}_1 + {{\overline{q^{'}_1v_1}}} \right] dp\\ & =-\overbrace{\frac{1}{g}{\int^{p_t}_{p_s}} {{\bar{q}}}_1{{{\Delta}}}{{\bar{v}}}dp}^{\substack{{{{{\mbox{Meridional}}}\, {component}}}\atop {{{{\mbox{of}}}\, {dynamic}\, {term}}}}} - \underbrace{\frac{1}{g}{\int^{p_t}_{p_s}} {{\bar{v}}}_1 {{{\Delta}}} {{\bar{q}}} dp}_{\substack{{{{{\mbox{Meridional}}}\, {component}}}\atop {{{{\mbox{of}}}\, {thermodynamic}\, {term}}}}} -\overbrace{\frac{1}{g}{\int^{p_t}_{p_s}} {{\overline{q^{'}_2v^{'}_2}}}- {{\overline{q^{'}_1v_1}}}}^{\substack{{{{{\mbox{Meridional}}}\, {component}}}\atop {{{{\mbox{of}}}\, {eddy}\, {term}}}}} dp \end{array}$$
(10)

Noting that for both zonal and meridional components, thermodynamic term > dynamic term ≫ eddy term, we can neglect the eddy terms. Now isolating thermodynamic and dynamic terms, we can write,

The thermodynamic component of IVT changes is formulated as,

$$\begin{array}{ll}{{{{{\bf{TIVT}}}}}}&={{{{{\rm{TIVTx}}}}}}\hat{i}+{{{{{\rm{TIVTy}}}}}}\hat{j}\\ &=-\frac{1}{g}\int\nolimits_{{p}_{s}}^{{p}_{t}}{\bar{u}}_{1}{{\Delta }}\bar{q}dp\hat{i}-\frac{1}{g}\int\nolimits_{{p}_{s}}^{{p}_{t}}{\bar{v}}_{1}{{\Delta }}\bar{q}dp\hat{j}\end{array}$$
(11)

The dynamic component of IVT changes is formulated as,

$$\begin{array}{ll}{{{{{\bf{DIVT}}}}}}&={{{{{\rm{DIVTx}}}}}}\hat{i}+{{{{{\rm{DIVTy}}}}}}\hat{j}\\ &=-\frac{1}{g}\int\nolimits_{{p}_{s}}^{{p}_{t}}{\bar{q}}_{1}{{\Delta }}\bar{u}dp\hat{i}-\frac{1}{g}\int\nolimits_{{p}_{s}}^{{p}_{t}}{\bar{q}}_{1}{{\Delta }}\bar{v}dp\hat{j}\end{array}$$
(12)

Finally, to express the TIVT and DIVT with respect to the PD mean IVT, we calculate the vector projection of each component onto IVT. This approach simplifies our analysis by illustrating the impact of each component on the mean IVT in the PD. Negative values indicate opposing effects on the PD IVT, whereas positive values signify strengthening of the mean IVT in the PD due to the specific vector being examined.

The vector projection of thermodynamic component (TIVT) onto IVT is calculated as,

$$\frac{{{{{{{{\bf{TIVT}}}}}}}}\cdot {{{{{{{\bf{IVT}}}}}}}}}{\left\vert {{{{{{{\bf{IVT}}}}}}}}\right\vert }$$
(13)

Likewise, the vector projection of dynamic component (DIVT) onto IVT is calculated as,

$$\frac{{{{{{{{\bf{DIVT}}}}}}}}\cdot {{{{{{{\bf{IVT}}}}}}}}}{\left\vert {{{{{{{\bf{IVT}}}}}}}}\right\vert }$$
(14)