We extract AR trajectories from two global state-of-the-art catalogs: the PIK Atmospheric River Trajectories version 1.0 (PIKART-1.0) catalog and the Tracking Atmospheric Rivers Globally as Elongated Targets Version 4 (tARget-4) catalog . Both catalogs are based on ERA5 reanalysis data and record 84 years (1940–2023) of AR activity with high spatiotemporal resolution of 0.5° and 6 h. As a general note of caution, we would like to stress that the ERA5 extension prior to 1979 (1940–1979) is known to suffer from reduced data quality, especially before the mid 1950s . The overall structure of the global atmospheric river transport network (ARTN) is qualitatively recovered for the pre-1979 period and for an ARTN derived from the MERRA2-based version of the PIKART-1.0 catalog (Fig. S1). Quantitatively, the pre- and post-1979 ERA5 networks share a Jaccard edge overlap of 0.69 and a Spearman rank correlation of 0.87 computed from edge weights on common edges, while the post-1979 ERA5 and MERRA2 networks agree more modestly (Jaccard 0.50, ρ=0.65).

PIKART-1.0 employs advanced image processing techniques to detect ARs: it identifies ARs as endogenous synoptic-scale IVT anomalies using a top-hat reconstruction technique , without relying on globally fixed magnitude thresholds, and allowing the detection scheme to adapt to regionally and temporally varying atmospheric moisture content. tARget-4 combines relative IVT thresholds with a fixed lower IVT limit. Both algorithms track ARs based on distinct criteria for spatial proximity and morphological similarity of AR shapes, and include physically-meaningful post-processing that filters detections based on geometric and transport-related criteria. The global AR distributions produced by both catalogs have been found to agree on important key regions but yield differing frequencies in regions with anomalous atmospheric moisture content, such as the tropics, mountainous regions, and the poles . Both are included in the Atmospheric River Tracking Method Intercomparison Project (ARTMIP) database .

Atmospheric river transport networks (ARTNs) are a mathematical representation of AR transport patterns that allows visualization, quantification, and modeling approaches. The observational transport patterns that define the ARTN's topology are extracted from a sufficiently large set of spatiotemporal AR trajectories. They consequently adopt a Lagrangian perspective where each AR is tracked along its trajectory. Nodes in the network are given by geographical locations. Edges encode the observed transport patterns. Here, we formally introduce the ARTN framework (Fig. S2).

Let G=(V,E) represent the directed, weighted transport network of ARs, where V is the set of nodes corresponding to spatial grid cells, and E is the set of edges representing AR trajectories traversing between nodes. Each node vi∈V corresponds to a discrete grid cell centered at coordinates lati,loni. The grid covers a spatial domain X (global extent in this study). While the grid could be defined with standard rectangular latitude-longitude grid cells, this would yield significant biases towards the poles due to spherical distortions . Discrete Global Grid Systems offer a more adequate gridding approach for global analyses that minimizes cell distortions . Here, we employ a hexagonal gridding scheme where the Earth is projected to an icosahedron (gnomonic projection) and tiled with hexagons . Hexagonal grid cells entail reduced biases for the network measures computed here (Fig. S3) and offer a set of additional advantages (e.g., intuitive cell neighboring). Every node is thus located at the center coordinates of a hexagon at a discrete spatial resolution that corresponds to average cell areas of 86 801.78 km² (≈2.65 ° at the equator).

We define a transport event as an AR passing through two arbitrary grid cells between two consecutive time steps. To establish this definition, ARs need to be represented by a set of 2D-coordinates, i.e., a locator s=lati,loni. This central constraint allows flexibility in defining different AR locators tailored to specific research objectives, but discards information on the AR's spatial extent. In this work, most ARTNs are computed from the AR centroid, an effective geometrical locator weighted by the IVT magnitude. Alternatively, ARs could be localized by their core (the position of an AR's maximum IVT), yielding an overall consistent network topology (Fig. S4). Locating them by their head (the foremost point across the AR's axis) yields less stable trajectories (Fig. S5) and shifts the geographic distribution of AR network metrics further downstream (Fig. S6).

A directed edge eij∈E connects node vi to node vj if at least one AR transport event occurs between these nodes over the observational time period. This rather permissive edge definition can be scrutinized by thresholding the AR transport matrix W (see below). For each directed edge, we calculate the transport weight wij as the (i,j)th entry of the transport matrix W=[wij]. Here, we choose wij to denote the number of ARs arriving at node vj from node vi (while it could, e.g., also reflect moisture transport or other relevant transport properties). By default, we omit self-links, i.e., wij=0∀i=j. Consequently, W is, by definition, an asymmetric matrix with integer entries that quantify the directed AR transport between each pair of nodes. As an important caveat to the method, some observed displacements of the AR locator result from AR deformation rather than from actual AR advancement. The higher the spatial resolution, the more of such AR movements will contribute to the respective edge weight. Omitting self-links reduces the influence of such deformations. Note that edges can generally exist between any pair of neighboring or non-neighboring nodes. As there are physical constraints to how fast ARs typically move (20–100 km h⁻¹) , most edges, however, connect nodes that are spatially close to each other. If spatial resolution is too coarse (>600 km = 5.41 ° at the equator), only little meaningful trajectory information can be obtained. This can have notable effects on network metrics (Fig. S7).

From the transport matrix W, we construct the transport network G by treating each non-zero element wij>0 as a directed weighted edge in the graph (similar to ). Alternatively, we can generate an undirected, symmetric, weighted graph by discarding edge directions, i.e., by symmetrizing the weights such that w̃ij=w̃ji=wij+wji. A directed, asymmetric, unweighted graph can be obtained by setting all nonzero entries Wij>0 to 1. This is equivalent to thresholding the transport matrix by a transport frequency of ϵ=1. Instead, we can as well choose a more informed threshold ϵ>1, e.g., based on some notion of what is the lowest AR transport frequency that we deem sufficiently high to define a “significant” transport pattern. High threshold values will erode the network and only retain the most frequent AR pathways, potentially discarding regions rarely visited by ARs but where ARs are still a vital control on freshwater supply (Fig. S8b/d). On the other hand, choosing a low (or no) threshold value increases computational complexity and could spuriously overstress rarely traversed pathways, e.g., in the computation of centrality measures (Fig. S8a/c and Sect. ). Some of the network metrics defined in Sect.  (e.g., node strength (Fig. S9a–e)) respond less sensitively to varying the threshold parameter than others (e.g., edge betweenness centrality (Fig. S9f–j)). More sophisticated approaches for network sparsification that identify an optimal value for ϵ can be tested in future studies. The resulting graph is represented as G=(V,E) and referred to as the AR transport network (ARTN).

Beyond the basic transport definition, edges in the ARTN G=(V,E) can be conditioned on additional criteria to tailor the network's construction to specific research objectives. We can define a subset of AR transport events TC⊆T, where T is the set of all AR transport events, and TC satisfies a criterion C. For instance, C can represent a specific temporal constraint (e.g., boreal summer), an AR property threshold (e.g., IVT≥IVTmin), or a spatial constraint (e.g., genesis within a predefined region Rg). Formally, let TC be defined as: 1TC={τ∈T∣C(τ)=True}. Thus, a directed edge eij∈E exists between nodes vi and vj if at least one transport event τ∈TC traverses from vi to vj. The transport weight wijc for this edge is computed analogously to the unconditioned case but restricted to the subset TC. The resulting transport matrix WC=[wijc] defines the refined network GC=(V,EC), where: 2EC={eij∣wijc>0}.

We construct upstream networks as a particular type of a conditioned ARTN by identifying only those segments of AR trajectories that eventually feed into a specified target region. This approach conditions AR pathways on entry into the region, retaining the portion of each trajectory leading up to and including the entry point, as well as the segment within the region.

For any conditional ARTN, thresholds have to be selected carefully as conditioning naturally reduces AR frequencies along each edge. We selected all thresholds such that they yield a comparable number of edges between different phases of a given climate oscillation (e.g., seasonality) and an overall sufficient number of edges for each upstream network.

Overall, approaches following a similar pipeline could be applied to any other synoptic-scale weather system .

The framework developed here seamlessly integrates multiple AR catalogs. The most straightforward approach is the computation of individual ARTNs from individual catalogs and direct comparison of any derived node-, edge-, path-, or community-based property. However, we can also represent the common transport patterns found among a number of K different AR catalogs by a single ARTN. Similar to climate model ensembles, a consensus weight wijcons can be determined for an edge connecting nodes i and j by aggregating edge weights wij(k) across all K AR networks: 3wijcons=1K∑k=1Kwij(k),if wij(k)≥ϵ∀k where wij(k)>0,0,otherwise. with the weight threshold ϵ. This simple approach emphasizes robust transport pathways supported by multiple catalogs by excluding inconsistent connections (see Fig. S8/9). We also tested a majority vote scheme (an edge is preserved if at least 75% of the networks agree on its existence and the edge weight is the median of the latter), as well as a rank-based consensus (within each network, edge weights are replaced by their percentile rank across all edges and an edge is preserved if its mean rank exceeds ϵ). The three considered approaches yielded overall consistent networks (Fig. S10).

We developed four null models to test whether a certain ARTN property (e.g., node strength) is significant for a given node/edge or if it could appear by chance (Fig. ). The key idea is inspired by time series surrogates : we preserve a certain set of properties in the real ARTN and randomize others to control for the effect of the non-randomized properties. With this approach, we pose the question whether a certain observed property that characterizes global AR transport could be obtained from an ensemble of AR trajectories that move randomly without any constraints imposed by large-scale atmospheric circulation, as well as land-atmosphere-feedbacks, ocean-atmosphere coupling, and overall climate dynamics. Each set of simulated random AR trajectories gives rise to a single synthetic, random ARTN. The individual models are explained below: i.

Fully Random Walker (FRW): AR trajectories are generated by performing an ensemble of random walks on a fully connected global network that has the same spatial resolution as the original ARTN. The walk starts at a randomly chosen node, and at each step, the walker moves to a randomly selected connected node (Fig. a). Preserved properties: number of AR trajectories, lifetimes. Null hypothesis: The value of network property X is fully explained by the number and lengths of the observed AR trajectories (and consequently not driven by atmospheric circulation)

Here, we define the measures employed in this study for network analysis. Most of them are covered in greater depth in standard text books on graph theory, e.g., . We stress that these network measures can be regarded as proxies of the underlying atmospheric processes and do not define novel physical quantities. Yet, they can in principle reveal previously unobserved circulation features.

Node strength quantifies the total transport activity associated with a node, incorporating both incoming and outgoing transport. For a node vi∈V, the in-strength siin and out-strength siout are defined as:4siin=∑j≠iwji,siout=∑i≠jwij,where wij represents the weight of the directed edge from node vi to node vj in the transport matrix W. The total strength of a node is the sum of its in- and out-strengths, i.e., si=siin+siout. We stress that node strength is conceptually distinct from a simple histogram, as ARs passing through the node are counted twice, resulting in subtle differences between the two (Fig. S12a/e versus Fig. S12b/f).

We apply the Infomap algorithm to detect hierarchical communities in our directed, asymmetric, weighted transport networks . Infomap is a flow-based method that detects community structures based on random walks on the network. Infomap is well-known for detecting hierarchical communities across scales. In its core, it relies on the map equation: 8L(M)=H(Q)+∑iH(Mi) where:

is the total probability of exiting any module.

Previous research has unambiguously identified the mid-latitudinal storm tracks as the regions of highest global AR activity . Here, ARs typically emerge from the large western ocean basins, propagate eastward and terminate along main coastlines. As a proof-of-concept, we derive node strengths (the sum of all edge weights entering or leaving a node) and test whether this pattern is recovered. Node strength reliably tracks the global regions of highest transport activity (Fig. b). Applying all developed null models to node strength shows that highly active nodes in the western ocean basins mark AR hotspots due to their role as AR genesis hubs (N. Atlantic, N. Pacific, S. Atlantic, S. Pacific, S. Indian Ocean), propagate eastward across hubs of strongly enhanced AR activity that pass all tests and terminate with their centroids typically located just off the coast. The core regions of the storm tracks pass all tests, suggesting AR activity here goes beyond AR genesis, termination and large-scale transport. Superimposed on the west-to-east gradient, a northward-propagation pattern tracks AR transport towards the poles (Fig. b).

Node divergence – the difference of incoming and outgoing edges for a given node – clearly confirms this large-scale transport pattern (Fig. c). Negative (positive) node divergence suggests a node acts predominately as an AR sink (source). Extending past familiar patterns, node divergence captures more complex genesis and termination hubs. In some regions with large continental lakes or seas (eastern Eurasia, eastern North America), intense moisture recycling (the Amazon basin, the La Plata basin) or strong coastal moisture gradients (North Africa/Mediterranean), nodes with significantly high negative and high positive divergence coexist in close spatial proximity, implying more intricate AR transport in and out of these regions.

If common AR pathways are traced repeatedly through the network, which nodes emerge as their most frequent long-term destinations (Fig. c)? By identifying the stationary distribution of an ensemble of random walkers that follow the ARTN, PageRank highlights regions that AR pathways repeatedly converge upon, effectively revealing the network’s structural “attractors”/topological sinks (beyond the basic notion of “high AR activity” provided by node strength). Most nodes exhibit higher PageRank scores than expected from random dynamics (blue pixels), with some of the highest values at known AR landfall hotspots like western North America. The Arctic and Antarctica show highly significant scores, stressing that intense, persistent ARs can reach these highly susceptible regions from many parts of the network with potentially severe hazards and impacts .

Elevated PageRank scores also emerge along major orographic barriers (southwestern Himalaya, southern Andes, northern Rocky Mountains), across monsoon regions such as Southeast Asia, and within key moisture recycling hubs including the Amazon basin and eastern boreal forests. Several of these regions also show significantly high (cycle) clustering coefficients (Fig. S17, e.g. the Amazon basin and the Himalaya), suggesting spatially confined, recurrent AR transport that translates to strongly interconnected regions in the ARTN. This is compatible with PageRank scores as a proxy for nodes at which ARs tend to reside for longer time periods, partly within isolated network components. The shifts of the identified AR hubs when using AR head coordinates are shown in Fig. S12.

In tropical monsoon areas, AR detection remains uncertain and formation mechanisms poorly understood. The high PageRank scores here may partially reflect spurious AR detection in the PIKART-1.0 catalog (especially along the tropical Pacific) but could also support the idea of tropical ARs as a “continuum of extratropical and monsoonal moisture plumes” with high potential to generate extreme precipitation . Future studies could scrutinize AR dynamics in these regions to resolve this uncertainty.

Where do ARs tend to strengthen or weaken their vapor transport? We derive net IVT changes across each node in the ARTN (Fig. a). AR intensification is not limited to ocean basins: Eastern North America, the La Plata basin, eastern Asia as well as parts of Australia can be clearly identified with continental nodes of enhanced IVT intensification. On the other hand, eastern Eurasia and the Amazon basin show less clear tendencies of IVT change and IVT into North Africa is overall balanced. Some nodes across which ARs clearly tend to intensify their vapor transport align strikingly well with well-known low-level jets (LLJs), e.g., the Great Plains LLJ, the Caribbean LLJ, and the southward nocturnal LLJ extension of the South American LLJ . The Western Hemisphere Warm Pool represents one of the main NH evaporative regions and has been previously identified as a global hub of AR moisture uptake , but Fig. a additionally reveals that the LLJs can be distinguished from a background of overall high moisture uptake.

Another region of markedly enhanced IVT gains emerges at the southern tip of South Africa, likely driven by moisture input from South America that drives cold-season ARs over western South Africa . The most prominent IVT depletion hubs (western North America, southwestern South America, the British Isles) align well with AR transport sinks (Fig. b/c) and correspond to regions with a strong dynamic IVT component, where high IVT is closely tied to heavy precipitation .

If individual ARs pass across multiple nodes that tend to intensify IVT, this could imply higher IVT at landfall (and potentially stronger AR hazards and impacts). We test the presence of such correlations and the implied degree of predictability by examining the relationship between the mean IVT classifier over each AR's life cycle sequence and its IVT at first landfall. ARs in the PIKART-1.0 (tARget-4) catalog exhibit a significant Pearson correlation of 0.39 (p<0.01) (0.22 (p<0.01)), respectively (Fig. S18). Given the substantial complexities involved in predicting the intensity of an AR at landfall, this significant relationship is encouraging.

We unravel where the identified link is most strongly expressed by performing a robust linear Huber regression between both variables and extracting the residuals of the regression. Each residual is divided by the respective overall correlation value noted above so that the highest predictability is scaled by the overall correlation (Fig. b, averaged over all AR trajectories and both catalogs at each landfall location). Coastal mid-latitudinal regions that are frequently exposed to high intensity ARs (e.g., California (marker size)) exhibit relatively high predictability scores. Predicting landfall IVT in polar regions and in the (sub)tropics appears more challenging. The PIKART catalog allows ARs to make landfall at non-coastal grid cells where predictability is overall lower. These results, that are based on a simple linear model, point towards the predictive potential of the developed approach, which could be exploited in future studies.

Edges in the ARTN encode crucial information about prominent AR pathways. As the emergent “roads” of global AR transport, they hold significant potential for improving forecasts of AR trajectories. We investigate where the most vital edges are located by computing the edge betweenness centrality (EBC) of each edge, i.e., which edges act as critical conduits of AR transport. Edges with low AR frequency are considered more “costly”. Thus, EBC is expected to carve out a set of edges that many ARs must cross. An edge with high EBC connects many pairs of nodes along their most likely AR routes. If that connection was disrupted, many paths that potentially link important but spatially distant regions would be blocked. We find a system of cohesive “AR highways”, along which ARs orient their flow, show clear tendencies of IVT gains/losses and which are modulated by large-scale atmospheric circulation (Fig. ).

The observed AR highways form a near-continuous belt encircling the globe in both hemispheres. All major branches are statistically significant (95%-confidence level, Fig. S19). The NH branch of the global AR highway pierces through northwestern North America with high EBC and spans across the North America continent, primarily supported by strong EBC edges in the tARget-4 network (Fig. a). Both catalogs also reveal a less prominent but notable branch towards California. The highway then veers towards northwestern Europe and splits into smaller branches reaching northern Africa, central Europe, the Arctic, and the Mediterranean. The main branch passes through the British Isles and northern Europe, continues across eastern Eurasia, and ultimately arrives in East Asia, where it merges with a fork-shaped (sub)tropical branch (identified only in the PIKART-1.0 network). After merging, it forms the well-known trans-Pacific AR pathway to the western United States. PIKART-1.0 also detects an AR highway across the central Pacific, linking Southeast Asia to the Caribbean. In the SH, both catalogs consistently identify a major AR highway extending from eastern South America across the Southern Ocean. Weaker offshoots connect this main SH highway to interior South America, South Africa, and South Australia. Some of these are not statistically significant (95%-confidence level, Fig. S19), likely owing to the fact that they act as “one-way roads”. They can, however, still be considered important entry points from major AR genesis regions and some of them (e.g., the one exiting interior South America) exhibit high AR frequency. EBC reacts more sensitively to changes in the network threshold or spatial resolution than, e.g., node strength (Figs. S7/S9).

The network of high-EBC edges can be regarded as a planetary-scale transport belt that guides ARs through the underlying ARTN. In this context, guiding refers to the dynamical constraint imposed by large-scale upper-tropospheric circulation patterns on the preferred pathways of AR propagation. Specifically, the location and spatial extent of this high-EBC backbone align well with the spatial structure associated with the circumglobal teleconnection pattern, whose dynamical expression is a quasi-stationary Rossby wave train in the upper troposphere . Generally, low-frequency upper-level atmospheric circulation modulates Rossby wave-breaking events which, in turn, control low-level weather systems such as (anti)cyclones and ARs . Wave breaking events along the circumglobal teleconnection pattern amplify upper-level anticyclonic anomalies and create persistent jet streaks that steer and intensify AR pathways along preferred longitudes . The resulting waveguide enhances the zonal coherence and longitudinal continuity of AR transport, supporting the emergence of a contiguous high-EBC AR backbone. This mechanism elucidates how ARs are efficiently navigating through the troposphere, e.g., from the North Atlantic across Eurasia and the North Pacific back to North America during boreal winter. While some previous work has conducted regional analyses of how Rossby wave breaking along the mid-latitudinal jets control AR propagation , our results provide the first evidence of a contiguous AR transport backbone along which these processes organize globally.

Along all AR highways, we derive net IVT gains and losses that can be induced by changes in moisture or winds (Fig. b). As expected, coastlines and mountainous regions deplete ARs of their IVT, likely through orographic precipitation (e.g., western North America, the British Isles, and Chile). Over the western ocean basins identified as AR sources (Fig. b), cyclogenesis plays a major role and ARs tend to fuel up their IVT (either through enhanced evaporation or strengthened winds or a combination of both) . Beyond oceanic sources, several continental regions stand out as key conduits along which ARs gain IVT (eastern North America, Central and East Asia) and some oceanic regions are associated with IVT losses (the North Atlantic branch towards Europe, segments of the southern Ocean branch).

Finally, the SH branch of the AR highways follows a wave-like pattern of IVT gain and loss (Fig. b, also prominent in Fig. a), reminiscent of the zonal wave-3 pattern – a dominant SH circulation mode that produces alternating ridges and troughs in geopotential height . In geostrophically balanced flow, horizontal pressure gradients primarily organize the large-scale waveguide along which ARs propagate, while ageostrophic circulations associated with ridges and troughs induce convergence and ascent, promoting moisture conversion to precipitation . These processes modulate AR pathways and moisture content along preferred longitudes, offering a physically consistent explanation for the observed pattern.

The notion of “AR highways” prompts the question of how closely individual ARs actually follow them. We classify ARs into three groups based on how closely their trajectories align with major AR highways: Conformists, Straddlers, and Strays (examples in Fig. S20a–c). Conformists follow highways most rigidly, while Straddlers deviate more often and Strays exhibit marked anomalies, sometimes venturing far off the main branches. Statistical tests confirm that Strays differ significantly in intensity (but not in lifetime), tending to transport more moisture, while Straddlers exhibit enhanced lifetimes (Fig. S20d/e).

To further quantify the tendency of ARs to propagate along the discovered highways, we compute (i) the fraction of AR trajectories (≥24 h) that traverse at least n highway segments during their lifetime, and (ii) the fraction of k-sequences, i.e., trajectories that follow a highway (10% of edges with highest EBC) for at least k consecutive steps (Fig. c). Both fractions substantially exceed random expectations (99% significance from 200 random-walk ensembles with the same number and lengths as in the PIKART catalog). Hence, highways increase the predictability of AR transport: ARs are drawn toward high-EBC edges, and once aligned, tend to remain on them.

While AR seasonality has been extensively studied in North America and Western Europe , research in other regions is only emerging . We delineate AR highways during December-January-February (DJF, boreal winter) from those in June–July–August (JJA, boreal summer) by conditioning the underlying ARTNs (Fig. d–f, conditioned ARTNs shown in Fig. S21). The band of highest AR activity towards western North America closely tracks the mean position of the eddy-driven jet across seasons, with a gradual migration from its most poleward position in JJA to its most equatorward position in DJF . AR transport patterns into Europe are more complex, but overall, previous studies showed that weakened storm tracks guide less ARs towards the European West coast in JJA than in DJF when some ARDTs record peak AR frequencies . We identify an overall similar backbone of high EBC edges in both seasons with a slightly less pronounced branch towards southern Europe during JJA (Fig. d).

In the SH, seasonal AR highways are not as sharply distinguishable but organise along a more continuous band of high EBC edges. During austral winter (JJA), ARs reach further north along the Chilean coastline and deplete significant amounts of IVT. Across the Southern Ocean, the identified AR highways do not differ considerably between both seasons apart from stronger branching towards southern Australia in JJA.

To showcase that the AR highways revealed here are modulated by climate teleconnections, we condition the ARTN on El Niño/La Niña episodes of the El Niño Southern Oscillation (ENSO) (Fig. g–i, conditioned ARTNs shown in Fig. S21). We classify El Niño (La Niña) episodes by detecting episodes for which the Oceanic Niño index exceeds (falls below) a value of 0.5 (-0.5) for at least five consecutive 3-month averages (Fig. S22). Previous studies found conflicting results on the role ENSO plays in reshaping AR pathways towards western North America, with some finding a significant equatorward (poleward) shift during El Niño (La Niña) episodes while others found other teleconnections to be more relevant, e.g., the Madden-Julian Oscillation or the Pacific-Japan teleconnection . We find that – at hemispheric and planetary scale – ENSO does substantially reshape AR highways. Towards northwestern North America, La Niña generates a distinctive branch (Fig. g). Across the North Atlantic, La Niña continues to direct ARs further poleward than El Niño with various smaller branches reaching into the Arctic. Across interior Europe, ARs tend to stay higher north, consistent with a poleward displaced jet. Conversely, El Niño channels AR highways further south with a prominent branch that crosses over the Mediterranean towards the Middle East and northern India before it ascends again towards Japan. In the (sub)tropics, La Niña generates more prominent AR highways that originate in the central-eastern Pacific and exhibit both an ascending branch towards western North America and a descending branch towards South America (similar to the findings in ). Across the Southern Ocean, no major differences between the dominant branches can be identified. In most regions, IVT gains and losses do not differ substantially along La Niña and El Niño AR highways (Fig. h/i). However, we do find stronger tendencies of ARs to weaken their IVT during La Niña at some of the main AR landfall hotspots, including western North America, Europe, and Chile. Overall, the regions where ARs exhibit the strongest tendencies of IVT change are relatively stable across the studied seasons and ENSO regimes with the most considerable deviations in the (sub)tropics (Fig. S23).

Most AR studies have a regional focus. Hence, our understanding of global AR basins – regions where individual AR trajectories tend to originate, evolve, and dissipate within the same broader area and AR activity is tightly interconnected – remains incomplete. Previous work has applied clustering algorithms to AR conditions (rather than trajectories) within a few confined regions . These regional, Eulerian methods overlook the connectivity of individual AR trajectories and can introduce spurious boundary effects by cutting across AR tracks that enter or exit the region.

Here, we instead detect AR basins as communities in the global ARTN from the more natural Lagrangian perspective using the Infomap algorithm . Physically, AR communities can be understood as enclosed geographical regions whose boundaries are determined by persistent steering flows, topography, coastal moisture gradients and thermodynamic limits on AR life cycles (depending on the rates of evaporation and precipitation over an AR's life cycle). We evaluate the performance and stability of the real AR communities against those obtained from the developed null models (Figs. S24–S26). Given the stochastic nature of Infomap, we assess its robustness by running the single-level mode for 100 different random seeds. We evaluate its overall stability through Adjusted Mutual Information and Adjusted Rand Scores and compare it to 95%-confidence levels obtained from the different null models. For each node, we furthermore quantify heterogeneity/label inconsistency as one minus the average Jaccard similarity between the sets of nodes it is grouped with across runs, capturing the consistency of its community assignments (Figs. S24/S25). We derive a setwise consensus partition by assigning each node to the most frequently co-occurring community set across all seeds. This strongly fragments community assignments for less stable communities (Fig. S24c). More sophisticated consensus clustering approaches would likely lead to a more sound consensus partition but are computationally intense. Instead, we consider the most central partition as the one whose community structure exhibits the lowest overall node-wise heterogeneity, weighted by community size (Figs. a, S24a). AR basins are not fully independent. For any pair of neighbouring basins, a certain degree of inter-basin traffic can be expected.

We uncover that global AR transport is organized across dozens of basins of various sizes and shapes (Fig. a). In the NH, the trans-Pacific and trans-Atlantic basins are most well-separated and stable (marked by high flow ratios and strong partition stability, see Fig. S24b–d). They closely match the stability of the GCW/TCW null models (Fig. S26), attesting high confidence: by definition, the GCW/TCW null models achieve high stability through their constrained AR genesis/termination regions which form strong communities in an otherwise random network.

The trans-Pacific basin splits near 175° E, suggesting that some ARs terminate early en route to western North America (in agreement with ). The northwest American coastline is intersected by five AR basins: two small basins over California and the Pacific Northwest (both likely enhanced during DJF, see Fig. e), a northern extension into Canada and two high-latitude basins that intersect with Alaska. The lower-latitude basins align well with previous classifications . The northern extension into Canada aligns with the entry point of inland-penetrating ARs responsible for triggering extreme precipitation over the continental interior . Over interior North America, community boundaries track topography, particularly the Rocky Mountains and the Appalachians.

Towards Europe, a large AR basin is bordered by four basins along the western coast: two southern basins via Iberia and the Mediterranean, a central basin covering large stretches of coastal and interior central Europe, and a northern branch reaching across Arctic coastlines. Interior Europe hosts additional AR basins, some of which transition towards Central Asia, indicating that various ARs penetrate far into the Eurasian continent. Similar to North America, mountainous regions such as the Ural Mountains and the Alps manifest as AR basin boundaries. AR communities over the Middle East and Northern Africa are identified less stably (Fig. S24b–d), indicating a stronger degree of fragmentation of AR tracks across these regions. While the larger hemispheric AR routes are well established, many of these basins have not previously been systematically identified as the strongly interconnected systems revealed here.

In the (sub)tropics, we find that few AR basins exceed 15 grid cells, highlighting fragmented AR motion. Notable exceptions include zonal basins over the Central Pacific and an isolated Caribbean basin. Conversely, SH AR basins are spatially extensive and more elongated. We identify the most robust and well-separated communities within a South Pacific basin that is directed towards South America and two basins over the Southern Ocean (Fig. S24b–d). The GCW replicates these robust basins, albeit more spatially confined around AR genesis hotspots (Fig. S25d–f). However, higher uncertainty remains over regions such as southern South America, Oceania, and Antarctica (Fig. S24b/d).

AR lifecycle characteristics differ considerably between the identified basins (Fig. S27 and Table S1). Mid-latitudinal oceanic basins (e.g., basin #7 and #3 in the NH or basin #1 and #2 in the SH) host intense, long ARs (IVT ≈490–550 kg m⁻¹ s⁻¹, length ≈3700–4000 km) with relatively long residence times (40–60 h; partly confounded by their large spatial extent). Baja California (basin #40, 49.7%), the southern Iberian Peninsula (basin #34, 44.8%), and the Mediterranean coast (basin #33, 41.5%) exhibit some of the highest AR inland penetration frequencies. ARs also appear to penetrate well into Alaska (basin #38, 35.9%) and northern Scandinavia (basin #26, 35.3%). Basin #11 (eastern China and Japan) receives relatively short but very intense ARs (IVT = 584 kg m⁻¹ s⁻¹, residence ≈27 h). Basin #8 (Central Asia) is entirely land-locked (97.5% land) with relatively high residence times (≈31 h) and moderate IVT (289 kg m⁻¹ s⁻¹), indicating potential moisture recycling. While not exhaustive, this analysis showcases how the identified basins provide a coherent spatial framework for stratifying AR transport patterns, exposing regional contrasts in moisture transport, persistence, and land interactions that hemispheric or storm-track-based views would obscure.

Finally, the identified AR basins are reorganized seasonally with pronounced hemispheric asymmetries (Fig. S27). The North Pacific and North Atlantic basins extend their reach into their respective continental interior during boreal winter and some NH basins migrate poleward in boreal summer (Fig. S28a/c). In contrast, SH basins preserve their zonal continuity year-round, reflecting the persistent austral westerlies. A more intricate cluster of small Mediterranean/Middle Eastern basins emerges during SON and MAM (Fig. S28b/d), in line with a previous study that identified MAM as the dominant season for intense ARs over the Middle East .

To trace moisture pathways, we construct upstream networks by tracking ARs backward from target basins (Fig. b/c). For instance, some ARs strengthen their IVT over western Canada, possibly fueled by the respective LLJ or the Great Lakes (Fig. b, top panel). A population of ARs that traverses Hudson Bay or western Greenland originates from the trans-Pacific branch, losing significant amounts of moisture on their passage across North America (Fig. b, middle panel). ARs that penetrate into continental Europe across the Atlantic branch tend to start reducing their IVT once their centroid has crossed ≈ 35 ° W or keep gaining IVT along a more southerly route across northern Africa/the Mediterranean (Fig. b, bottom panel). IVT losses could initially be induced by ascent in an extratropical cyclone warm conveyor belt or a reduction of zonal winds, followed by orographic uplift.

In the SH, a spatially extensive Southern Ocean AR basin is fed by three distinct AR populations (Fig. c, top panel): ARs that surpass Chile with significant IVT losses, ARs that originate from interior South America over the La Plata basin and ARs originating from South Africa. Southern Australia draws ARs with IVT gains from a northerly route and with more IVT losses from a southerly route (Fig. c, center panel). ARs that may make landfall over eastern Antarctica are strongly diverted from their zonal movement, incurring large IVT losses (Fig. c, bottom panel).

With these examples, we clearly demonstrate how the ARTN not only delineates previously unrecognized basins of interconnected AR transport but also characterizes water vapor flows along the network. The ARTN framework thus complements computationally intense parcel-tracking approaches by offering a lightweight yet physically meaningful view into global moisture transport.

According to the prevailing paradigm, there are five main basins of global AR transport: the North Pacific, North Atlantic, South Indian Ocean, South Pacific, and South Atlantic, each serving as a major corridor for long-range moisture delivery . Figure  suggests the existence of a significantly higher number of AR (sub-)basins. We investigate how these are embedded within larger basins using hierarchical Infomap community detection (Fig. ). At the 0th level, the algorithm simply distinguishes both hemispheres (assigning the tropics to the NH, not shown). At the first level, the (sub)tropics are separated from both extratropical hemispheres (Fig. a). With every additional level, these large-scale communities are broken down into more fine-grained communities. Their intra- and inter-community flow (i.e., AR transport within and between communities) is shown in Fig. e–h and the hierarchical branching is visualized through a Sankey diagram (Fig. i).

At level-2, the NH AR community fragments into an Atlantic and a Pacific domain. Some tropical communities disappear as they are fragmented into sub-communities that are smaller than 15 grid cells (grey). The SH remains stable, underscoring its strong internal connectivity. The community flow is still mostly concentrated as internal flow in the two extratropical hemispheres, similar to level-1 (Fig. f).

Only at level-3, the flow homogenizes more equally across the globe (Fig. g) as communities become even more fine-grained, especially in the NH (Fig. c). In contrast, the SH only breaks down into less, larger communities. This asymmetry likely arises from how the hierarchical Infomap algorithm naturally splits communities in a physically meaningful way: communities are split only if doing so significantly reduces the map equation, i.e., if sub-communities offer a better compression of random walk dynamics. In the SH, AR pathways tend to follow long, coherent, and relatively uninterrupted tracks across the Southern Ocean (Fig. ). These transport patterns translate to high internal connectivity (indicated by high internal flow, Fig. g) and less strongly defined substructures. Conversely, the NH exhibits more frequent AR branching and termination due to larger continental landmasses which entail numerous orographic barriers as well as thermal gradients at coasts, which in turn promotes more frequent and meaningful community subdivisions at higher levels.

At the final hierarchical level, the SH breaks up into various communities (Fig. d). A few of these align well with the AR basins identified in Fig. a. For instance, the large-scale AR basin directed towards the Chilean coast is sub-divided into multiple sub-basins, including a southern and two northern basins in Fig. d with strong inter-community flow. While AR flow has strongly homogenized (i.e., no single community contains the large majority of AR tracks, Fig. h), the most dominant communities are still the ones in the Southern Ocean (e.g., the purple and the olive basin). These two communities contain some of the globally most active AR hubs (Fig. b). In the NH, fragmentation has rendered many communities smaller than the synoptic-scale whereby the remaining ones form a continuous conduit that aligns with the AR highway (Fig. ), underscoring the robustness of the main moisture transport route.

Figure j–m show that, at every hierarchical level, most of the top-20 communities (by how much AR transport is contained) are significantly better separated than expected from the FRW model (95% confidence). Only some smaller communities fall below this threshold. The discrepancies between Figs.  and likely reflect differences in flow magnitude: many small communities in Fig.  lie within the range expected at random, even though the median flow ratio exceeds the 95% quantile (Fig. S24). Thus, some smaller AR basins may be less strongly connected than the main global ones. By contrast, larger, high-flow basins – such as southern California, western Greenland, the Middle East, southern Australia, and western Antarctica – emerge as robust, understudied regions warranting further attention.