Atmospheric rivers (ARs) play a crucial role in the global hydrological cycle, and AR-driven precipitation can cause costly floods or alleviate droughts. Because there is no first-principles definition of ARs grounded in geophysical fluid mechanics, AR identification is currently performed by a large array of expert-defined, threshold-based algorithms. The AR Tracking Method Intercomparison Project (ARTMIP) has revealed uncertainty introduced by the choice of detection algorithm (O’Brien et al. 2022).

In prior work, we constrained AR detection algorithms to satisfy a key physical property: in boreal winter, ARs dominate the meridional transport of latent energy. At each latitude, we select a subset of the ARTMIP ensemble such that the identified ARs account for all the transient poleward latent heat transport. Here, we develop AR detection catalogs that satisfy this physics-based constraint. Following Lora et al. 2020, we assess consensus and disagreement among individual AR detection algorithms and these catalogs. We contextualize the constrained set of AR projections in terms of the broader spread of ARTMIP algorithms.

Using neural networks, we can also generate AR detections with this physics-based constraint in datasets without an associated ARTMIP labeling campaign. We leverage this method to assess AR behavior in a long-term historical dataset (ERA-20th Century Reanalysis, ERA20CR) and in high-resolution future projections (the High-Resolution Model Intercomparison Project, HighResMIP). Given the physical constraint, we present a comprehensive assessment of ARs’ future projections and historical trends, in CMIP6, ERA20CR, and HighResMIP. We explore changes in future AR behavior induced by different climate model resolutions and future emissions pathways. We assess how future changes in poleward energy transport will affect AR-induced droughts and floods.