Weather-climate extremes emerge from interactions across scales from planetary climate variability to synoptic weather systems, mesoscale convection, and local processes. The Madden–Julian Oscillation (MJO) and El Niño–Southern Oscillation (ENSO), the dominant modes of tropical intraseasonal and interannual variability, interact with and influence atmospheric rivers, tropical easterly waves, and tropical cyclones, creating multiscale pathways to extreme precipitation, flooding, and unusually active or inactive hurricane seasons. Representing these interactions and translating them into skillful predictions of rare, localized extremes remain major challenges for Earth system models.

Rapid advances in AI emulators offer new opportunities, but their ability to represent the physical processes and multiscale interactions underlying extremes must be rigorously evaluated using observations and process-based models. We introduce the Multiscale Object Tracking System and AI Capabilities for Extremes (Mosaic4E), an integrated framework that enables users to define extremes, identify and track their multiscale drivers, benchmark AI emulators such as ACE2 and NeuralGCM against Earth system models such as CESM and E3SM, and support prediction and downscaling from global variability to local impacts. Applications include the record-breaking 2024 Southern California and the 2025 Pacific Northwest floods, and extreme active and inactive Atlantic hurricane seasons during 1975–2025.

Emerging AI capabilities calls for a new paradigm that strategically aligns Earth system models, AI emulators, and observing systems. The upcoming Tropical Pacific Experiment (TEPEX) offers an opportunity to advance this paradigm by integrating targeted observations, high-resolution coupled Earth system modeling, and AI to improve understanding and prediction of multiscale drivers of weather-climate extremes.