A Unified Global Climate Mode Prediction Model (UniCM) Opens New Paths for AI‑Empowered Climate Science

Recent advances such as Pangu‑Weather, GraphCast, GenCast and Aurora have pushed AI‑driven weather forecasting to near‑or‑better than traditional numerical models for short‑range predictions. However, climate science demands understanding and forecasting of multi‑month to multi‑decadal changes across interacting climate modes.

Key climate phenomena—ENSO, IOD, TNA, NPMM, SPMM, among others—do not act in isolation but exchange energy through remote‑correlation, forming a globally coupled system that influences monsoons, droughts, floods and compound extremes. The challenge is to let AI capture these long‑term, multi‑modal couplings.

To address this, the authors propose UniCM, a unified climate‑mode prediction framework that places several major ocean‑atmosphere modes from the Pacific, Indian and Atlantic oceans into a single learning system. UniCM employs a dual‑branch multi‑view Transformer: the Globalformer branch learns bottom‑up spatio‑temporal dynamics from fine‑grained physical fields (sea‑surface temperature, wind stress, thermocline depth, upper‑ocean temperature), while the Modeformer branch learns top‑down nonlinear interactions among the climate modes.

A cross‑view coupling mechanism injects modal‑level information back into the physical‑field prediction process, enabling the model to capture bidirectional feedback—both "physical fields generate climate modes" and "climate modes regulate physical fields"—and thus learn long‑range, cross‑scale coupling that is hard to model explicitly.

Empirical evaluation shows UniCM achieving state‑of‑the‑art skill on ENSO forecasts. On reanalysis datasets (GODAS, ORAS5, SODA) UniCM outperforms representative baselines XRO (Nature 2024), CNN (Nature 2019), ResoNet (AAS 2024) and DESN (NPJ‑CAS 2026) across a 24‑month horizon, extending effective ENSO prediction to about 19 months and maintaining ACC > 0.5 for a 14‑month lead time. For the Indian Ocean Dipole, UniCM delivers a usable forecast window of roughly 7 months, and across seven key modes (ENSO, IOD, IOB, SIOD, SPMM, NPMM, TNA) it improves average skill by ~20 % over the baselines. The model also accurately reconstructs observed lag relationships, such as NPMM leading ENSO by ~4 months, confirming that it captures genuine physical coupling rather than spurious correlations.

Analysis of UniCM’s internal attention maps provides an interpretable view of its decision process. The model highlights spatial precursors before major ENSO events and reveals the pivotal roles of NPMM and TNA in complex multi‑modal anomalies, findings that align with existing climate teleconnection research.

Beyond superior forecasts, UniCM demonstrates how AI can become a scientific discovery partner: by exposing emergent predictability hidden in coupled climate dynamics, it offers new avenues for early warning of extreme events and supports sectors such as renewable energy, fisheries, agriculture and disaster mitigation. The authors anticipate extending the framework to intra‑seasonal oscillations, decadal variability, and climate‑mode evolution under global warming, heralding a new paradigm where AI jointly predicts and explains Earth system behavior.