UniCM: A Unified Global Climate Mode Prediction Model Paving a New AI‑Driven Path for Climate Science

Recent breakthroughs such as Pangu‑Weather, GraphCast, GenCast and Aurora have pushed AI weather forecasting to near‑or‑better accuracy than traditional numerical models, yet climate science demands forecasts spanning months to decades and an understanding of coupled ocean‑atmosphere dynamics.

To address this, Li Yong’s team at Tsinghua University introduced UniCM, a unified climate‑mode prediction framework that ingests multiple key ocean‑air modes from the Pacific, Indian and Atlantic oceans into a single model. UniCM employs a dual‑branch, multi‑view Transformer: the bottom‑up Globalformer learns spatio‑temporal evolution from fine‑grained physical fields (sea‑surface temperature, wind stress, thermocline depth, etc.), while the top‑down Modeformer captures nonlinear interactions among climate modes.

UniCM’s cross‑view coupling injects system‑level information back into the physical‑field prediction stream, enabling the model to learn both "physical fields generate climate modes" and "climate modes regulate physical fields". This bidirectional feedback captures long‑range, cross‑scale coupling mechanisms that are hard to model explicitly.

On the ENSO forecasting task, UniCM outperforms representative baselines (XRO (Nature 2024), CNN (Nature 2019), ResoNet (AAS 2024), DESN (NPJ‑CAS 2026)). Across a 24‑month horizon it extends effective skill to ~19 months, maintaining ACC > 0.5 for a 14‑month lead and accurately reproducing the 1997‑98 El Niño and the 2020‑2023 triple La Niña events, demonstrating that the model learns systemic dynamics rather than memorising index curves.

Beyond ENSO, UniCM is evaluated on seven coupled modes (ENSO, IOD, IOB, SIOD, SPMM, NPMM, TNA). It improves average forecast skill by ~20 % and captures lag structures such as NPMM leading ENSO by four months, confirming that the model learns genuine physical coupling rather than spurious correlations. The results suggest that long‑term predictability resides in the interactions among modes, not in isolated phenomena.

The internal attention mechanism provides an interpretable map of the model’s decision process. Analysis reveals that UniCM highlights spatial precursors and modal interactions preceding major ENSO events, such as the pivotal role of NPMM before the 1997 El Niño and the hub function of TNA in complex multi‑modal anomalies. This demonstrates that UniCM can serve as a scientific discovery partner, helping researchers generate hypotheses from massive climate datasets.

UniCM exemplifies the broader AI‑for‑Science trend: moving from point‑task prediction toward end‑to‑end scientific discovery, where AI assists hypothesis generation, experimental design and theory formulation. Extending the framework to intra‑seasonal oscillations, decadal variability and climate‑change contexts could further enhance early‑warning systems for disasters, renewable‑energy planning, fisheries and agriculture.