Bin Mu

Tropical cyclone (TC) forecasting is essential for disaster preparedness, yet accurate intensity prediction remains challenging, particularly during rapid intensification. This study proposes TianXing‐TC, a lightweight probabilistic post‐processing framework that leverages ensemble forecasts from the AI weather prediction model TianXing and integrates multi‐source TC‐related information. The framework first generates perturbed TianXing ensemble forecasts and extracts TC‐centered environmental trajectories along the predicted storm tracks. A TC Correction Network then fuses these environmental trajectories with geostationary satellite cloud‐pattern imagery and storm‐state statistical factors to predict maximum sustained wind, minimum sea‐level pressure, longitude, and latitude up to 96 hr ahead. Compared with the raw TianXing ensemble, TianXing‐TC substantially reduces systematic intensity underestimation and improves track forecasts, particularly at longer lead times. Homogeneous comparisons with official forecasts show that TianXing‐TC consistently reduces intensity forecast errors and achieves competitive track forecast skill, with the largest long‐lead track improvements in the western North Pacific basin. Stratified verification further indicates useful short‐range skill during rapid‐intensification events, although rapid intensity changes remain underestimated. Integrated‐gradient analyses and case studies reveal physically plausible attribution patterns related to environmental flow, convective organization, and storm‐state persistence. These results suggest that AI weather model generated ensemble trajectories, when combined with satellite and statistical information, can provide an efficient complementary approach for TC forecast post‐processing.

Artificial intelligence (AI) is rapidly transforming Earth science, offering unprecedented capabilities to tackle the most pressing challenges in the field. This work explores significant advances and emerging challenges across the AI for atmosphere–ocean sciences, while outlining critical ways forward. We review deep-learning methods and their application in weather and climate forecasting, which outperforms dynamical models in accuracy and computational efficiency. The role of AI in detecting complex phenomena, enhancing data assimilation and reconstruction, bias correction and downscaling coarse model outputs is also examined. However, the ‘black-box’ nature of complex AI models necessitates a focus on explainable AI to build trust and extract mechanistic insight. The most promising path forward is identified as the development of hybrid physics–AI modeling, which integrates the data-driven power of AI with the foundational constraints of physical laws to ensure generalizability and causal consistency. A new framework for AI-based model intercomparison is essential for rigorous benchmark performance. Finally, we contextualize these technical developments by discussing the usefulness and applicability of AI to society, including the improvement of multi-hazard early-warning systems and green energy production. We conclude by envisioning the future of AI agents for Earth science—autonomous, goal-oriented systems capable of designing and running experiments, generating and testing hypotheses, and learning dynamics from multisource data. This synthesis underscores that AI is not merely a tool, but a paradigm shift, which will significantly improve how we understand and adapt to a changing climate.

Climate mode interactions are inherently causal. While deep learning excels at climate prediction, particularly for El Niño‐Southern Oscillation (ENSO), its interpretability remains limited to validating known correlations. Here we propose a paradigm integrating causal inference into data‐driven modeling to enable predictions based on genuine causal relationships. Following this paradigm, we develop ENSO‐CausalNet, achieving skillful ENSO prediction of the Niño 3.4 index up to 22 months ahead. Results reveal that dominant physical processes affecting ENSO vary with lead time, elucidating distinct causal pathways through which Bjerknes feedback and extratropical Pacific, Atlantic, and Indian Ocean air‐sea interactions drive ENSO variability. When input dimensionality increases, the model may learn incomplete causal relationships, resulting in degraded prediction skill. These findings demonstrate that forecast capability depends critically on comprehensive causal understanding, confirming the model's physical validity. This paradigm provides robust ENSO predictions extensible to other climate systems, enabling mechanistic analysis and scientific discovery.

Existing convolution recurrent neural networks (ConvRNNs)-based memory cells majorly take advantage of gated structures and attention mechanisms to extract discontinuous latent associations for spatial-temporal sequence forecast (STSF) problems, which may lead to serious over-fitting and spurious relationships with correlated noise. It is a consensus that incorporating cause-effect relationships in modeling can alleviate these problems. In this paper, we propose a Causality Attention Unit (CAU) to assist ConvRNNs by complementing the causal inference ability in a plug-and-play way. Specifically, CAU serially consists of the attention module and causality module. The former is constructed by a spatial-channel attention layer, which preliminarily generates the correlated future with the correlations between historical memories and the current state. The latter borrows the idea of transfer entropy ( TE ) to detect the latent cause-effect relationships and precisely correct the correlated future. A space-time exchange strategy for accelerating the calculation of TE in CAU is also designed. CAU can be easily combined with the existing ConvRNN cells, and we construct a simple general model to predict long-term spatial-temporal series, which consists of encoder/decoder and stacked CAU paralleled to stacked ConvRNN cells. After determining the optimal model structure, we carry out a series of experiments to evaluate model performance, including comparisons with other advanced models, training loss analysis, and multiple ablation and sensitivity studies. Experimental results show that our proposed model can effectively improve the performances of existing ConvRNNs to the state-of-the-are level on representative public datasets, including Moving MNIST, KTH, BAIR, and WeatherBench. The ablation and sensitivity studies verify the superiority of CAU. The learned causal maps precisely distinguish the pixel attributions and motion characteristics in sophisticated entangled scenarios.

The North Atlantic Oscillation (NAO) is a major climatic phenomenon in the Northern Hemisphere, but the underlying air–sea interaction and physical mechanisms remain elusive. Despite successful short-term forecasts using physics-based numerical models, longer-term forecasts of NAO continue to pose a challenge. In this study, we employ advanced data-driven causal discovery techniques to explore the causality between multiple ocean–atmosphere processes and NAO. We identify the best NAO predictors based on this causality analysis and develop NAO-MCD, a multivariate air–sea coupled model that incorporates causal discovery to provide 1–6 month lead seasonal forecasts of NAO. Our results demonstrate that the selected predictors are strongly associated with NAO development, enabling accurate forecasts of NAO. NAO-MCD significantly outperforms conventional numerical models and provides reliable seasonal forecasts of NAO, particularly for winter events. Moreover, our model extends the range of accurate forecasts, surpassing state-of-the-art performance at 2- to 6-month lead-time NAO forecasts, substantially outperforming conventional numerical models.

A seasonal predictability barrier has long been noticed in ENSO forecasting with numerical models. Previous studies explored the impact of seasonal optimal initial perturbation evolutions in sea surface temperature anomalies (SSTA) on ENSO forecasting using the intermediate coupled model (ICM) via the conditional nonlinear optimal perturbation (CNOP) method. In this paper, we investigate the joint effects of SSTA and sea level anomalies (SLA) from the perspective of the optimal growth initial error (OGE). After determining the four seasonal OGEs about SSTA and SLA (i.e., SSTA-OGE, SLA-OGE and Joint-OGE), we first demonstrate the patterns, evolutions and the resulting spring predictability barrier (SPB) of the above OGEs. Then, we analyze the mechanism of OGE evolutions and SPB. Finally, we conduct observing system simulation experiments to determine the best (economic) observation network. Our experimental results indicate that the ENSO evolution error induced by SSTA-OGE and Joint-OGE presents season dependency, but SLA-OGE has no impact on ENSO evolution. Moreover, Joint-OGEs induce error evolutions and the SPB with more significant intensity than SSTA-OGEs and SLA-OGEs. From mechanism analyses, the evolutions of SSTA-OGEs are mainly dominated by Bjerknes feedback. Further, the evolution dynamics of Joint-OGEs primarily contain the continuous heating between the upper ocean combined with Bjerknes feedback and thermal diffusion in response to the discharge process. In addition, comprehensive and economical sensitive areas are identified through Joint-OGE, including the central-eastern equatorial Pacific and the western and north-eastern tropical Pacific boundary, which contribute to the ENSO prediction benefits reaching 58.31% on average.