AI Learns to Recover Hidden Climate Predictability from Sparse Data

Climate scientists have long faced an uncomfortable truth: the atmosphere–ocean system is effectively infinite-dimensional, yet we can only ever observe fragments of it. Sparse station networks, satellite retrievals and reanalysis products offer partial projections of a state that can never be fully measured. A new study published in Climate Dynamics argues that this incompleteness may not be the fatal blow to seasonal prediction that it appears to be. Drawing on decades-old mathematics of chaos theory, a research team led by Hao Li and Jianping Li of the Ocean University of China demonstrates that large-scale climate predictability can be substantially recovered from sparse observations using a framework called Dynamical System Deep Learning, or DSDL.

The theoretical foundation of the approach is Takens’ delay embedding theorem, a landmark result from 1981 showing that the phase-space structure of a deterministic chaotic system can be reconstructed from time-delayed coordinates of a single observed variable. In other words, even if most of a system’s state variables are hidden, the shadow they cast on the ones you can see may be enough to reconstruct the underlying dynamics. The DSDL framework operationalizes this idea by building two mathematical objects: a non-delayed attractor constructed from multivariate time series, and a delayed attractor for the target variable to be predicted. A nonlinear mapping between the two, learned from data and extended with higher-order monomial terms, then maps the current state of the system to its future state.

To test whether this works when parts of the system are simply missing, the team subjected DSDL to a punishing gauntlet of five canonical nonlinear systems: the classic three-variable Lorenz’ 63 model, a seven-dimensional hyperchaotic circuit system, the Lorenz’ 96 model with eight variables, a multiscale coupled version of Lorenz’ 96 that explicitly represents interactions among slow, intermediate and fast components, and the Kuramoto–Sivashinsky equation with 64 spatial variables, a benchmark of spatiotemporal chaos. In each case, only a subset of state variables was retained as observable, while the rest were treated as unobserved. The multiscale coupled Lorenz’ 96 system is particularly important because it mimics key features of the real atmosphere, where slowly evolving large-scale variables interact with rapidly varying small-scale processes.

The results were compared against a battery of representative data-driven forecasting methods, including artificial neural networks, echo state network reservoir computing, long short-term memory networks, next-generation reservoir computing, and SINDy, a sparse-regression method for discovering governing equations. Performance was quantified with two metrics: the Effective Prediction Time, which measures how long a forecast stays within one standard deviation of the reference trajectory, and the Coefficient of Variation, which measures how consistent that performance is across repeated experiments. All results were averaged over 100 independent trials. Across every system and every missing-variable configuration tested, DSDL achieved the longest effective prediction horizons and the lowest variability, indicating that its advantage is not an artifact of any particular dynamical regime.

Some of the individual results are striking. In the multiscale coupled Lorenz’ 96 system, DSDL maintained the most persistent predictability precisely when the small-scale variables were unobserved, the hardest configuration. In the Kuramoto–Sivashinsky equation, the framework retained a prediction horizon approaching 40 dynamical time units even after 12 of its 64 spatial variables were removed, substantially outperforming the alternatives. SINDy could not even be evaluated for that system, because estimating the high-order spatial derivatives it requires from incomplete observations is hopelessly unstable without prior knowledge of the governing equations. The authors also extended DSDL to handle temporal gaps between training and prediction periods, showing meaningful skill in experiments up to ten steps ahead, a capability that matters for seasonal forecasting where predictors and targets are separated in time.

The real test, of course, is the actual climate system. The team applied DSDL to one-month-lead hindcasts of summer precipitation over China for the period 2000 to 2022, using the CN05.1 gridded precipitation dataset built from 2,416 stations, supplemented with GPCC data over Taiwan, and drawing predictors from sea surface temperature in the NOAA ERSST v6 analysis and 500-hPa geopotential height from the ECMWF ERA5 reanalysis. The globe was partitioned into 36 regions of 30 by 60 degrees, and sea surface temperature and circulation variables from each region, along with their time-lagged terms and second-order nonlinear combinations, were assembled into a large candidate predictor pool. A cross-validated stepwise regression procedure then automatically identified the key precursor variables for each precipitation grid point, keeping the selection transparent and traceable rather than hand-tuned.

The hindcast skill was concentrated exactly where theory says it should be. Over eastern China’s monsoon region, where seasonal precipitation is strongly modulated by large-scale circulation and slowly varying air–sea coupling, DSDL reproduced the dominant drought–flood anomaly patterns with positive anomaly spatial correlations in most years and operational Ps scores that remained stable around 70 to 80 across the 23-year period. Anomaly temporal correlations were highest over the Yangtze River basin, North China and the Northeast, and weakest over parts of the Northwest and Southeast, where precipitation is dominated by localized topographic and convective processes that sparse large-scale predictors cannot capture.

The framework’s performance in extreme years is likely to draw the most attention. In July 2020, persistent Meiyu rainfall produced severe flooding along the middle and lower reaches of the Yangtze River, with drier conditions over parts of southern China. DSDL reproduced both the major wet anomaly center and the north–south contrast between the Yangtze basin and southern China, earning positive spatial correlation scores and high Ps values. Two years later, during the severe 2022 summer drought over the Yangtze basin, the model recovered the widespread dry anomaly structure across southern China and captured the drought conditions over the Southwest, alongside wetter anomalies to the north. Capturing the sign and spatial organization of such high-impact seasonal anomalies at a one-month lead is exactly the kind of skill that operational forecasters prize.

The deeper message of the study is conceptual as much as technical. Conventional data-driven forecasting architectures are typically built on the implicit assumption of relatively complete observations and continuous trajectories, and they tend to treat missing variables as truncated boundary conditions or noise, suffering catastrophic skill loss when the state space is incomplete. DSDL instead exploits the fact that large-scale low-frequency climate variability is dynamically coupled to the observable variables through teleconnections and air–sea interactions, so that sparse local observations retain recoverable information about the global attractor. In this view, climate predictability does not depend solely on fully observed state-space information; it can be partially preserved within limited observations through nonlinear dynamical coherence.

The authors are careful to spell out the limitations. The climate application covers only one-month-lead summer precipitation over China, and whether the recoverable predictability persists at one-season or multi-season leads remains open. The study also did not compare DSDL against operational dynamical seasonal prediction systems or hybrid statistical–dynamical models, and the physical mechanisms linking the automatically identified precursor regions to Chinese rainfall, including possible connections to ENSO, Indo-Pacific air–sea interaction and the western Pacific subtropical high, still require dynamical diagnosis. Skill is weaker over topographically complex and arid inland regions, and the model smooths some local extreme anomalies. Scenarios involving temporal discontinuities and irregular observations remain untested. Even so, the demonstration that a dynamics-informed framework grounded in nonlinear embedding theory can extract seasonal predictability from deliberately incomplete information points toward a new generation of hybrid forecasting methods, one in which the mathematics of chaos and the machinery of deep learning work together rather than in competition.

Subject of Research: Recovering climate predictability from sparse observations using dynamical system deep learning based on nonlinear embedding theory

Article Title: Recovering climate predictability from sparse observations via dynamical system deep learning

Article References: Recovering climate predictability from sparse observations via dynamical system deep learning. (n.d.). https://doi.org/10.1007/s00382-026-08413-8

Image Credits: AI Generated

DOI: 10.1007/s00382-026-08413-8

Keywords: climate predictability, dynamical system deep learning, Takens embedding theorem, sparse observations, chaos theory, seasonal prediction, East Asian summer monsoon, precipitation hindcasts, machine learning, air-sea interactions, Lorenz systems, Kuramoto-Sivashinsky