Dynamics-based deep learning keeps forecast skill when observations are missing

Why it matters

Real climate observations are always incomplete. The results suggest that partial data can still carry usable information about slow, large-scale climate behavior. This could help seasonal forecasting where the full state of the system is unknown, though the real-world test here is small.

What they did

The authors hid some variables in five chaotic test systems: Lorenz 63, a seven-dimensional hyperchaotic system, Lorenz 96, a multi-scale Lorenz 96 and the Kuramoto–Sivashinsky equation. They compared DSDL with ANN, RC-ESN, LSTM, NG-RC and SINDy, using 100 repeated experiments. They then used DSDL for one-month-lead hindcasts of June–August rainfall over China for 2000–2022. Predictors were sea surface temperature and 500-hPa height from 36 global regions, picked by cross-validated stepwise regression, with a 30-year moving training window.

Key findings

• In all five test systems with missing variables, DSDL had the longest effective prediction time among the six methods.
• DSDL also had the lowest coefficient of variation across 100 repeated experiments, so its skill was the most stable.
• In the Kuramoto–Sivashinsky system, DSDL kept a prediction horizon near 40 dynamical time units after 12 spatial variables were removed.
• For China rainfall, DSDL reproduced the main drought–flood patterns in most years, with the best skill over the eastern monsoon region. Ps scores stayed around 70–80.
• It captured the broad structure of the July 2020 Yangtze wet anomaly and the 2022 southern drought, at least in part.

Limitations

• The climate test covers only one-month-lead hindcasts of summer rainfall over China, so longer seasonal leads are untested.
• DSDL was not compared with operational dynamical seasonal forecast systems or other climate-oriented models. Skill is weaker over complex terrain and arid inland areas, and some years show spatial smoothing that underestimates or displaces extreme anomaly centers.
• Only missing state variables were studied, and the physical mechanisms behind the selected predictors were not explored.

Glossary

• Dynamical System Deep Learning (DSDL): A forecasting framework based on delay-embedding theory that maps the current state to a future state.
• Takens’ embedding theorem: A result saying a system’s hidden dynamics can be rebuilt from time-delayed copies of partial observations.
• Effective prediction time (EPT): How long a forecast stays within an error threshold before it first exceeds it.
• Coefficient of variation (CV): The standard deviation divided by the mean of EPT across repeated experiments; lower means more stable.