The Madden–Julian Oscillation (MJO) is the dominant component of intraseasonal variability in the tropics. It manifests as an equatorial planetary wave that originates over the Indian Ocean and propagates eastward at about 5 m/s, with a pronounced signal at zonal wavenumbers 1–3 and periods of 40–90 days. The MJO modulates monsoon evolution, midlatitude predictability, and the onset of El Niño events, making it a critical factor for forecasts from subseasonal to seasonal scales. However, the MJO is inherently intermittent and disorganized; its multiscale nature, driven by the nonlinear coupling of convection and large-scale circulation, remains a major source of forecast uncertainty. This complexity perpetuates the gap in our ability to model and predict tropical weather transitions, challenging both theoretical approaches and numerical models. This work stands apart: there is almost no one nearby, the closest neighbors in the archive work on diffusion models, machine learning, and the large-scale structure of the universe, but they do not connect generative models with low-dimensional climate theory.
The model is built on a standard U-Net architecture with transformers, extended with a temporal dimension, and trained on ECMWF ERA5 reanalysis. Data cover the tropics (35°N–35°S) on a 5.625°×4.375° grid (64×16 points) from 1960 to 2022, with daily resolution. Fields used include UBC (zonal wind difference at 850 and 200 hPa), ZBC (geopotential difference), Q400 (specific humidity at 400 hPa), and OLR (outgoing longwave radiation). Climatology and interannual variability are removed using a Butterworth filter with a 120-day cutoff to isolate intraseasonal anomalies. Low-dimensional conditions include a modified RMM index (RMM-UBC), seasonal embeddings doyc and doys (sine and cosine of day of year), and the Niño 3.4 SST index (n34sst). The model was trained on sequences of 16 frames (days) with two conditioning pathways: global via ResNet blocks and cross-attention for spatial and temporal attention. For long sequence generation, Brick-Wall Denoising with a shift of 3 frames was used. Training took about 12 hours on two A100 GPUs, and generating a 60-year record took about 30 minutes.
The model can reproduce the training record, generating intermittent MJO sequences whose amplitude and phase closely follow the pc1 and pc2 conditions. Power spectra by wavenumber and frequency show a pronounced MJO signal (k=1–3, periods 30–90 days), as well as convectively coupled Kelvin, Rossby, mixed Rossby–gravity, and inertia–gravity waves. However, the model is biased compared to ERA5: equatorial CC-Kelvin and MRG waves are less pronounced, which is explained by incomplete mapping of low-dimensional conditions to high-dimensional variability and the non-Gaussianity of the data. Ensemble sampling (10 realizations) showed that the ensemble mean smooths the MJO signal, while the standard deviation reflects seasonal variations with a winter maximum in the warm pool of the western and central Pacific, consistent with disorganized convective processes. The difference between a single realization and the ensemble mean reveals freely generated Kelvin and Rossby waves that vary from sample to sample. When prompted with idealized conditions, the model generates a perpetual MJO with a constant period of 65 days and an amplitude of 1.2 standard deviations, with composites showing eastward propagation and a quadrupole structure consistent with theory. Adding seasonal modulation revealed differences between winter and summer MJOs: summer spectra are more asymmetric and have more pronounced CC-Kelvin waves. Under ENSO modulation with a two-year cycle, composites showed increased intensity, extent, and eastward propagation of the MJO after El Niño compared to La Niña, consistent with observations.
This work differs from archive neighbors in that it is the first to connect generative video diffusion with low-dimensional theoretical MJO indices, turning climate metrics into tunable generative drivers. This allows testing hypotheses about tropical dynamics by decomposing processes through controlled idealized prompts. The approach bridges the gap between low-dimensional theory and high-dimensional atmospheric complexity, offering a practical framework for integrating deep learning with physical interpretability in Earth system modeling. Unlike purely predictive or explanatory methods, here the probabilistic nature of diffusion maps low-dimensional states onto a manifold of physically consistent high-dimensional realizations.
Prospects include improving the model to reduce biases in MJO representation, for example through appropriate data scaling (power or quantile transformations) to address non-Gaussianity. One could systematically evaluate sensitivity to model parameters and determine which low-dimensional latent variables are best suited for prompting MJO variability and its modulations. The approach could be extended to other climate problems where low-dimensional theory is established but a link to spatiotemporal complexity is needed. Ultimately, 'climate prompting' could clarify the connection between planetary-scale MJO dynamics, embedded convectively coupled equatorial waves, and external modulations.
The method could improve tropical atmosphere forecasting by providing a fast generator of realistic MJO scenarios for ensemble forecasts. It is also useful for studying the physics of tropical variability and testing theoretical models.
Next steps include mitigating model biases through data transformations and more systematic evaluation of hyperparameter sensitivity. It also remains to determine the optimal low-dimensional variables for prompting and to test the approach's applicability in forecasting tasks.
This work directly addresses an unresolved problem in atmospheric sciences: how to reconcile low-dimensional theoretical MJO models with the full complexity of tropical convection and circulation. The proposed approach offers a bridge through generative modeling, allowing investigation of MJO modulation mechanisms by seasons and ENSO under controlled conditions.
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The study uses a conditional video diffusion model to generate the MJO, trained on ERA5 reanalysis and conditioned on low-dimensional indices (RMM, season, ENSO). The generated sequences reproduce key MJO characteristics, including composites, spectra, and associated waves. Prompting with idealized conditions allows for process decomposition.
Where its strength is
The connection between low-dimensional MJO theory and high-dimensional atmospheric complexity through a generative model enables controlled exploration of physical drivers.
Where it could go next
Worth trying
Nearby works in our archive
The archive is sparse around this work: the closest review is only 0.49, against a usual 0.64. It looks like few have drilled nearby.
Generating a 60-year MJO record takes about 30 minutes, comparable in computational cost to models of intermediate complexity. Moreover, the model can create counterfactual scenarios, such as a perpetual MJO with a constant period, which are impossible to observe in nature.
The idea of 'prompting' a climate model resembles weather control in stories where scientists set parameters and obtain predictable yet realistic responses. However, here there is no manipulation of the real atmosphere, only exploration of its virtual twins.