cs.CVMar 23, 2026

Climate Prompting: Generating the Madden-Julian Oscillation using Video Diffusion and Low-Dimensional Conditioning

Authors: Sulian Thual, Feiyang Cai, Jingjing Wang, Feng Luo

Organizations: School of Computing, 100 McAdams Hall, Clemson, SC, USA.

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Generative Deep Learning is a powerful tool for modeling of the Madden-Julian oscillation (MJO) in the tropics, yet its relationship to traditional theoretical frameworks remains poorly understood. Here we propose a video diffusion model, trained on atmospheric reanalysis, to synthetize long MJO sequences conditioned on key low-dimensional metrics. The generated MJOs capture key features including composites, power spectra and multiscale structures including convectively coupled waves, despite some bias. We then prompt the model to generate more tractable MJOs based on intentionally idealized low-dimensional conditionings, for example a perpetual MJO, an isolated modulation by seasons and/or the El Nino-Southern Oscillation, and so on. This enables deconstructing the underlying processes and identifying physical drivers. The present approach provides a practical framework for bridging the gap between low-dimensional MJO theory and high-resolution atmospheric complexity and will help tropical atmosphere prediction.

Figures & tables

ConditioningDescription
pc1First principal component of the RMM-UBC index (MJO phase/amplitude)
pc2Second principal component of the RMM-UBC index (MJO phase/amplitude)
doycCosine component of the day-of-year (seasonal cycle)
doysSine component of the day-of-year (seasonal cycle)
n34sstNiño 3.4 Sea Surface Temperature anomaly (ENSO state)
Table 1: Low-dimensional conditioning variables used for climate prompting.
ModelFieldsConditionings
MJO Onlyolr, ubc, zbc, q400pc1, pc2
MJO with Seasonsolr, ubc, zbc, q400pc1, pc2, doyc, doys
MJO with Seasons/ENSOolr, ubcpc1, pc2, doyc, doys, n34sst
Table 2: Trained Models.
Model Architecture (U-Net)
Levels (Hierarchy)4Dim Multipliers(1, 2, 4, 8)
Bottleneck Size2×82\times 8Attention Heads8
Positional EncodingRotary (Temp)Sampling Factor2×2\times
Parameter (Training)
Field width (lon)64Field height (lat)16
Field Channels2 to 4Field Frames (time)16
Table 3: Model parameters for architecture, training, and sampling.
Figure 1: Model generated MJO sequence. a , Conditionings pc1 and pc2, as a function of time. b , Conditionings doys, doys and n34sst. c,d,e,f Hovmollers of sample fields, as a function of time and longitude: zonal winds (UBC), geopotential (ZBC), moisture (Q400) and convection (OLR). Fields are averaged over equatorial band (15N-15S). Here the model is prompted with observed conditionings from ERA5 1960-2022 (thus it regenerates the training record). All conditionings and fields are standardized. The principal components deduced directly from the sample fields match the conditionings (gray dashed lines in a).
Figure 2: Model generated power spectra. a , Power spectra of symmetric and asymmetric components 15N-15S, for UBC=U850-U200 (zonal wind velocity). MJO is within band k=1-3, w=0.01-0.03 (30-90 days). Gray lines indicate analytical dispersion curves for major equatorial waves. b , Repeated for ZBC=Z850-Z200 (geopotential). c , Repeated for Q400 (moisture). d , Repeated for OLR (convection). Here the model is prompted with observed conditionings from ERA5 1960-2022, as in Fig. 1 .
Figure 3: Ensemble Sampling. a , Conditionings pc1, pc2 as a function of time. b,c,d , Hovmollers (15N-15S) for Observed Fields UBC, Q400, OLR from the ERA5 training record, as a function of time and longitude. e,f,g , Hovmollers for ensemble mean. h,i,j , Hovmollers for ensemble standard deviation. k,l,m , Hovmollers for ensemble difference, that is difference between one member and the ensemble mean. Here the model is prompted with observed ERA5 conditionings 1960-2022, but sampled 10 times.
Figure 4: Model Prompt for a perpetual MJO. a , Conditionings pc1, pc2 as a function of time. b-c , Hovmollers (15N-15S) of OLR, UBC as a function of frame (i.e. time) and longitude. d-e , Power spectra (symmetric and antisymmetric 15N-15S) for OLR and UBC, following Fig. 2 . This uses model version trained for MJO alone (Tab. 2 ).
Figure 5: MJO Composites. a, c, e, g , Composites of UBC, ZBC, Q400 and OLR, by MJO phase 1-8. b,d,f,h , Composites of pc1, pc2, by MJO phase 1-8. Here the model is prompted for a perpetual MJO as in Fig. 4 . MJO phases are deduced from pc1, pc2. This is deduced from the perpetual MJO flow in Fig. 4
Figure 6: Model Prompt for a perpetual MJO with seasonal modulation. a , Conditionings pc1, pc2 as a function of time (sinusoides with period 73 days). b , Conditionings doyc, doys for seasons (sinusoides with period 1 year). c-d , Hovmollers (15N-15S) of OLR, UBC as a function of time and longitude. e-f , Seasonal Power spectra (symmetric 15N-15S) for OLR and UBC, with seasonal multiplier centred on boreal winter. g-h , Seasonal power spectra with seasonal multiplier centred on boreal summer. This uses model version trained for MJO with Seasons (Tab. 2 ).
Figure 7: Model Prompt for a perpetual MJO with seasonal and ENSO modulation. a , Conditionings pc1, pc2 as a function of time (sinusoides with period 73 days). b , Conditionings doyc, doys for seasons (sinusoides with period 1 year), and n34sst for ENSO (sinusoide with period 2 years). Note all conditionings are 2-year periodic. c-f , Hovmollers (15N-15S) of Phase-Composite for UBC, ZBC, Q400, OLR, as a function of 2-year phase and longitude. Phase Composites are binned averages along the 2-year phase, which extracts the periodic signals. This uses model version trained for MJO with Seasons/ENSO (Tab. 2 ).
Figure 8: Architecture of the video diffusion model. The model predicts the noise ϵθ\epsilon_{\theta} added to the input atmospheric fields xtx_{t} . It implements a 3D hierarchical U-Net architecture with tokens (conditionings) embedded within the Encoder/Decoder blocks via global and cross-attention pathways. Skip connections concatenate encoder output with corresponding decoder input. Encoders downsample while Decoders upsample the feature maps.
More in this topic

Climate Modeling

Climate modeling uses physical simulations and data-driven methods to understand, predict, and project Earth’s climate and its response to external forces.

See all 69 papers

Explore similar work

CardsList

  1. Monthly Diffusion v0.9: A Latent Diffusion Model for the First AI-MIP

    Apr 15, 2026Kyle J. C. Hall, Maria J. MolinaClimate EmulationClimate Modeling

    Department of Atmospheric & Oceanic Sciences, University of Maryland, College Park, MD 20740

  2. Geospatial Diffusion-based Evolution Synthesis (GeoDES) for Storm-Centered Weather Augmentation

    Jul 21, 2026Sonia Cromp, Satya Sai Srinath Namburi GNVV, Youran Wang +4Synthetic Data AugmentationVideo Diffusion Models

    University of Wisconsin-Madison · GE HealthCare · University of Georgia +2

  3. Tropical Cyclone Forecasting via Latent Rectified Flow using Satellite Imagery and Atmospheric Fields

    Aug 8, 2026Meheru Zannat, Sk. Md. Masudul AhsanRectified FlowTropical Cyclone Intensity Forecasting

    Department of Computer Science and Engineering Khulna University of Engineering & Technology Khulna 9203, Bangladesh