Madden-Julian Oscillation (MJO)

• The Madden–Julian Oscillation (MJO) is a dominant tropical convective system characterized by eastward-propagating envelopes of enhanced and suppressed convection.
• Recent diagnostic methods, including the RMM index and transformer-based autoencoder frameworks, improve phase identification and reduce spurious propagation in MJO tracking.
• Advanced prediction techniques using physics-guided U-Net, deep learning, and stochastic models extend forecast skills by up to 35 days, enhancing operational climate projections.

The Madden–Julian Oscillation (MJO) is the dominant component of intraseasonal variability in the tropical atmosphere and an important driver of global weather and climate extremes. It is a planetary-scale convective system characterized by large-scale envelopes of enhanced and suppressed convection that propagate eastward and contain numerous mesoscale convective systems (MCSs). Despite its central role in subseasonal variability, accurately defining its life cycle, monitoring its state, and predicting its propagation remain difficult: operational dynamical models are typically skillful for only about 3–4 weeks, and even the widely used Real-time Multivariate MJO (RMM) index has been argued to mix mathematical artifacts with physical states (Zhou et al., 20 Oct 2025, Zhou et al., 14 Dec 2025, Yang et al., 22 Apr 2026).

1. Canonical structure, indices, and phase-space representations

A standard diagnostic of MJO state is the RMM index, built from two principal components, RMM1 and RMM2, derived from empirical orthogonal functions of near-equatorially averaged outgoing longwave radiation (OLR), 850-hPa zonal wind, and 200-hPa zonal wind. Its amplitude is commonly written as

$A=\sqrt{RMM1^2+RMM2^2},$

and an active event is often defined by the threshold $A \geq 1$. In this convention, event duration is the number of consecutive active days, while event size can be defined as the sum of daily amplitudes over the event interval (Minjares et al., 27 Jul 2025, Corral et al., 2023).

The RMM framework has been operationally influential because it compresses convection and circulation into a low-dimensional phase space, but recent work has emphasized its limitations. In particular, the established linear-projection method is reported to conflate location and intensity, to suffer from spurious propagation, and to misplace convection in physically suppressed regions. A transformer-based autoencoder framework, PhysAnchor-MJO-AE, was introduced to learn a similarity-preserving latent representation of daily MJO states and to cluster them into six canonical phases rather than the classical eight-angle phase diagram. The resulting six-phase “anatomical map” identifies Eastern Africa, Indian Ocean, Maritime Continent, Philippine Sea, Western Pacific, and Central Pacific states, and explicitly isolates two transitional phases: organizational growth over the Indian Ocean and a northward shift over the Philippine Sea. About 75% of all MJO-active days are assigned to these six canonical phases, and the associated monitoring framework reduces spurious propagation from 23% to 0.24% and convective misplacement from 17% to 3% relative to the classical RMM index (Zhou et al., 14 Dec 2025).

Alternative indices target related aspects of the same phenomenon. VPM1 and VPM2 replace OLR with 200-hPa velocity potential, while ROMI is based solely on OLR. Spectral analysis of RMM1, RMM2, VPM1, and VPM2 indicates that the raw indices contain a high-frequency component with $E(f)\propto f^{-3}$, whereas a 3-day running average yields a stretched-exponential form,

$E(f)\propto \exp\!\left[-(f/f_0)^{1/2}\right],$

which has been interpreted as characteristic of Hamiltonian distributed chaos with spontaneously broken time translational symmetry. The corresponding fundamental period, $T_f=50\pm 7$ days, is consistent with the canonical intraseasonal scale of the MJO (Bershadskii, 2018).

2. Dynamical theories and reduced mathematical descriptions

Reduced models of the MJO span several mechanistic traditions. In the MJO skeleton model of Majda and Stechmann, nonlinear traveling-wave solutions exist in four branches—dry Rossby, moist Rossby, MJO, and dry Kelvin—and the MJO branch is a pulse-like convective disturbance with a narrow region of enhanced convection and a wide region of suppressed convection. Its dispersion relation depends explicitly on amplitude: larger amplitude implies longer wavelength for a given wave speed, while the nonlinear MJO propagates more slowly and at lower frequency than the linear MJO. In the weak-forcing limit, the convective activity reduces to a solitary $\mathrm{sech}^2$ pulse, and exact solution construction depends crucially on conservation of total energy and on the Hamiltonian structure of the reduced traveling-wave equations (Chen et al., 2015).

A different conceptualization describes the MJO as a nonlinear Burgers kink in the meridional vorticity equation. In that framework, buoyancy is parameterized by convective available potential energy, and the large-scale envelope satisfies

$$\frac{\partial \eta}{\partial t}-s\eta\frac{\partial \eta}{\partial x}=D\frac{\partial^2\eta}{\partial x^2}.$$

The resulting traveling kink moves toward the moisture source, with propagation speed controlled by asymmetry in zonal surface winds. This model also predicts convection at the equator to the east of the MJO that is not correlated with Kelvin and Rossby waves (Blender, 2023).

Shallow-water theories have proposed yet other mechanisms. One study of nonlinear Wind Induced Surface Heat Exchange (WISHE) forcing found that a slow eastward-propagating Kelvin-like signal arises because a Yanai wave group forces a Kelvin response whose phase speed is reduced relative to the free Kelvin mode; in the model, the slow signal propagates at about $8$–$12\,\mathrm{m/s}$, compared with a free Kelvin speed of about $22\,\mathrm{m/s}$, while the observed MJO is near $A \geq 1$0 (Solodoch et al., 2010). Another shallow-water model with triggered convection and steady radiative cooling interprets the MJO-like envelope as an interference pattern of westward and eastward inertia-gravity waves. In that setting, the envelope propagation speed is one half of the speed difference between the WIG and EIG waves, and quadrupole vortex structures emerge in longitude–latitude composites (Yang et al., 2012).

Stochastic conceptual models preserve the same planetary-scale variables while introducing multiplicative randomness to represent bursty convection and multiscale coupling. A stochastic skeleton model with time-dependent observation-based forcing reproduces the observed statistics of event lifetime, extent, and amplitude, but fails to capture the seasonality of MJO events and their dependence on El Niño, La Niña, and neutral ENSO conditions (Ehstand et al., 27 Jan 2025). A related stochastic conceptual model for coupled ENSO and MJO uses a three-box ocean and a low-order Fourier representation of the atmospheric MJO; with state-dependent noise in moisture and convection, it reproduces observed non-Gaussian ENSO diversity and MJO spectra and captures interactions between wind, MJO, and ENSO (Moser et al., 2024).

3. Multiscale organization, convection, and the life cycle

A central issue in MJO theory is whether predictability resides primarily in large-scale fields or depends essentially on smaller-scale convective organization. Deep-learning experiments designed explicitly to test this question found that large-scale patterns dominate the learned signals: models using only large-scale input match the forecast skill of models using all scales, especially for RMM and ROMI. At the same time, small-scale signals remain informative because deeper feature maps reconstruct the large-scale envelope of small-scale activities. This has been interpreted as support for both a large-scale view of the MJO and a multiscale envelope view in which small-scale convection organizes into planetary-scale structure (Yao et al., 4 Oct 2025).

This multiscale interpretation is consistent with observational analyses of the relation between the MJO and MCSs. Satellite-based composites show that the MJO modulates MCS frequency, intensity, and organization through moisture, instability, and vertical wind shear. Active MJO phases feature higher MCS frequency, greater MCS intensity, and increased organization, especially over the Indian Ocean through the Maritime Continent, while suppressed phases exhibit reduced occurrence and vigor. The interaction is bidirectional: enhanced MCS populations are associated with coherent large-scale circulation anomalies, momentum transports, and thermodynamic anomalies that reinforce the MJO convective envelope and support its eastward propagation. In this view, MCSs are not merely passive responses to the MJO environment but active contributors to its maintenance and evolution (Yang et al., 22 Apr 2026).

Recent data-driven work has also reframed the MJO life cycle itself. Clustering of learned “MJO fingerprints” in latent space yields a six-phase anatomy rather than a purely angular phase-space trajectory. The newly isolated Indian Ocean transitional phase represents large-scale organization and rapid convective growth, while the Philippine Sea phase represents a northward shift of convection that had been obscured by meridional averaging in the classical index. This suggests that onset, growth, mature propagation, and monsoon-linked reorganization are more cleanly separated in the full two-dimensional fields than in linear EOF projections (Zhou et al., 14 Dec 2025).

4. Propagation, barriers, event statistics, and nonlinear regime behavior

One of the most persistent operational difficulties is the “Maritime Continent barrier,” the tendency for forecasts to stall or weaken the MJO as it passes the Maritime Continent, particularly around phases 4–5. Raw dynamical forecasts often fail to sustain correct eastward propagation and amplitude into the western Pacific. Case studies and composite diagnostics show that successful correction requires not only amplitude adjustment but repair of the full spatial-temporal structure of OLR and wind anomalies so that coherence is maintained across the Maritime Continent and into the western Pacific (Zhou et al., 20 Oct 2025).

At the event level, the MJO does not behave like a single regular oscillator. Statistical analyses of active episodes defined by $A \geq 1$1 show that both event durations and event sizes follow a double power-law distribution. A characteristic duration of about 27 days separates two scale-free regimes, and after about 27 days there is a sharp increase in extinction probability. Because this increase is independent of the starting and ending phases, it has been interpreted as evidence for an internal mechanism of exhaustion rather than control by a fixed external barrier (Corral et al., 2023).

Work on extreme MJO events refines this picture. Using a power-law fit to the event-size distribution, extreme events were defined by size $A \geq 1$2, with a corresponding duration threshold of $A \geq 1$3 days. These events most frequently initiate in phases 2–3 through the year, often complete a full phase cycle, and are stronger than weak events in their OLR, eddy streamfunction, and velocity potential anomalies. During La Niña, extreme events tend to last longer than during El Niño, a modulation not seen in weak events (Minjares et al., 27 Jul 2025).

A further complication is that MJO propagation may be intrinsically regime-switching. Cloud-system-resolving ensembles comprising 4,000 simulations of two MJO events reveal multiple regimes with distinct timings of MJO propagation under a single atmosphere–ocean background. The emergence of bimodal propagation behavior depends critically on the equatorial asymmetry of climatological sea surface temperature, and regime selection is probabilistic, influenced by whether tropical–extratropical interplay promotes moistening associated with westward-propagating tropical waves over the western Pacific. This formulation treats MJO propagation as a deterministic chaotic phenomenon produced by cross-scale nonlinear interactions rather than as a purely linear eastward mode (Takasuka et al., 29 Jun 2025).

5. Prediction skill, post-processing, and machine-learning frameworks

Subseasonal MJO prediction remains difficult. Operational dynamical models are typically skillful for only about 3–4 weeks, with skill usually evaluated using a bivariate correlation threshold of 0.5. One commonly used form is

$A \geq 1$4

Within this framework, recent work has focused both on direct prediction and on AI-based correction of dynamical forecasts (Zhou et al., 20 Oct 2025).

A prominent post-processing approach is the Physics-guided Cascaded Corrector for MJO (PCC-MJO). Its first stage is a physics-informed 3D U-Net that corrects spatial-temporal errors in OLR, $A \geq 1$5, and $A \geq 1$6 using spatial kernels designed to capture multi-scale features and temporal kernels tuned to MJO time harmonics. The corrected fields are then projected onto the leading Wheeler–Hendon EOF patterns, and a second-stage LSTM refines the RMM time series by directly maximizing bivariate correlation. Applied to operational forecasts from CMA, ECMWF, and NCEP, the method extends the skillful forecast range by 6, 8, and 2 days, respectively, and explainable AI based on Integrated Gradients shows attribution patterns with spatial correlations of 0.94 for RMM1 and 0.93 for RMM2 against observed EOF structures (Zhou et al., 20 Oct 2025).

Other machine-learning systems target direct prediction. A U-Net-inspired deep convolutional neural network predicts RMM skillfully to 21 days and ROMI to 33 days, with skills comparable to leading subseasonal-to-seasonal systems such as NCEP, and winter RMM skill reaching 25 days. Spectral analysis of the learned latent space indicates that planetary and large-scale fields are the primary source of this predictability, although small-scale-only inputs remain skillful for 1–2 weeks after retraining because the network reconstructs large-scale envelopes from their spatial organization (Yao et al., 4 Oct 2025). The FuXi-S2S machine-learning model extends boreal-winter MJO skill to 35 days, compared with 28 days for ECMWF S2S, and its improvement during strong initial phase 3 is linked to more accurate prediction of the area-averaged meridional gradient of low-frequency background moisture over the tropical western Pacific (Cao et al., 22 Aug 2025).

Probabilistic forecasting has motivated alternative methods. An autoregressive Gaussian-process model calibrated with empirical correlations outperforms ANN models during the first five lead days and, after a posteriori covariance correction, extends probabilistic coverage by more than three weeks (Chen et al., 21 May 2025). Reservoir computing trained on a causal real-time band-pass filtered RMM1 time series yields skillful prediction for about a month from pre-developmental stages, while best-performing cases reproduce RMM evolution over two months, exceeding the expected inherent predictability limit cited in that study (Suematsu et al., 2022).

Framework	Reported skill result	Distinctive feature
PCC-MJO	CMA: 33 to 39 days; ECMWF: 18 to 26 days; NCEP: 18 to 20 days	Universal post-processor with 3D U-Net plus LSTM
DCNN forecast model	RMM: 21 days; ROMI: 33 days	Spectral analysis of learned features
FuXi-S2S	35 days vs 28 days for ECMWF S2S	Improved western Pacific moisture-gradient prediction
Gaussian-process model	Better than ANN models for first five lead days	Direct uncertainty quantification and covariance correction
Reservoir computing	One month routinely; best cases over two months	Causal filtering and time-delay embedding

6. Teleconnections, extremes, and broader climatic significance

The MJO modulates weather and climate extremes well beyond the tropical Indo-Pacific. Over South America during austral summer, extreme MJO events induce more intense rainfall anomalies of larger spatial extent than weak events and strongly affect the South American rainfall Dipole. Composite analyses show stronger and more extensive OLR, upper-level eddy streamfunction, and velocity-potential anomalies during extreme events, especially for initiation in phases 2–3 and 6–7. Enhanced convection over the Indian Ocean is linked to increased rainfall over southeastern South America and reduced rainfall over the South Atlantic Convergence Zone, with the dipole structure evolving as the MJO propagates eastward (Minjares et al., 27 Jul 2025).

Humid heatwaves are also strongly modulated by the MJO and the boreal-summer intraseasonal oscillation. Across much of the tropics and subtropics, humid heatwave likelihood can double or halve depending on the phase, over both land and ocean. The primary control is through specific humidity rather than dry-bulb temperature, and in the subtropics the dominant mechanism is horizontal advection of the climatological moisture gradient in the planetary boundary layer by MJO-related anomalous winds. This establishes a direct subseasonal pathway from phase-dependent tropical circulation anomalies to wet-bulb temperature extremes (Rocuet et al., 22 Sep 2025).

On intraseasonal monsoon variability, the spectral form of smoothed RMM and VPM indices has been used to argue that the MJO pumps analogous Hamiltonian distributed chaos into the Asian–Australian monsoon system, reinforcing the view that monsoon intraseasonal variability is dynamically linked to the same underlying tropical disturbance rather than constituting unrelated noise (Bershadskii, 2018). Coupled stochastic models further suggest that state-dependent interactions among MJO convection, winds, and ENSO-related latent heat are needed to reproduce the observed non-Gaussianity of tropical climate variability and the diversity of extreme events (Moser et al., 2024).

Taken together, these results place the MJO at the center of a hierarchy of tropical and global variability: a planetary-scale convective envelope with internal event statistics, multiscale embedded organization, sensitivity to background state and chaos, and demonstrable influence on rainfall, humid heat, and monsoon variability. A plausible implication is that progress in MJO science depends simultaneously on better physical theories of propagation, better state representations than linear phase-angle diagnostics alone, and prediction systems that correct or learn the full spatial-temporal fields rather than only low-dimensional indices.