Research

My research asks how the large-scale tropical circulation is shaped by solar forcing, air–sea coupling, and eddy–mean flow interaction, and what those dynamics imply for climate variability, predictability, and model bias. I work across a hierarchy of tools — analytical theory, conceptual models of a few variables, idealized numerical models, coupled general circulation models (GCMs), and, most recently, machine-learning atmospheric emulators — with energy, momentum, and mass budgets closed diagnostically rather than inferred from correlations. My characteristic approach is to translate a phenomenological question, such as why the tropical rain band sits north of the equator or why the monsoon is predictable in some years and not others, into a structural question that a minimal model can answer definitively, and then to verify the answer in comprehensive models and observations.

Two threads organize the work. The first concerns the seasonal cycle, the largest climate signal on Earth, whose imprint on the annual mean is commonly assumed to average out; I have shown in several contexts that it does not. The second concerns the Asian summer monsoon: what controls its year-to-year variability, where its predictability comes from, and where it breaks down. Both threads are exercises in identifying a mechanism rather than a correlation, and both now converge on a current question — whether AI climate models reproduce those mechanisms or only their appearance.

Seasonal rectification and the position of the tropical rain band

The annual-mean intertropical convergence zone (ITCZ) sits north of the equator, and coupled models have long struggled to reproduce this, generating a spurious double ITCZ instead. I showed that the strength of the seasonal cycle is a first-order control on this asymmetry, even when annual-mean insolation is hemispherically symmetric. Varying Earth's obliquity from 0° to 30° in a fully coupled GCM, enhanced austral-summer heating pushes Southern Hemisphere SSTs above the tropical convective threshold and activates a transient southern rain band, which engages the wind–evaporation–SST feedback. No comparable response occurs in the opposite season. Because the air–sea coupled response is seasonally selective, it rectifies into the annual mean.

The argument is carried across a model hierarchy: a coupled GCM with varying obliquity, a two-dimensional mixed-layer-ocean–Matsuno–Gill model, and a three-variable conceptual model. Across CMIP6, the multi-model mean overestimates the amplitude of the southeastern tropical Pacific SST seasonal cycle by 1.3 K, or 42%, and inter-model variations in that bias are closely related to errors in annual-mean ITCZ asymmetry (r = 0.77). Seasonal SST evolution is therefore one contributor to the double-ITCZ bias. The analysis does not identify which process produces the excessive seasonality: the respective roles of the mixed layer and thermocline remain unresolved, as do the effects of upwelling and cloud-radiative feedbacks.

Zhang, Xie, Lutsko, Okumura, Yang & Shaw (2026), Nature Geoscience, in press.

Ocean memory and monsoon predictability

The Northwestern Pacific anomalous anticyclone (AAC) is the dominant mode of East Asian summer monsoon variability, and its predictability rests on the ocean. El Niño peaks in winter, but the Indian Ocean stores the signal through surface heat flux adjustments and slow oceanic Rossby waves, then discharges it the following summer via the Indo-western Pacific Ocean capacitor.

Two results sharpened this picture. Using tropical Pacific pacemaker and initialized large-ensemble experiments, I showed that the AAC behaves as a preferred mode of the summer atmosphere, energized by barotropic conversion from the confluent mean flow, and identified oceanic precursors of AAC events unrelated to the El Niño–Southern Oscillation, extending the basis for seasonal prediction beyond the classic El Niño pathway. I then resolved why ENSO's monsoon impacts are robust after El Niño but unreliable after La Niña: La Niña events often persist for two to three years, and the concurrent cold state in the central Pacific opposes the delayed capacitor effect of the preceding winter, degrading the coherence of monsoon anomalies.

Related work quantified how tropical ocean dynamics contributed to the successive record-hot years of 2023 and 2024, and traced the unusual growth of the 2023–24 El Niño to ocean dynamics.

Zhang, Xie, Kosaka & Lutsko (2024), J. Climate [link]  ·  Zhang, Xie, Kosaka, Lutsko, Okumura & Miyamoto (2024), Nature Communications [link]  ·  Xie, Miyamoto, Zhang et al. (2025), npj Clim. Atmos. Sci. [link]  ·  Peng et al. (2025), Nature Geoscience [link]

Dynamical foundations

Underpinning both threads is a line of research on the tropical atmospheric circulation itself. I characterized and explained the seasonal superrotation of Earth's tropical upper troposphere, and I developed an axisymmetric single-layer model of the Hadley cell with continuously adjustable parameterized eddies, yielding quantitative scaling laws for how eddies and equatorial heating control the circulation's strength and extent.

The obliquity experiments in the Nature Geoscience study grew out of a conjecture at the end of the 2022 paper and overturned part of its premise — an example of the closed loop between my theoretical and modeling work. This grounding is what allows me to connect statistical relationships in coupled simulations to testable dynamical mechanisms.

Zhang, Lutsko, Hill & Xie (2025), J. Atmos. Sci. [link]  ·  Zhang & Lutsko (2022), J. Atmos. Sci. [link]

Where the threads converge: physics-based evaluation of AI climate models

Machine-learning models such as NeuralGCM now produce multi-decadal climate simulations at a small fraction of the cost of conventional GCMs, and the spread across their ensembles is increasingly read as internal variability. Whether that spread arises from correct dynamics has received little systematic testing. Skill, forced response, variance, and spectra establish that a model has plausible states and plausible variability amplitudes; none of them identifies what maintains the variability.

In a manuscript now in final preparation, a project I initiated during my postdoc, I develop a mechanism-based benchmark using the AAC as the test case. An ensemble of AMIP-like NeuralGCM simulations is evaluated not only on whether the AAC emerges as the leading mode of ensemble spread, but on whether it is maintained by barotropic energy conversion from the mean-flow confluence, with CESM2 atmosphere-only ensembles as a positive control and ERA5 as the observational reference. The dual criterion distinguishes a model that contains the right physics from one that merely reproduces a familiar pattern.

The convergence is methodological rather than topical. The monsoon thread supplies the test case and its energy source; the seasonal work established the habit of asking which process is responsible rather than which variables covary. Together they define a template I intend to extend to data-driven models of the coupled ocean–atmosphere system, where a mode must reproduce not only its observed pattern but also its propagation pathway and its source of memory.

Manuscript in preparation.

In progress

With Tiffany Shaw and Da Yang at the University of Chicago, I am examining why Earth's two hemispheres reflect nearly equal amounts of sunlight, focusing on the mediating role of baroclinic-eddy cloud responses.

A complete list of published work is on the publications page.