A computational experiment suggests that getting the Tibetan Plateau’s land-surface temperature right may dramatically change how climate models simulate California’s most extreme winter precipitation events.
What if a supercomputer trying to understand California’s winter precipitation was looking in the wrong place?
That is the intriguing possibility raised by new research published in Science Advances. A team led by Yongkang Xue at the University of California, Los Angeles, used numerical weather and climate simulations to investigate two extraordinary California precipitation seasons, winter 2016–2017 and winter 2022–2023, and found that a seemingly remote piece of the atmosphere-land system may have played an important role: unusually strong early-winter heating over the Tibetan Plateau.
The computational experiment is particularly interesting from a high-performance computing perspective because the researchers did not simply ask a climate model to reproduce what happened. They used controlled ensemble simulations to ask a much harder question:
What happens to California’s precipitation when the model’s representation of Tibetan Plateau land-surface temperature is changed?
The answer was striking.
After correcting the Tibetan Plateau temperature initialization, the simulations reproduced approximately 56% of the observed January 2017 extreme precipitation anomaly and 38% of the March 2023 anomaly over California and adjacent regions.
The result does not mean a supercomputer has discovered a single variable that can perfectly predict California floods. The researchers explicitly describe the work as a single-model case study and call for multimodel investigations.
But it does demonstrate something potentially more consequential for computational Earth-system science: model initialization can determine whether a remote physical mechanism becomes visible at all.
The computational problem: California was not supposed to behave this way
California’s winter precipitation is strongly influenced by large-scale atmospheric circulation and atmospheric rivers, long, narrow corridors of concentrated water vapor that can transport enormous quantities of moisture toward the West Coast.
Yet the winters Xue and colleagues examined presented an interesting forecasting puzzle.
Both 2016–2017 and 2022–2023 occurred during La Niña conditions, which are traditionally associated with relatively dry conditions in California. Nevertheless, both periods produced extraordinary precipitation.
That raises a fundamental computational question.
If a model is initialized with the observed state of the climate system, why can’t it reproduce the extreme precipitation?
The researchers approached the problem with numerical experiments using the National Centers for Environmental Prediction Global Forecast System, coupled with the second-generation Simplified Simple Biosphere land-surface model, known as GFS/SSiB2.
The atmospheric model was run at T126L64 resolution, corresponding to approximately 100 × 100 kilometers horizontally, with 64 vertical levels extending to 2 hPa.
That is nowhere near the kilometer-scale resolution increasingly used for specialized regional simulations. But at global-climate scale, the computational domain is enormous, and the model must represent atmospheric circulation, land-surface processes, ocean conditions, and interactions across the entire planet.
And the researchers weren’t running one simulation.
They were running ensembles.
Ten computers’ worth of possibilities, or more accurately, ten model realizations
The control experiments, designated CTRL2017 and CTRL2023, were initialized using land-surface and atmospheric information from the NCEP Climate Forecast System Reanalysis.
This included variables such as soil moisture, land temperature, and snow cover.
Each experiment consisted of a 10-member ensemble, allowing the researchers to examine the modeled response while reducing the influence of individual realizations of internal atmospheric variability.
This is one of the fundamental reasons HPC matters in modern climate research.
A single simulation gives researchers one trajectory through an enormously complicated nonlinear system.
An ensemble gives them a small computational population of alternative trajectories.
The distinction matters because atmospheric dynamics are chaotic. Tiny differences in initial conditions can grow rapidly, making it difficult to determine whether a particular event results from a predictable external influence or simply from the system’s internal variability.
The control experiments provided an important warning.
They did not reproduce the California precipitation extremes particularly well.
And they also exhibited substantial errors in Tibetan Plateau temperature.
That coincidence became the computational clue.
The model may have been initialized incorrectly where nobody was looking
The Tibetan Plateau is thousands of kilometers from California.
At first glance, changing its land temperature might seem unlikely to affect precipitation on the other side of the Pacific.
But the atmosphere doesn’t respect political or continental boundaries.
Large-scale heating anomalies can alter pressure fields and atmospheric circulation, generating planetary-scale wave responses that propagate through the atmosphere.
The researchers therefore designed another set of experiments.
Rather than simply accepting the model’s initial Tibetan Plateau temperature state, they modified the land temperature over the plateau using observed monthly mean anomalies and model errors relative to the 1980–2023 period.
The resulting experiments were designated LT2017 and LT2023.
Again, each consisted of a 10-member ensemble.
The goal was not merely to make the model produce more California precipitation. It was to test whether correcting the Tibetan Plateau’s thermal state could activate a physically plausible chain of atmospheric responses connecting Asia to North America.
And that is where the experiment became particularly interesting.
Follow the wave
The simulations point toward a large-scale atmospheric wave train connecting the Tibetan Plateau and the Rocky Mountain region.
The proposed sequence is approximately:
Tibetan Plateau heating → planetary-scale wave response → Rocky Mountain circulation → northeastern Pacific circulation → atmospheric-river modulation → California precipitation.
The mechanism involves changes in the large-scale atmospheric circulation and subsequent Rossby wave breaking over the northeastern Pacific and western North America.
In other words, the model wasn’t simply saying:
“Tibet got warmer, therefore California got wetter.”
The computational hypothesis was considerably more complicated.
Heating over the plateau altered the atmospheric circulation. That circulation generated a wave train extending downstream. The resulting circulation changes modified the environment in which atmospheric rivers formed and propagated toward the West Coast.
That provided a dynamical pathway by which a land-surface anomaly thousands of kilometers away could influence precipitation over California.
Atmospheric rivers become the computational messenger
The atmospheric-river component provides another useful HPC diagnostic.
Researchers examined changes in integrated vapor transport (IVT) and integrated moisture flux convergence (IMFC), quantities that help describe how atmospheric rivers transport and concentrate water vapor.
For the March 2023 case, the simulations indicated approximately a 15% enhancement in IVT and about a 30% increase in integrated moisture flux convergence associated with the Tibetan Plateau-induced wave response.
The January 2017 experiment showed an even stronger response in the relevant atmospheric-river diagnostics, with IVT increasing from approximately 90.8 to 126.0 kilograms per meter per second, while moisture-flux convergence increased by roughly 62%.
Those changes matter because atmospheric rivers are not simply atmospheric plumbing carrying moisture toward California.
Their impacts depend on where and how that moisture transport interacts with the larger-scale circulation.
A relatively modest change in moisture transport can therefore become consequential if the atmospheric circulation simultaneously changes where the moisture is concentrated and where it is forced upward.
That is exactly the kind of nonlinear interaction that numerical experiments are designed to expose.
The surprise wasn’t more computing power. It was better initialization.
There is a subtle HPC lesson buried inside this result.
When a model fails to reproduce an extreme event, the obvious response is often to ask whether the simulation needs higher resolution, a more sophisticated physical parameterization, a larger ensemble, or simply more computational horsepower.
Those are legitimate questions.
But this experiment points toward another possibility: The model may have enough computing power. It may simply have been given the wrong starting state.
The researchers’ control experiments contained substantial Tibetan Plateau temperature errors.
Once the land-temperature initialization was adjusted, the simulated atmospheric response changed substantially, and the model reproduced a significant fraction of the observed California precipitation anomalies.
This is a reminder that the computational pipeline for Earth-system modeling is not simply:
More FLOPS → better prediction.
It is closer to:
Observations → data assimilation/reanalysis → initialization → ensemble generation → numerical integration → diagnostics → physical interpretation.
If the initial state is wrong in a strategically important part of the Earth system, throwing additional floating-point operations at the simulation does not necessarily fix the problem.
The supercomputer can calculate the wrong answer extraordinarily accurately.
Why this matters for predictive skill
Seasonal-to-subseasonal prediction sits in an awkward computational space.
Weather forecasts operate over relatively short periods, while conventional climate projections examine much longer timescales.
Between them lies a difficult regime in which researchers want to know whether a particular atmospheric state provides useful predictive information weeks or months in advance.
California winter precipitation is particularly challenging because extreme events can depend on interactions among ocean conditions, atmospheric circulation, land-surface states, snow, moisture transport and internally generated atmospheric variability.
The researchers argue that the Tibetan Plateau may provide one previously underappreciated source of predictability.
That is potentially significant because land-surface conditions are among the components of the Earth system that can carry memory forward in time.
Soil temperature, soil moisture and snow conditions don’t necessarily reset instantly when the atmosphere changes.
They can therefore become part of the initial-condition problem for subseasonal-to-seasonal prediction.
A supercomputer as a laboratory
Perhaps the most interesting aspect of the study is that the computer simulation is functioning less like a forecasting machine and more like a laboratory.
Scientists cannot experimentally heat the Tibetan Plateau and wait to see what happens to California.
But they can construct a numerical world in which the Tibetan Plateau temperature is altered while attempting to hold other aspects of the experiment sufficiently controlled to isolate the response.
That allows them to ask a counterfactual question:
If the Tibetan Plateau had been initialized differently, would the downstream atmospheric circulation have evolved differently?
The answer from these simulations is yes.
The experiment therefore moves beyond correlation.
The researchers had previously observed statistical relationships between Tibetan Plateau conditions and downstream atmospheric behavior. The numerical experiments provide a way to investigate whether the proposed relationship is dynamically plausible.
That is a fundamentally computational form of scientific experimentation.
But don’t declare victory yet
There is an important caveat, and the paper itself emphasizes it.
This was a single-model case study.
The results are therefore model-dependent, and the authors say multimodel studies will be necessary to determine how robust the mechanism is.
The Tibetan Plateau heating mechanism also explains only part of the observed precipitation anomalies. Other processes, including internal atmospheric variability and changes involving snow, vegetation and soil moisture, may contribute as well.
That distinction is critical.
The study does not establish that Tibetan Plateau heating is the explanation for California’s extreme precipitation.
It establishes that, within this modeling framework, correcting Tibetan Plateau temperature initialization produces a substantial downstream response and reproduces a meaningful portion of the observed anomalies.
That’s a much more interesting scientific result than a simplistic claim of causation.
The next HPC experiment could be even bigger
The logical next step is not necessarily another 10-member ensemble.
It is broader computational experimentation.
Multiple atmospheric models.
Multiple land-surface models.
Higher spatial resolutions.
Larger ensembles.
Different initialization systems.
Longer hindcast periods.
And, critically, many more extreme precipitation cases.
If the same Tibetan Plateau–Rocky Mountain wave pathway appears across independent models, the evidence for a robust mechanism becomes much stronger.
If it disappears in some models, that would be equally valuable information.
It would tell researchers where model physics, land-surface initialization, resolution, or atmospheric dynamics are influencing the result.
That is where HPC becomes more than an accelerator.
It becomes the experimental apparatus.
From Tibet to California, one initialization variable changes the question
The broader implications of this study are profound. A global climate model functions as a complex dynamical system, and its output is fundamentally contingent upon its initial conditions. In these experiments, a temperature bias over the Tibetan Plateau inhibited the model's ability to replicate extreme precipitation events thousands of miles away. By correcting this initialization, researchers successfully induced an atmospheric wave train that significantly altered circulation patterns and moisture transport, ultimately leading to a more accurate representation of California's precipitation anomalies.
The computer functioned as more than a simple forecasting instrument; it enabled researchers to conduct a counterfactual experiment that would be impossible to replicate in the physical world. This finding raises a compelling question for the next generation of supercomputing-based Earth-system models: how many events currently categorized as unpredictable might actually be foreseeable, provided the models are initialized correctly in the regions that have previously been overlooked? For HPC researchers, this may prove to be the most significant implication of the study.