A new study published in the journal Climate Dynamics offers a fresh path forward. A team of researchers led by Shu Gui, Dayong Wen, and Jie Cao of Yunnan University, together with colleagues including Liang Duan, has shown that the atmospheric heat source over the Southeast Asian low-latitude highlands, a region abbreviated as SEALLH, can serve as a powerful early warning signal for the timing of the Meiyu onset. By combining this physical precursor with machine learning models trained in an unconventional way, the team achieved hindcast skill that outperformed more direct approaches, and their results highlight a subtle but crucial lesson for anyone applying artificial intelligence to climate prediction: a model that predicts well but predicts for the wrong physical reasons is a model that cannot be fully trusted.

The Southeast Asian low-latitude highlands encompass the elevated terrain of the Indochinese Peninsula and adjacent regions, sitting south and southwest of the better-known Tibetan Plateau. Like the plateau itself, these highlands act as an elevated heat source in the boreal spring and early summer. As the sun climbs northward, the land surface warms, convection intensifies, and enormous quantities of latent and sensible heat are released into the middle and upper troposphere. This diabatic heating, the net heating of an air parcel by processes other than direct radiation, including condensation of water vapor in deep convective clouds and turbulent transfer from the surface, helps drive the large-scale circulation transitions that eventually pull the East Asian summer monsoon northward and set up the Meiyu front over the Yangtze and Huaihe river valleys.

Previous work by members of the same team had already established that anomalies in the atmospheric heat source over these highlands in late spring are closely tied to year-to-year swings in the Meiyu onset date. When the heating is anomalously strong or weak, it alters the position and intensity of the western North Pacific subtropical high, the great anticyclonic circulation that steers moisture-laden air toward eastern China. It also modulates moisture convergence over East Asia, effectively setting the stage, months in advance, for whether the Meiyu rain band will arrive early or late. What remained unclear was whether this promising physical linkage could actually be converted into quantitative prediction skill, and if so, how best to do it.

The researchers turned to the interannual increment approach, a strategy that has gained traction in seasonal climate prediction. Rather than training models to predict the onset date itself, which varies within a fairly narrow window and carries strong decadal trends that can mislead statistical learning algorithms, the method predicts the year-to-year increment, the difference between the onset date in one year and the previous year. This differencing strips away slowly varying background signals and focuses the learning problem on the genuine interannual variability that forecasters care about, often yielding sharper and more robust prediction models.

To build their prediction system, the team trained five categories of machine learning models using historical simulations from the Coupled Model Intercomparison Project Phase 6, the international archive of state-of-the-art climate model experiments. Training on climate model output rather than observations alone gives the algorithms a much larger sample of physically consistent atmospheric behavior to learn from, helping to guard against overfitting to the short observational record. The trained models were then validated against ERA5, the high-resolution reanalysis product from the Copernicus Climate Change Service that blends observations with a numerical weather prediction model to provide a best estimate of the historical atmospheric state. The predictors were the diabatic heating anomalies over the Southeast Asian low-latitude highlands, and the target was the Meiyu onset date increment.

The team designed two distinct hindcasting strategies and pitted them against each other. In the first group, models directly predicted the interannual increment of the Meiyu onset date in a single step. In the second group, the models instead predicted the onset dates of two consecutive years separately, and the increment was then derived by taking the difference between the two hindcasts. This seemingly small architectural difference turned out to matter enormously, and the reason lies in the physics that the models implicitly learned.

When the researchers examined what the best-performing models were actually doing, they found that the top models in both groups better captured the real-world chain of cause and effect linking the highland heating to large-scale circulation anomalies, moisture convergence, and ultimately the onset date, compared with the bottom-performing models. But the top models in the direct-prediction group harbored a hidden flaw: their predicted 500-hectopascal zonal wind anomalies, the east-west winds in the mid-troposphere that are central to the Meiyu circulation, were physically inconsistent with the observed relationships. In other words, these models could reproduce the onset date reasonably well while violating the very circulation physics that makes the prediction meaningful. Such inconsistency is a red flag for any operational forecast system, because a model that gets the right answer through the wrong mechanism is unlikely to remain reliable when conditions shift outside its training experience.

Remarkably, the two-step strategy largely corrected this defect. By hindcasting two consecutive onset dates and differencing them, the second group of models produced predictions whose associated circulation anomalies were far more physically coherent, and this coherence translated directly into skill. The top-performing models in the second group achieved a temporal correlation coefficient of 0.62 between hindcast and observed onset date increments, compared with 0.52 for the best models in the direct group. In the world of seasonal climate prediction, where correlation coefficients above 0.5 are often considered useful, this improvement is substantial. The study also found that the Extra Trees model, an ensemble method that builds many randomized decision trees and averages their outputs, delivered robust performance not only for predicting the current year’s onset but also for hindcasts made one year ahead, a lead time at which most dynamical seasonal models offer little guidance for the East Asian monsoon.

The implications extend well beyond the Meiyu. The work demonstrates that diabatic heating over the Southeast Asian low-latitude highlands deserves a place among the standard precursors used for East Asian summer monsoon prediction, complementing the traditional emphasis on sea surface temperature anomalies in the tropical Indian and Pacific oceans. Just as importantly, it offers a template for how machine learning should be deployed in climate science. Rather than judging models solely by their prediction scores, the researchers evaluated whether the models preserved known physical linkages among heating, circulation, and rainfall, and they showed that enforcing this physical consistency through clever problem design can itself be the key to better forecasts. As machine learning floods into weather and climate prediction, from nowcasting to decadal projection, the lesson from the Yunnan team is clear: the most skillful artificial intelligence is the kind that understands, or at least respects, the physics of the atmosphere it is trying to predict. For the farmers, dam operators, and millions of residents of the Yangtze valley, that combination of highland heat and disciplined algorithms may soon mean a more reliable heads-up before the rains arrive.

Subject of Research: Seasonal prediction of the Meiyu onset date using diabatic heating anomalies and machine learning

Article Title: Skillful prediction of Meiyu onset using diabatic heating over Southeast Asian low-latitude highlands

Article References: Gui, S., Wen, D., Cao, J., Dong, Z., Cai, L., Yang, R., & Duan, L. (2026). Skillful prediction of Meiyu onset using diabatic heating over Southeast Asian low-latitude highlands. Climate Dynamics, 64(10), Article 412. https://doi.org/10.1007/s00382-026-08366-y

Image Credits: AI Generated

DOI: 10.1007/s00382-026-08366-y

Keywords: Meiyu onset, diabatic heating, Southeast Asian highlands, machine learning, East Asian monsoon, Climate Dynamics, interannual increments, CMIP6, ERA5, Extra Trees, seasonal prediction, Yangtze River valley