University of Tokyo reports finer global model removes artificial rain bursts

The University of Tokyo reported on October 8 that a global atmospheric simulation run on Japan’s Fugaku supercomputer eliminated unrealistic, concentrated bursts of rain by narrowing its grid spacing to 220 meters. The result could help researchers evaluate weather and climate models, but the team says its computing demands are far too high for everyday forecasting.

The university’s report, published by Phys.org, describes an improvement in how a model represents rainfall, rather than a demonstrated improvement in public forecasts or warning times. Related results from the same researchers were already presented in a 2024 conference abstract, making the latest announcement a report on work with a documented public history.

What a 220-meter grid changes

Atmospheric models divide the globe into a grid to calculate weather and climate processes. According to the university, conventional global climate models typically use spacing of tens to hundreds of kilometers. Global storm-resolving models bring that down to roughly one to ten kilometers, allowing more detailed investigation of clouds and storms.

The Tokyo experiment takes that spacing down to 220 meters. Shuhei Matsugishi, a project researcher at the university’s Atmosphere and Ocean Research Institute, said the finer grid represents individual rising and descending air currents inside convective clouds more explicitly. Running such a simulation globally lets researchers examine how clouds interact with one another and with larger atmospheric circulation.

Eight simulated hours required much of Fugaku

The computational scale was enormous. According to the university, the experiment represented nearly one trillion three-dimensional grid points. Simulating eight hours of atmospheric conditions for August 5, 2016 required using more than half of Fugaku simultaneously. The simulated date is historical, separate from the October 8 report.

“At present, a global simulation at 220-meter (720-foot) resolution is far too computationally expensive to replace operational weather forecasting systems,” Matsugishi said. The demonstration therefore remains a research experiment rather than an announced replacement for the systems producing routine forecasts.

Matsugishi identified potential uses including studying tropical convection and heavy rainfall, and providing a detailed reference against which coarser climate models can be evaluated and improved. These are research applications the team proposes; the report does not establish an implementation timetable for operational forecasting.

The university also reports that biases in the distribution of cloud cover remained. Removing one rainfall error therefore does not mean that all important model errors disappeared. The experiment’s eight-hour window also leaves open how the results would extend to other weather situations and longer simulations.

Related rainfall results were public in 2024

A 2024 European Geosciences Union conference abstract by Matsugishi, Tomoki Ohno and Masaki Satoh described the same eight-hour, 220-meter global simulation. It identified the model as the Nonhydrostatic Icosahedral Atmospheric Model, or NICAM, and reported experiments at grid spacings of 3.5 kilometers, 1.7 kilometers, 870 meters, 440 meters and 220 meters.

The researchers initialized finer experiments from coarser simulations. Their abstract reported that rainfall covered a wider area as resolution increased, with smoother rainfall distributions when averaged over larger areas. It also found that changes in high clouds depended on the turbulence scheme, showing that grid size was not the only consequential modeling choice.

Independent research shows why resolution is not enough

Separate research by Stony Brook University’s Joonghyun In and Marat Khairoutdinov provides context for those limitations. Their March 23, 2026 study evaluated 13 DYAMOND-Winter global storm-resolving models using grids of five kilometers or finer to simulate January–February 2020. It did not test Tokyo’s 220-meter experiment.

Comparisons with observational and reanalysis products found substantial differences between models and persistent systematic errors, particularly over land. Models often exaggerated extreme rainfall, although they improved most atmospheric fields compared with selected coarser climate-model experiments over the study’s short comparison period. The authors concluded that higher resolution alone does not automatically eliminate longstanding biases.

The Tokyo team plans further work on convective clouds, including how turbulence, cloud microphysics and processes that remain unresolved should be represented as global grids become finer. Reducing uncertainty in weather and climate predictions is the eventual aim; the current result establishes a reported improvement in a limited research simulation.