In a groundbreaking achievement for meteorological science, researchers from Japan's leading universities have harnessed the power of the Fugaku supercomputer to simulate the full lifecycle of a typhoon—from a nascent weak vortex to a devastating super typhoon—using an unprecedented horizontal resolution of 100 meters. This ultra-high-resolution large eddy simulation (LES), published in Geophysical Research Letters on January 14, 2026, marks the first time scientists have directly computed the intricate dynamics of small-scale eddies that play a pivotal role in the storm's rapid intensification.
This feat not only illuminates the previously elusive mechanisms behind typhoon evolution but also holds profound implications for improving intensity forecasts in Japan, a nation frequently battered by these powerful storms. Typhoons, known locally as台風 (taifū), cause billions in damages annually, with events like Typhoon Hagibis in 2019 resulting in over $15 billion in losses and dozens of fatalities. Accurate prediction of when and how intensely a typhoon strengthens could save countless lives and safeguard infrastructure.
The Fugaku Supercomputer: Japan's Computational Powerhouse
Fugaku, developed jointly by RIKEN—a premier national research institute—and Fujitsu, held the title of the world's fastest supercomputer from 2020 to 2022. Powered by Fujitsu's A64FX ARM-based processors, it boasts a theoretical peak performance of 442 petaflops and has been instrumental in diverse fields from COVID-19 drug discovery to materials science. In meteorology, Fugaku's ability to handle massive datasets at extreme resolutions enables simulations unattainable on lesser machines.
For this typhoon study, researchers configured a 2,000 km by 2,000 km domain with 60 vertical layers, totaling around 1.2 billion grid points at 100-meter spacing. This demanded enormous computational resources—equivalent to modeling every cubic meter of airspace explicitly for turbulent eddies. Traditional weather models operate at kilometers-scale, parameterizing sub-grid processes like turbulence, but Fugaku's prowess allowed direct numerical simulation of large eddies while sub-filter scale effects were modeled.
Fugaku's contributions extend beyond typhoons; it has simulated urban tornadoes within typhoon systems in collaboration with Yokohama National University and Fujitsu, achieving real-time predictions. These advancements underscore Japan's strategic investment in high-performance computing to tackle climate-related disasters.
Understanding Typhoon Rapid Intensification: A Persistent Forecasting Challenge
Rapid intensification (RI) occurs when a tropical cyclone's maximum sustained winds increase by at least 30 knots (about 55 km/h) within 24 hours. This phenomenon is notoriously difficult to predict because it involves multiscale interactions—from planetary waves to convective bursts and vortex dynamics. In Japan, RI typhoons like Faxai (2019) and Hagibis amplified destruction by unexpectedly strengthening just before landfall.
Operational models like the Japan Meteorological Agency's (JMA) MSM (Meso-Scale Model) at 5 km resolution struggle with RI timing, often predicting onset too early. This study addresses that gap by resolving processes at 100 meters, where small-scale turbulence governs energy transfer to the core.
Globally, RI accounts for 70-80% of major hurricane landfalls in the Atlantic, per NOAA data. In the Northwest Pacific, where Japan lies, super typhoons (winds > 240 km/h) form rapidly, exacerbating risks amid climate change, which is projected to increase RI frequency by 5-20% per IPCC AR6.
Unveiling the Simulation: Methods and Technical Feats
The team employed a cloud-resolving nonhydrostatic model in LES mode, initialized with a weak axisymmetric vortex over idealized warm ocean waters (26°C SST). The domain spanned 2000 km, capturing the full storm evolution over four days. Fugaku processed this at 100 m horizontal and variable vertical resolution, explicitly resolving eddies down to the grid scale.
Key parameters included a Coriolis parameter for 20°N latitude, Rayleigh damping aloft, and large-scale nudging for ambient conditions. Computations ran for 120 hours, outputting diagnostics like minimum sea-level pressure (MSLP), max wind, and vorticity fields.
Comparisons used a 2 km reference simulation with the same model (Meso-NH?). Both reached similar maturity (MSLP ~920 hPa, max winds ~75 m/s), but RI timing diverged sharply.The full paper details the model's physics and validation.
Key Findings: The Lifecycle from Vortex to Super Typhoon
The 100 m LES faithfully reproduced the canonical stages: spin-up of the primary circulation, formation of a ring of convection, eyewall establishment, and explosive RI leading to a compact super typhoon. Vorticity fields revealed intense counter-clockwise rotation near the surface eye, with clockwise anticyclonic eddies encircling it—forming a polygonal eyewall characteristic of intense TCs.
RI, defined here as MSLP drop >30 hPa/24h, commenced at t=68 hours in LES vs. t=42 hours in 2 km run. Final intensity matched observations of real super typhoons.
Photo by BoliviaInteligente on Unsplash
The Critical Role of Small Eddies in Delaying RI
High-resolution revealed two eddy populations hindering inflow: sub-kilometer 'tiny eddies' (<1 km radius) generated by surface friction and shear, and mesovortices (~10 km) around the eyewall. These created a 'blockage'—divergent outflow aloft suppressed convection spin-up.
In 2 km models, these eddies are parameterized away, allowing premature axisymmetrization and RI. LES showed persistent asymmetry until t~68h, when eddies dissipated, enabling rapid deepening.
This multiscale view explains RI forecast errors: resolving eddies ~O(100m) is essential for timing.
Comparison with Low-Resolution Models and Observations
- Intensity Peak: Both sims ~920 hPa MSLP, 75 m/s winds—matches JMA/MSM for idealized RI cases.
- RI Timing: Delayed 26h in LES; coarser models overpredict early strengthening.
- Structure: LES eyewall radius ~20 km, polygonal; observed in Hurricane Patricia (2015).
- Observation Proxy: Dropsonde/recon data from NW Pac TCs validate vorticity patterns.
While idealized, results align with real TC RI composites (e.g., Rogers et al. 2013).JST press release provides figures.
The Collaborative Research Team and Institutional Synergy
Led by Associate Prof. Junshi Ito (Tohoku University Graduate School of Science), with Yuta Sakurai and Leia P.S. Tonga (Tohoku grads), Hiroshi Niino (U Tokyo Atmosphere and Ocean Research Institute, emeritus prof.), and Yoshihito Miyamoto (Keio University SFC Associate Prof.). This interdisciplinary effort pooled expertise in cloud-resolving modeling, TC dynamics, and HPC.
Tohoku U's fluid dynamics group, U Tokyo's AORI (world-class ocean-atmos lab), and Keio SFC's computational innovation drove the project. Funded via Fugaku priority projects, it exemplifies Japan's university ecosystem fostering cutting-edge research.
Implications for Typhoon Forecasting and Disaster Preparedness
Japan faces 25-30 typhoons yearly, 3-4 landfalling, causing ¥1 trillion+ damages/decade. RI misforecasts amplify evacuation failures. LES insights suggest hybrid models incorporating explicit small-eddy physics could refine ECMWF/JMA RI probs.
Operational shift to km/100m nests on Fugaku successors (Fugaku-next?) viable. For climate: Higher-res GCMs needed for RI frequency under warming (e.g., +10-20% by 2100).
Broader Impacts: From Academia to Policy
This work advances LES frontiers, previously limited to ~1 km domains. Applications: Tornado/eyewall burst nowcasting, urban wind hazards. Universities like Tohoku training next-gen modelers via such projects.
Policy: MEXT/RIKEN prioritizing weather HPC aligns with SDGs 13 (climate action). Global collab potential with NOAA/JMA.
Future Directions and Ongoing Challenges
Next: Real-case LES (e.g., Typhoon Faxai), coupling ocean/terrain, ensemble RI probs. Challenges: Exascale compute for global LES, data assimilation for ops. Fugaku's legacy ensures Japan's lead.
Universities eye AI-accelerated LES for faster turnaround.
