Two brothers working 7,400 kilometers apart have quietly built what may be the most efficient clear-air turbulence prediction system ever demonstrated. PSTNet — a physics-structured neural network with just 552 trainable parameters — can ingest a single NASA weather observation and produce a multi-altitude turbulence intensity map in under seven seconds on consumer hardware.
The system, developed by Boris Kriuk at the Hong Kong University of Science and Technology (HKUST) and Fedor Kriuk at the University of Technology Sydney (UTS), was shown this week running live over the Central Asia–Himalaya corridor, one of the most turbulence-prone airspaces on Earth. It correctly identified severe cells at tropopause altitudes, terrain-blocked zones and the characteristic turbulence minimum in the lower stratosphere — all from a data footprint smaller than a single Excel spreadsheet.
The timing is not accidental. Clear-air turbulence, invisible to radar, has increased measurably over major flight corridors in recent years, with climate research linking the trend to strengthening wind shear in the jet stream. Airlines reported a 17% increase in turbulence-related injuries in 2025 compared to the prior five-year average, according to IATA data. Yet current operational turbulence forecasts — primarily the Graphical Turbulence Guidance (GTG) system and its international equivalents — rely on numerical weather prediction models that run on supercomputer clusters and update on multi-hour cycles.
PSTNet takes a fundamentally different approach.
How It Works
The model trains from scratch on each run. Using NASA POWER satellite-derived weather data — wind speed, temperature, pressure — combined with terrain elevation gradients, PSTNet generates training samples based on known turbulence physics: Richardson number thresholds, mountain wave parametrizations, and vertical wind shear relationships. It then trains for 200 epochs on roughly 1,000 samples and produces turbulence intensity estimates across a 24×24 grid at eight standard flight levels from FL100 (10,000 ft) to FL450 (45,000 ft).
The result is a 576-cell heatmap at each altitude that a pilot, dispatcher, or route-planning algorithm can read instantly.
“We deliberately constrained the model to 550 parameters because we wanted to prove that physics does the heavy lifting, not scale,” said Boris Kriuk, who leads the atmospheric modeling and architecture design at HKUST. “When your network is structured around the actual regimes of the atmosphere, you don’t need millions of parameters to avoid producing nonsense. The constraints are the intelligence — the network just learns the residuals that pure theory can’t capture.”
Why It Matters Beyond Aviation
The dual-use implications are what elevate PSTNet from a clever demo to a strategically significant piece of work. Turbulence estimation sits at the intersection of aviation safety, drone operations, defense trajectory planning, and climate monitoring. A model that runs on a laptop in under seven seconds, with no internet dependency beyond an initial weather API call, can be deployed in environments where cloud-connected supercomputer forecasts are simply unavailable — remote airfields, maritime operations, or forward-deployed military units.
“The core realization was that estimating turbulence for a flight path and estimating turbulence for a ballistic trajectory are the same mathematical problem — you need intensity at a point given sparse atmospheric state,” said Fedor Kriuk, who leads the systems engineering and real-time deployment pipeline at UTS Sydney. “Once we framed it that way, the architecture wrote itself. We built one model that serves both problems, and it runs anywhere — a laptop, an edge device, a ship.”
The system currently visualizes results through a specialised interface with severity legends, and interactive flight level selection. The team has indicated plans to validate PSTNet against pilot reports (PIREPs) and eddy dissipation rate (EDR) measurements from equipped commercial aircraft — the standard benchmark for turbulence prediction accuracy.
What Comes Next
Independent atmospheric scientists who reviewed the output noted that the vertical turbulence profile — high near terrain, moderate through the troposphere, peaking again at jet stream altitudes, and dropping sharply in the stratosphere — is physically consistent and non-trivial to reproduce with so few parameters.
The open questions remain validation depth and temporal evolution. A single snapshot is useful; a continuously updating nowcast would be transformational. The Kriuk brothers say both are on the roadmap.
For now, 552 parameters, 6.5 seconds, and one weather observation have produced something the aviation industry has spent decades and billions trying to achieve: a turbulence map that makes physical sense, runs anywhere, and costs nearly nothing.
The next step is systematic validation against pilot encounter reports and onboard EDR measurements — the standard path from research demonstration to operational certification. If the accuracy holds at scale, PSTNet could redefine what’s possible in lightweight atmospheric prediction.
