Aviation Technology

An Airline Taught a Machine to Feel Turbulence Before Anyone Else

On August 7, 2025, All Nippon Airways became the world’s first airline to implement an AI-based turbulence prediction service, developed by BlueWX in collaboration with Keio University. Built using deep-learning techniques and a decade of turbulence data, the model achieved a reported 86% prediction accuracy. Before formally integrating it into ANA’s weather-data infrastructure, the airline conducted operational trials involving 2,500 pilots, who supported its continued use — citing its alignment with flight reports and its effectiveness under real-world operating conditions.

But the announcement was not the beginning of the project. ANA and Keio University began developing the system in 2019, trial use started in 2021, and BlueWX was established in July 2023. By the time ANA announced full implementation in 2025, the technology had already passed through years of development, operational testing, pilot feedback, and continuous refinement.

The 86% figure may be the headline. The more important story is how ANA pursued it.

Six Years, Not Six Months

The timeline is the point.

2019 — ANA and Keio University begin a joint research initiative on turbulence prediction.

2021 — Trial use begins, with continuous updates improving the model’s accuracy and performance.

July 2023 — BlueWX is founded as an industry-academia collaboration between ANA and Keio University to commercialize the technology.

2025 — Following operational trials with 2,500 pilots, the system is formally integrated into ANA’s weather data infrastructure and announced publicly.

Six years from research to fleet-wide deployment. Four years of that spent with pilots flying against the model and reporting back.

In most industries, that pace would be considered a failure of execution. In aviation, it’s the process working correctly.

Why 2,500 Pilots Matter More Than 86 Percent

An accuracy figure is a laboratory statement. It says the model matched a dataset. It does not say the model is useful to a captain deciding whether to request a different altitude over the Pacific at 3 a.m.

What ANA did instead was expose the model to the population most capable of destroying it: 2,500 line pilots flying real routes in real conditions, comparing its forecasts against what the airplane actually did.

That’s a hostile testing environment in the best sense. Line pilots have no institutional stake in an algorithm succeeding, extensive pattern recognition built from thousands of hours, and immediate feedback on whether a prediction was right. A model that produces plausible outputs but doesn’t correspond to the atmosphere gets found out quickly.

The pilots reported that the system’s forecasts closely matched flight reports and outperformed traditional prediction methods in real-world operations. That endorsement is worth considerably more than the accuracy percentage, because it came from people whose judgment the model was attempting to inform rather than replace.

The Problem It’s Solving

Turbulence is not a comfort issue. According to NTSB data cited by the companies, turbulence accounts for 30 to 50% of aviation incidents — alongside economic costs including cargo and equipment damage, delays, and diversions.

It’s also getting worse. BlueWX and ANA have both framed the work against a backdrop of turbulence intensifying in part due to global warming, a concern increasingly reflected in atmospheric research on clear-air turbulence in particular.

Clear-air turbulence is the hardest case. It generates no radar return, produces no visual cue, and can occur in cloudless sky — which is why it’s disproportionately responsible for injuries to unbelted passengers and cabin crew. A prediction model that improves warning for exactly that category addresses the part of the problem existing tools handle worst.

Hiroyuki Kometani, ANA’s Executive Vice President of the Operation Division, framed turbulence prediction as a long-standing challenge for airlines — one the carrier set out to solve with current AI and deep-learning methods rather than incremental improvement of existing forecasting.

The Validation Principle

Here is what the ANA case actually demonstrates, and why it matters beyond one airline’s weather tooling.

A promising algorithm does not earn trust through a laboratory result or a polished demonstration. It earns trust when working pilots challenge it repeatedly in the operational environment — and it continues to prove useful.

That principle is currently being tested across aviation. AI is entering NOTAM interpretation, flight planning, air traffic management, maintenance prediction, and cockpit decision support. Each of those tools will arrive with a headline number attached. Very few will arrive with six years of development and 2,500 line pilots behind them.

The distinction between the two is not technical sophistication. It’s willingness to spend years finding out whether the thing works before telling anyone it does.

Aviation has learned this the expensive way, repeatedly. The industry’s safety record is built less on brilliant innovations than on the disciplined, unglamorous validation that follows them — the certification programs, the operational trials, the incident reporting, the iterative refinement that turns a good idea into something a crew can rely on at altitude.

What Comes Next

BlueWX has indicated the ANA rollout marks the beginning of global expansion to other airlines, and the company is developing wind speed and direction forecast models aimed at optimizing flight routes — work tied to the industry’s carbon neutrality target for 2050. Better wind forecasting is one of the more direct levers on fuel burn available to airlines.

Whether the turbulence model performs as well across other carriers, fleets, and route structures is an open question. A model trained substantially on one airline’s data operating in one region may or may not generalize. That’s a legitimate uncertainty, and the appropriate response is the same one ANA applied: trial it, expose it to pilots, and find out.

For general aviation pilots, the direct relevance is limited today — this is an airline operations tool integrated into airline weather infrastructure. But the trajectory is worth noting. Turbulence forecasting improvements tend to propagate. Better atmospheric modeling eventually reaches the weather products in flight planning apps, and the AI methods proving out in airline operations are the same ones being applied to the tools GA pilots use.

The Bottom Line

ANA’s turbulence prediction system is a genuine achievement: the world’s first AI-based turbulence forecasting service in airline operations, at 86% accuracy, addressing a hazard responsible for a large share of aviation incidents and one that appears to be worsening.

But the achievement worth studying isn’t the model. It’s the method — six years of development, four years of trials, 2,500 pilots challenging the output against what the airplane actually did, and continuous refinement before anyone announced anything.

Technology may accelerate innovation. Disciplined validation is what turns innovation into aviation safety.

Frequently Asked Questions

What is ANA’s AI turbulence prediction system? On August 7, 2025, All Nippon Airways became the first airline in the world to implement an AI-based turbulence prediction service, developed by BlueWX in collaboration with Keio University. The model uses deep learning techniques trained on more than a decade of turbulence data and achieved a reported 86% prediction accuracy. It is integrated into ANA’s weather data infrastructure to support flight planning and operations.

How was the turbulence prediction model validated? ANA conducted operational trials involving 2,500 of its pilots before formally integrating the system. Pilots compared the model’s forecasts against actual flight reports and reported that it closely matched real-world conditions and outperformed traditional prediction methods. Development began in 2019 as a joint initiative between ANA and Keio University, with trial use starting in 2021 and continuous refinement through to the 2025 rollout.

Who is BlueWX? BlueWX Company Limited is a Tokyo-based company founded in July 2023 through an industry-academia collaboration between All Nippon Airways and Keio University. It specializes in AI-driven weather forecasting solutions for aviation. Following the ANA implementation, the company has indicated plans to expand the turbulence prediction system to other airlines globally, and is developing wind speed and direction forecast models to support route optimization.

How significant is turbulence as an aviation safety issue? Turbulence accounts for an estimated 30 to 50% of aviation incidents according to NTSB data cited by ANA and BlueWX, and also generates economic costs including cargo and equipment damage, delays, and diversions. Clear-air turbulence is particularly difficult to forecast because it produces no radar return and no visual cue. Both companies have framed the work against evidence that turbulence is intensifying, in part due to global warming.

Does AI turbulence prediction affect general aviation pilots? Not directly at present — the ANA system is an airline operations tool integrated into airline weather infrastructure. However, improvements in atmospheric modeling tend to propagate over time into the weather products available through general aviation flight planning applications, and the AI methods being validated in airline operations are increasingly applied to tools used across general aviation.

Sources: