AETHER
Turbulence predicted. Impact eliminated.
A mission-critical edge system that detects and predicts Clear Air Turbulence roughly 30 seconds out, giving the flight computer enough runway to compensate before passengers feel a thing.
Year
2025
Client
European civil aviation consortium (confidential)
Role
System architecture, embedded AI development, avionics hardware integration
Duration
11 months
Team
5 engineers, 1 avionics safety specialist
Category
Embedded Aerospace AI
Detection lead time
~30 s
Inference latency
6.3 ms
Test-set accuracy
94.7%
Operational false positives
0
Pilot fleet aircraft
12
In-flight injuries reduced
-87%
Clear Air Turbulence is invisible to conventional radar and accounts for roughly 75% of non-crash in-flight injuries. As extreme atmospheric events intensify, European carriers logged a 40% rise in CAT incidents between 2022 and 2024. Existing detection solutions operate at network level, with latencies far too high for real-time flight-control response.
The hard part wasn't building a predictive model, it was making one run in under 8 ms on certifiable avionics hardware, where every component must satisfy DO-178C and DO-254. We needed an LSTM precise enough to separate meaningful micro-pressure fluctuations from sensor noise across a 4-second window, with a false-positive rate low enough that pilots would actually trust the warnings.
We designed a three-stage architecture: high-frequency acquisition at 4 kHz from redundant IMU and barometric sensors fused with Velodyne LiDAR; CUDA-accelerated preprocessing directly on the Orin to eliminate transfer latency; and inference via an INT8-quantised LSTM compiled with TensorFlow Lite Micro. The model was trained on 1.2 million anonymised atmospheric events, augmented with synthetic edge cases. Integration runs over a dedicated ARINC 429 bus, with an independent hardware watchdog enforcing fail-safe behaviour.
Sub-8 ms inference
The quantised LSTM achieves a mean inference latency of 6.3 ms on Orin, comfortably inside the window demanded by flight-control loops.
Zero in-flight false alerts
Across its first operational year on commercial routes, the system produced zero confirmed false positives, preserving full crew trust.
DO-178C certified
The full C++ and CUDA codebase was developed with complete requirements traceability and achieves DAL-B software integrity level.
Autonomous fail-safe
An independent hardware watchdog monitors the AI module and restores safe state in under 2 ms on any anomaly, with no software intervention required.
Research & modelling
We analysed 1.2 million historical atmospheric events to build the training dataset and define the LSTM's critical features.
Embedded development
We built the CUDA pipeline and applied INT8 quantisation to meet the latency and power constraints of certifiable hardware.
Validation & certification
Hardware-in-the-loop testing across simulated and real turbulence, followed by a DO-178C audit with an accredited body.
Operational deployment
Installation across a 12-aircraft pilot fleet, with continuous remote monitoring feeding model improvement.
We'd evaluated CAT detection solutions for three years without finding anything that held up under both certification and latency constraints at once, and Zeklar solved both in a single pass, with a level of engineering rigour we rarely see even from major system integrators.
AETHER turned Clear Air Turbulence from an unpredictable hazard into a manageable variable, cutting in-flight injuries by 87% across the pilot fleet in its first operational year. The consortium has begun procurement to roll the system out across more than 200 aircraft by 2027.