The most dangerous risk is the one radar cannot see.
Clear-air turbulence is invisible to conventional radar and causes the majority of non-crash in-flight injuries, with events rising as extreme atmospheric phenomena intensify. Existing solutions operate at network level, with latencies incompatible with flight-control loops, and accurate predictive models rarely survive the certification and power constraints of avionics hardware. The result is that the risk arrives with no usable warning.
On-board early forecasting
An LSTM network separates meaningful micro-pressure fluctuations from sensor noise across a few-second window, forecasting the atmospheric event roughly 30 seconds ahead, the runway flight control needs to react.
Certifiable edge inference
The INT8-quantised model runs in a handful of milliseconds on avionics-grade edge hardware, with CUDA-accelerated on-device preprocessing to eliminate transfer latency and stay inside control-loop constraints.
DO-178C / DO-254 compliance
The full codebase is developed with complete requirements traceability up to DAL-B integrity level, so the system clears certification audits with no compromise on latency.
Autonomous hardware fail-safe
An independent hardware watchdog monitors the AI module and restores safe state within milliseconds on any anomaly, with no software intervention, preserving crew trust through a near-zero false-positive rate.
Sample
High-frequency acquisition from redundant IMU and barometric sensors, fused with LiDAR, builds a continuous atmospheric picture around the aircraft.
Predict
The quantised LSTM analyses the time window on edge hardware and estimates the probability and intensity of the incoming atmospheric event.
Signal
The forecast travels over the dedicated avionics bus to the flight computer with deterministic latency and calibrated confidence thresholds.
Compensate
The flight-control system adjusts attitude and notifies the crew before passengers feel any impact.
AETHER is an embedded avionics system in which inference lives on board, between a redundant sensor front-end and the flight computer. The architecture is built around determinism and fail-safe: every stage has bounded latency and an independent hardware watchdog oversees the AI module.
Cloud (training and fleet)
Cloud is used only for offline model training and fleet-level aggregate analysis; the production model is quantised and distributed to aircraft as a signed artifact, never executed in cloud on board.
Embedded on board
Standard configuration: the whole pipeline runs on the avionics-grade edge hardware integrated in the aircraft, with local inference and output on the ARINC 429 bus. No flight data leaves the aircraft in real time.
Certified air-gapped
For military or government operators: the module operates fully isolated, with weight updates only via a ground-maintenance procedure on signed media, consistent with DO-178C traceability.
Certification standards
The full codebase is developed with complete requirements traceability up to DAL-B under DO-178C (software) and DO-254 (programmable hardware), with evidence ready for certification audit.
Data sovereignty
Flight data stays on board: inference is local and no atmospheric or attitude payload is transmitted in real time. Fleet analysis uses only aggregated data exported on the ground under operator control.
Access and audit
Weight updates and configurations are signed and traced through the maintenance procedure, with an immutable log of model version, confidence thresholds and watchdog activations for safety review.
Encryption and integrity
On-board storage is encrypted and every model artifact is signed and verified before execution; verified boot prevents any unauthenticated code from running on the module.
Does AETHER control the aircraft?
No. AETHER produces a calibrated-confidence forecast that it hands to the flight computer and crew; attitude compensation remains the responsibility of the certified flight-control system, not the AI module.
How do you handle false positives?
Confidence thresholds are calibrated for a near-zero false-positive rate: frequent, unjustified alarms would erode crew trust, so the system is tuned to signal only high-probability events.
Is it compatible with existing avionics?
Yes. The output uses the standard ARINC 429 bus and the module is designed as a subsystem integrable with no architectural changes to the flight computer.
What happens if the AI module fails?
A hardware watchdog independent of the software detects the anomaly and restores safe state within milliseconds: on failure the aircraft reverts to nominal behaviour without the predictive contribution, never degrading safety.
How heavy is the certification path?
AETHER is built with DAL-B requirements traceability and DO-178C/DO-254 artifacts from the start: the evidence package ships with the product, reducing the integrator's certification effort versus a model developed outside the standard.
The mark encodes the mechanism, not the sector: l'ala a delta con la scia di vortice.