Certifying AI for Safety Critical Aircraft Systems

On October 27, Project Wayfinder’s Chief Engineer Cedric Cocaud joined colleagues from the aviation industry on a webinar organized by Avionics to address the question of how to achieve type certification for safety critical eVTOL systems. The answer, as you might expect, is a deeply nuanced one.

Regulators are supportive as the industry innovates

Across the board, the speakers agreed that while there is no clear, one-size-fits-all path to certification today, regulatory bodies are working swiftly and creatively to support ongoing progress and to create a more consistent path for the future. Industry and certification authorities are working hand-in-hand to ensure that technological progress and evolution isn’t slowed as the industry moves toward ensuring even greater efficiency and safety for all.

Safety critical systems versus propulsion systems and aircraft design

Martin Peryea, CTO of Jaunt Air Mobility, pointed out that certification of safety critical systems are independent from aircraft and propulsion systems, be those battery, hydrogen, hybrid or otherwise. This also goes for aircraft design: eVTOL, fixed wing, etc. As new propulsion solutions lead to an opening of the design space for aircraft altogether, certification requirements will be considered independently.

AI is only as good as the data that feeds it

AI is the cornerstone of autonomy in aviation, and autonomy is key to future industry growth and increased safety. Cedric pointed to the importance of data, in terms of quality, volume and accurate labeling as crucial to improving AI and its reliability. This is the central focus of Acubed’s Project Wayfinder team as it builds scalable, certifiable autonomy systems to power the next generation of aircraft.

Amanda Melles, Chief Airworthiness Computer Scientist for the MANNARINO Design Assurance Organization, highlighted that safety critical system certification is a particularly nuanced one as software has unique failure conditions from hardware, costs remain high even as the market and demand grows rapidly and complexity in airborne software increases across the board.

The challenge of certifying AI

All the webinar participants agreed that certifying AI for safety critical systems is currently a work in progress for the industry. In the case of autonomous landing, the concept for using AI is specific: work is first done on the ground until the AI reaches target performance levels, at which point it is frozen and ported to the aircraft for in-flight testing. This is what Project Wayfinder is currently doing through its data collection flight campaign in the Bay Area.

Fig. 1: Project Wayfinder has been collecting data in the Bay Area through its flight test lab (pictured here) since July 2020

To achieve certification, companies will need to prove that the AI-based system is predictable, well behaved, and can fully perform the operational envelope for which it was developed. Enough data is required not only to demonstrate that all scenarios are covered to support the highest levels of performance and safety, but also to demonstrate exactly when and how the system performance degrades beyond the target operational scope. “These are the fundamentals for certification,” concluded Cedric, “which makes for some exciting times ahead.”

Watch the full webinar HERE.

- Paige Wilson