Concerns with using Machine Learning in Airworthiness Applications
H V Carter, Jason Rupert, Alexander Chan, Chris Vinegar · 2023
Machine Learning is seeing accelerating growth in its use across a wide swath of applications ranging from wall mounted thermostats to automobiles. Even the historically conservative aviation industry is beginning to explore the use of machine learning in benign applications, e.g., predictive health monitoring, and beyond, including plans for flight critical applications such as navigation solutions. With the potential push of machine learning into flight critical applications, airworthiness practitioners concerned with software should ask: are there concerns with using machine learning in flight critical airworthiness applications? The goal of this paper is to answer this question with a resounding 'yes', by identifying some specific concerns. These concerns include data-driven development, uncertain (statistical/probabilistic) output, and extending to functional behavior, repeatability, non-rigorous development, and specific deployment environments. Beyond just identifying the concerns, this paper proposes approaches to build justified confidence in the use of machine learning in flight critical applications through combined evaluation of performance assurance, development assurance, and mitigations.