ML meets aerospace: challenges of certifying airborne AI.

Artificial Intelligence (AI) technologies can potentially revolutionize the aerospace industry with applications such as remote sensing data refinement, autonomous landing, and drone-based agriculture. However, safety concerns have prevented the widespread adoption of AI in commercial aviation. Currently, commercial aircraft do not incorporate AI components, even in entertainment or ground systems. This paper explores the intersection of AI and aerospace, focusing on the challenges of certifying AI for airborne use, which may require a new certification approach. We conducted a comprehensive literature review to identify common AI-enabled aerospace applications, classifying them by the criticality of the application and the complexity of the AI method. An applicability analysis was conducted to assess how existing aerospace standards - for system safety, software, and hardware - apply to machine learning technologies. In addition, we conducted a gap analysis of machine learning development methodologies to meet the stringent aspects of aviation certification. We evaluate current efforts in AI certification by applying the EASA concept paper and Overarching Properties (OPs) to a case study of an automated peripheral detection system (ADIMA). Aerospace applications are expected to use a range of methods tailored to different levels of criticality. Current aerospace standards are not directly applicable due to the manner in which the behavior is specified by the data, the uncertainty of the models, and the limitations of white box verification. From a machine learning perspective, open research questions were identified that address validation of intent and data-driven requirements, sufficiency of verification, uncertainty quantification, generalization, and mitigation of unintended behavior. For the ADIMA system, we demonstrated compliance with EASA development processes and achieved key certification objectives. However, many of the objectives are not applicable due to the human-centric design. OPs helped us to identify and uncover several defeaters in the applied ML technology. The results highlight the need for updated certification standards that take into account the unique nature of AI and its failure types. Furthermore, certification processes need to support the continuous evolution of AI technologies. Key challenges remain in ensuring the safety and reliability of AI systems, which calls for new methodologies in the machine learning community. [ABSTRACT FROM AUTHOR]