Finding Common Ground For Airborne Machine Learning Certification

Abstract

Machine-learning (ML) and, more broadly, artificial intelligence (AI) have become ubiquitous in the consumer market by enabling the development of complex functions inferred from data rather than hard-coded from requirements. Aviation is now adopting ML to address challenges such as perception in degraded-visual environments and optimal flight-path prediction, that conventional engineering cannot readily solve, or that ML can implement more efficiently. Before ML can be fielded in safety-critical aeronautical systems, however, both technical and procedural barriers must be overcome. Regulators in the United States and the European Union are developing complementary, yet presently different, policies, guidance and verification methods to support this transition. This paper compares the emerging certification frameworks of the FAA and EASA, highlights areas of convergence and divergence, and proposes recommendations that may facilitate either formal harmonization or, where necessary, pragmatic coexistence of the two approaches.