Dmitriev, KonstantinKaakai, FatehIbrahim, MohamadDurak, UmutPotter, BillHolzapfel, FlorianGroher, IrisVogel, Thomas2023-02-132023-02-132023https://dl.gi.de/handle/20.500.12116/40204Machine Learning (ML) technology can provide the best results in many highly complex tasks such as computer vision and natural language processing and quickly evolving further. These unique ML capabilities and apparent potential can enable the next epoch of automation in airborne systems including single pilot or even autonomous operation of large commercial aircraft. The main problems to be solved towards ML deployment in commercial aviation are safety and certification, because there are several major incompatibilities between ML development aspects and traditional design assurance practices, in particular traceability and coverage verification issues. In this paper, we study the qualification aspects of tools used for development and verification of ML-based systems (ML tools) and propose mitigation measures for some known ML verification gaps through ML tools qualification. In particular, we review the DO-330 and DO-200B tool classification approach with respect to ML-specific workflows and propose to extend the tool qualification criteria for ML data management and ML model training tools.enMachine LearningCertificationDesign AssuranceTool QualificationText/Conference Paper10.18420/se2023-ws-19