scope:

This document reviews current aerospace software, hardware, and system development standards used in the certification/approval process of safety-critical airborne and ground-based systems, and assesses whether these standards are compatible with a typical Artificial Intelligence (AI) and Machine Learning (ML) development approach. The document then outlines what is required to produce a standard that provides the necessary accommodation to support integration of ML-enabled sub-systems into safety-critical airborne and ground-based systems, and details next steps in the production of such a standard.

NOTE: This document and the upcoming standard it is to help inform are concerned only with "offline" learning applications of AI and ML. In offline learning, ML models are trained on historical data within a dedicated learning environment. When the trained models are then implemented into a production system, learning functionality is turned-off. The production system implementing AI may collect data for retraining at a later date, but any retraining or further learning will happen in the separate learning environment, and any resulting changes to the ML models will then need to be reimplemented into production as a new version of the system utilizing AI. This is in contrast to "online" learning, where a system utilizing AI will continue to learn and adapt its operation while in production. Consideration of such systems is not out of scope for SAE-G34, but the committee will not consider online learning until after publication of this document and its related standard.