Our team shares lessons on certifying AI-based aviation systems. We are pursuing an FAA STC for FLEX AI™.
Bringing an AI-based system to market in commercial aviation is one of the most technically and regulatorily demanding challenges in any industry. The Federal Aviation Administration's certification framework was developed over decades to address deterministic, rule-based systems — systems whose behavior under any given set of inputs can be precisely predicted and verified. AI systems, by their nature, do not fit neatly into this framework. Their behavior emerges from training data and statistical inference rather than explicit programming, which creates genuine challenges for traditional certification approaches.
This article shares what we have learned — both about the regulatory landscape as it currently stands and about the direction in which it is evolving.
The FAA's primary framework for certifying avionics and aircraft systems is DO-178C (Software Considerations in Airborne Systems and Equipment Certification) for software and DO-254 for hardware. These standards were designed for deterministic software and do not directly address the characteristics of machine learning systems — specifically, the fact that an ML model's behavior is determined by its training data rather than by explicit code.
In 2020, RTCA (the standards body that develops DO-178C and related standards) published DO-178C Supplement: Machine Learning in Airborne Systems (informally known as the ML Supplement), which provides initial guidance on how ML systems should be developed, validated, and documented for airborne applications. This supplement represents the FAA's first formal acknowledgment of ML as a distinct category of airborne software, and it has become the primary reference document for AI certification projects.
The ML Supplement identifies several characteristics of machine learning systems that require special attention in the certification process. Data management is perhaps the most significant: the quality, representativeness, and traceability of training data must be rigorously documented and justified. For a structural health monitoring system, this means demonstrating that the training dataset includes sufficient examples of the failure modes the system is designed to detect, drawn from a representative range of aircraft types, operational conditions, and structural configurations.
Model explainability is another area of focus. Certification authorities generally require that system behavior be explainable — that it be possible to understand why the system produced a particular output. This is straightforward for rule-based systems but challenging for complex neural networks. The FLEX AI Platform addresses this through a combination of attention visualization techniques and a parallel rule-based verification layer that provides an independent check on AI-generated alerts.
Operational domain definition — clearly specifying the conditions under which the system is designed to operate and perform inside its defined operating conditions — is also critical. An AI system that performs well within its operational domain may behave unpredictably outside it. Certification requires that the operational domain be precisely defined and that the system include mechanisms to detect when it is operating outside that domain.
We are pursuing an FAA STC for FLEX AI™. The process involves extensive documentation of our training data pipeline, model architecture, validation methodology, and operational domain definition, as well as a comprehensive flight test program.
One of the most valuable aspects of the process was the collaborative relationship we developed with our FAA certification team. The agency's engineers are genuinely engaged with the challenge of certifying AI systems and are actively working to develop more efficient certification pathways. We found them to be thoughtful partners in problem-solving rather than obstacles to be overcome.
The FAA has signaled clearly that AI and machine learning will play an increasingly central role in aviation safety systems, and the agency is actively investing in developing the regulatory infrastructure to support this. The FAA's Artificial Intelligence and Machine Learning (AI/ML) Roadmap, published in 2023, outlines a multi-year program to develop updated certification standards, guidance material, and internal expertise for AI systems across all aviation domains.
For companies developing AI-based aviation systems, the key takeaway is that early and frequent engagement with certification authorities is essential. The regulatory framework is still evolving, which means there is genuine opportunity to help shape it — but it also means that certification programs require more flexibility and collaboration than they would for a mature, well-defined technology category.
Platinum Eagle Aerospace is committed to being an active participant in this process. We believe that the development of clear, efficient certification pathways for AI-based aviation systems is one of the most important contributions the industry can make to the long-term safety and efficiency of commercial aviation.
Learn more about FLEX AI™ structural health monitoring.