Chandrasekar S supervised by Dr. Karthik Vaidhyanathan received his Master of Science by Research in Computer Science Engineering (CSE). Here’s a summary of his research work on A Semi-Automated Framework for the Certification of Machine Learning based Airborne Applications
Machine learning components are increasingly being integrated into airborne systems, yet existing aviation certification standards such as DO-178C were not designed with data-driven components in mind. This is a challenge for low-criticality systems at Design Assurance Level (DAL) C and DAL D, where no specific certification guidance for ML-enabled systems (MLS) currently exists. We aim to address this by proposing a structured, evidence-based approach to ML certification for DAL C and DAL D airborne applications.
We review research papers and grey literature alongside existing standards and guidelines to derive an ML certification standard organised around five process areas: planning, development, verification and validation, quality assurance, and configuration management. The standard covers both DAL C and DAL D, with requirements adjusted to reflect the difference in assurance levels between the two. From this standard, we design a semi-automated certification framework that combines automated ML testing with human review for tasks that require contextual judgment. The automated testing covers data-specific checks such as size, distribution, drift, consistency, and integrity, as well as model-specific checks covering performance, robustness, stability, and explainability. We derive a score-based certification checklist from this framework and define a Certification Assurance Profile that aggregates the checklist scores into a structured evidence record for each ML component. We evaluate the checklist through two phases: an expert review with certification authorities and software vendors, followed by semi-structured interviews with eleven industrial experts from non-airborne domains. We also implement a web-based semi-automated checklist tool that allows vendors to fill in the checklist, view automated test results, and track the Certification Assurance Profile as items are completed. We also propose an extension to LoCoML, an existing low-code ML deployment framework, where the certification process is modelled as a pipeline with each node corresponding to one of the five process areas.
The expert review and interviews produced two revised versions of the checklist, with findings showing broad agreement on the overall structure while also identifying areas that need further refinement. A planned third phase will apply the validated checklist to three ML-enabled systems covering tabular, computer vision, and time series data modalities. The LoCoML pipeline extension has also been left as a direction for future work.
Together, these contributions provide a practical pathway for certifying MLS in low-criticality airborne systems and lay the groundwork for further automation of the certification process.
July 2026