Ph.D. Thesis Proposal 

Alberto Cardenas Melgar

(Faculty Advisor: Professor Dimitri Mavris)

"An Uncertainty-Aware and Explainable Assurance Framework for Safety-Critical Aviation Deep Learning Models Using Conformal Prediction"

Friday, June 5

12:30 p.m.

Weber, CoVE

Abstract: 

Machine Learning (ML) models are increasingly being implemented into aviation systems to support tasks such as trajectory prediction, anomaly detection, collision avoidance, aircraft inspection, and operational decision-making. However, the adoption of AI/ML in operational aviation remains constrained by challenges related to safety assurance, uncertainty quantification, explainability, robustness, and certification. Traditional ML evaluation methods primarily focus on predictive performance metrics and often fail to provide the valid safety guarantees, interpretability, and operational trustworthiness required for uncertain safety-critical aviation environments. Recent efforts by the European Union Aviation Safety Agency (EASA), the Federal Aviation Administration (FAA), and related research initiatives have highlighted the need for learning assurance frameworks capable of addressing uncertainty, robustness, and continuous safety monitoring in AI-enabled aviation systems.

The research focuses on aviation safety-event prediction and operational monitoring problems using sequential flight data and deep learning architectures, including recurrent and convolutional neural networks. This dissertation proposes a comprehensive framework for trustworthy aviation AI that integrates uncertainty-aware machine learning, conformal prediction, and explainable artificial intelligence (XAI) oriented for time-series aviation applications.

First, the dissertation investigates conformal prediction methods for uncertainty quantification under distribution shift and non-stationary operational conditions. Online, semi-online, and traditional conformal prediction approaches are evaluated against Bayesian and ensemble-based uncertainty methods using metrics related to calibration, conditional coverage, interval efficiency, and robustness. Special attention is given to the relationship between conformal prediction set size, model generalization, and operational reliability.

Second, the dissertation develops uncertainty-aware multi-objective optimization strategies for training aviation ML models. In addition to predictive performance objectives, the proposed optimization frameworks incorporate uncertainty-related objectives such as coverage gap and interval width metrics. The study evaluates whether uncertainty-aware optimization produces models with improved reliability, robustness, and operational trustworthiness compared to conventional performance-only optimization approaches.

Third, this work explores explainability methods for predictive uncertainty in aviation AI systems. Existing XAI approaches, including KernalSHAP, TimeSHAP, and TSHAP are adapted and analyzed for sequential aviation data and conformal prediction outputs. The dissertation investigates whether local explanation techniques can effectively characterize uncertainty sources, identify failure modes, and support operational safety assessment for aviation decision-support systems.

Finally, the dissertation proposes an integrated evaluation framework aligned with emerging aviation AI assurance concepts, including EASA’s trustworthiness principles, learning assurance processes, and continuous safety paradigms. The framework combines statistical evaluation, uncertainty quantification, explainability analysis, robustness assessment, and operational performance metrics to support the development of certifiable and trustworthy AI systems for aviation.

The contributions of this dissertation include:
(1) a comparative evaluation of conformal prediction approaches for aviation time-series applications,
(2) uncertainty-aware multi-objective optimization methodologies,
(3) explainability techniques for predictive uncertainty in sequential models, and
(4) a unified assurance-oriented framework for trustworthy aviation AI. Collectively, these contributions aim to bridge the gap between machine learning performance evaluation and the safety, reliability, and certification requirements.

Committee:

Dr. Dimitri Mavris (advisor), School of Aerospace Engineering
Dr. Keegan Moore, School of Aerospace Engineering
Dr. Karen Feigh, School of Aerospace Engineering
Dr. Alexia Payan, School of Aerospace Engineering
Dr. Yao Xie, Industrial and Systems Engineering