Towards Fair and Explainable AI in Aviation: Case Study on Runway Configuration

Artificial intelligence (AI) is playing an increasingly transformative role in Air Traffic Management (ATM), yet its integration raises critical concerns about fairness, transparency, and robustness. This study addresses these concerns through the lens of Responsible AI (RAI) in aviation by focusing on a decision-support model for runway configuration management. We propose and evaluate a suite of fairness-aware and explainable AI techniques applied to an offline reinforcement learning-based Runway Configuration Assistance (RCA) tool. First, we quantify model bias using a feature-sensitive F1-score permutation framework. Next, we implement two in-processing bias mitigation methods: a modified Meta Fair Classifier (MFC) adapted for multi-class classification, and a feature-wise adversarial debiasing approach that does not require predefined sensitive attributes. We further explore combinations of these with pre-processing strategies such as relabeling. In parallel, we enhance model transparency by employing two interpretability techniques: Layer-wise Relevance Propagation (LRP) and Kernel-SHAP. These methods enable insight into both the global and local behavior of the model. The proposed methods are validated using operational data from three major US airports. The results demonstrate significant improvements in fairness metrics and interpretability, with minimal compromise to predictive performance, supporting the case for responsible AI adoption in safety-critical aviation applications.