Causal modeling and risk assessment of runway excursion events via interpretable machine learning

Runway excursions remain a persistent threat to flight safety and operational reliability. The causal chains behind these events are often unclear under complex weather and workload conditions. This study identifies and quantifies the triggers and pathways that lead to runway excursion risk. We analyze 1345 Boeing 737-700 flights using Quick Access Recorder data. An XGBoost model with SHAP values screens nonlinear drivers. We then apply the Independent Component Analysis Linear Non-Gaussian Acyclic model to infer directed and weighted links among sequential exceedances. A Bayesian network estimates trigger probabilities, average treatment effects, and joint risks. Our main contribution is the integration of interpretable learning and causal discovery in a single, queryable graph. Results confirm key triggers: angle of attack asymmetry, elevated radio altitude near touchdown, and delayed thrust reverser deployment. The network shows path dependence, convergence from multiple sources, and escalation along high-impact chains. In a counterfactual policy test, we increase the share of landings with normal bank and a normal 50 ft to touchdown distance by 1.5 percent and reduce the share with small bank and long float by 1 percent. This lowers the estimated probability of severe excessive total landing distance from 11.7 percent to 10.0 percent. A 1 percent increase in the favorable scenario lowers this probability by 0.14 percentage points. Findings remain stable in robust bootstrap tests and noise sensitivity checks. These insights support targeted training and procedures that prioritize touchdown precision and timely deceleration. The framework provides quantitative and reproducible guidance for reliability-centered runway safety management.