Machine learning surrogate modeling with quantile adjustment for landing gear surface pressure fluctuation prediction

Many engineering datasets are difficult for machine-learning surrogate modeling because the selected features are informative but insufficient, and the response distribution is often biased toward noncritical regions. This paper investigates that issue through an aeroacoustic case study in which limited steady-state flow variables are used to predict a transient hotspot-dominated landing-gear noise field. Multiple regression-based surrogate models are benchmarked, including functional, neural, boosting-based, and decision-tree-based methods. The results show that conventional Random Forest outperforms more advanced boosting and deep tabular models, indicating that model suitability depends strongly on how the learning characteristics match the available feature information and response structure. The study also shows that when a baseline predictor captures the overall response trend but systematically underestimates peaks, an upper quantile can serve as an effective bias-adjustment strategy to improve hotspot prediction.