Application of Explainable Artificial Intelligence (XAI) in Combination with Bootstrap to Improve Processes in Model-Based Aero-Engine Development

In aerospace engineering, data-driven surrogate models are increasingly employed to mitigate the computational and temporal costs of simulations, numerical analyses, and experiments. Two major challenges accompany this trend. First, the training of surrogate models often requires a sufficient amount of data, the determination of which is inherently difficult. Second, these models often exhibit high complexity, limiting both the traceability of their outputs and the extraction of useful insights. Explainable Artificial Intelligence (XAI) methods have therefore emerged as promising tools to enhance the interpretability, explainability, and transparency of such models. In this work, a combination of the established Shapley Additive Explanations (SHAP) approach with a bootstrap-based method is investigated. The proposed framework provides insights into the contribution of individual features and enables an assessment of data sufficiency with respect to surrogate model performance. Building upon these findings, the Bootstrap-Informed Feature Importance (BIFI) method is proposed. BIFI offers a model-agnostic, robust identification of relevant features.The method is analyzed in the context of Design of Experiments (DOE) processes used for surrogate model construction. Evaluation on four synthetic datasets of increasing complexity, as well as a dataset from aero-engine development, demonstrates that BIFI-based DOEs can improve surrogate model quality measured in terms of R\((^2)\) and MSE by up to 90%. Consequently, the proposed method enables more efficient utilization of simulations, computations, and experiments while reducing the required number of samples.