Drilling parameters are intricately linked to the mechanical interactions between the drilling device and lunar regolith, significantly affecting sampling characteristics. Achieving high coring efficiency requires a deep understanding of how these parameters influence the coring rate. While experiments can characterize coring performance, they are time-consuming and costly due to the need for numerous repetitive tests. To address this, this paper integrates experiments with machine learning to develop a surrogate model that replaces experimental results, allowing for extrapolation and interpolation to analyze the sensitivities of drilling parameters. Specifically, Taguchi orthogonal experiments are designed to gather a representative sample set, and a Bayesian neural network (BNN) is employed to create a predictive model which can quantify uncertainty. Active learning is then applied to focus additional sampling on uncertain regions, refining the model iteratively to achieve an accurate approximation of experimental results. This approach reduces the number of experiments needed to develop a high-fidelity surrogate model. Validation results demonstrate strong extrapolation accuracy and generalization, with an R² of 0.89 and a MAPE of 5.74% on the independent test set. Based on this model, this paper analyzes the influence of drilling parameters on sampling performance, providing insights for optimizing drilling parameters to maximize coring efficiency.