Plasma etching experiments are costly, making the rapid development of reliable process surrogate models with limited samples essential for improving process development efficiency. Conventional uncertainty sampling primarily focuses on regions where the model has limited knowledge, but it may repeatedly select nearby candidate points in small-sample settings, resulting in redundant information. Conversely, sampling based solely on spatial distribution cannot effectively refine regions with complex response behavior. To address these limitations, this paper proposes a Distance-Uncertainty Weighted Active Learning Kriging (DUW-ALK) method. The proposed method employs an Automatic Relevance Determination (ARD) Matérn-5/2 Kriging model as the probabilistic surrogate. It integrates epistemic uncertainty, characterized by the predictive standard deviation, with the spatial novelty of candidate points relative to the training set through normalization and dynamically adjusts their relative contributions to balance design space coverage and local refinement. Comparative results on two-dimensional computational fluid dynamics (CFD) simulation data and four-dimensional experimental deep reactive ion etching (DRIE) data demonstrate that spatial distance information reduces redundant local queries, while posterior uncertainty helps identify response regions requiring additional samples. The synergy between these two signals enables DUW-ALK to achieve improved sample efficiency throughout the learning process and higher predictive accuracy across different data structures. The proposed method provides a simple and readily implementable approach to sequential experimental design for high-cost processes with limited sample availability.

active learning,kriging model,plasma etching,small-sample modeling