In semiconductor manufacturing, the process parameters for deep reactive ion etching (DRIE) are typically coupled. The trade-off between lateral dimension control and a high aspect ratio (AR) makes process parameter determination heavily reliant on extensive experiments, thereby increasing process development costs. For process optimization under limited experimental data, this paper proposes a bi-objective optimization method that combines the ordinary Kriging (OK) surrogate model with the non-dominant sorting genetic algorithm II (NSGA-II). First, experimental data are obtained through experimental design. OK models are established for the one-sided critical dimension (CD) loss and AR, respectively, and their predictive performance is evaluated using leave-one-out cross-validation (LOOCV) metrics. Then, with the objectives of reducing CD loss and increasing AR, NSGA-II is used to search for candidate non-dominated solutions. The optimal process parameter combination is determined by an equally weighted normalized score, and its trade-off performance is evaluated based on the measured results. Experimental results on different wafer sizes and multiple feature dimensions show that the determined process parameter combination can meet both the dimensional control and aspect ratio requirements, and exhibits robustness and generality.

deep reactive ion etching,process parameter,Kriging,bi-objective optimization