Deep Reinforcement Learning-Based Parameters Optimize Prediction Model for Smooth-Vertical Sidewall Profile in Deep Reactive Ion Etching Process
Fang Wang, Hao Yu, Miao Yu, Yue He, Ke Sun, Yi Sun, Heng Yang, Xinxin Li
Abstract
This paper reports a deep reinforcement learning-based, multi-parameters optimize prediction model for achieving smooth-vertical sidewall profile in deep reactive ion etching (DRIE) process for the first time. Leveraging the adaptive learning capabilities of Deep Q-network (DQN), the proposed model dynamically learns from DRIE process feedback, predicting and refining the optimal key parameter recipes, such as etching cycle time, passivation cycle time. A reward-penalty strategy, based on critical sidewall profile indicators like angle and uniformity, steers the learning process to output optimal recipes once indicators ideal states are achieved. Extensive orthogonal experimental validation demonstrates that an etching/passivation ratio within the range of 1.09 to 1.4 can achieve nearly 90-degree sidewalls, while ratios between 1.09 and 1.5 ensure scallop depth below 80 nm. Applying the model-generated recipes in actual DRIE, the model yields 0.95 accuracy in sidewall angles, 0.94 in scallop size, and 0.92 in critical dimension, significantly enhancing sidewall verticality and smoothness. The model provides a robust solution for addressing the complex parameter space of DRIE, paving the way for more consistent and scalable microfabrication processes.