Invited Paper PS+EM-WeM12
Etching Recipe Optimization Using Machine Learning

Wednesday, October 24, 2018, 11:40 am, Room 104A

The development of semiconductor fabrication processes has been prolonged due to constantly evolving nano-scale 3D devices. This lengthy development period has driven up the cost of semiconductor devices, and the process development needs to be sped up in order to reduce the cost.

Along with time modulation of plasma generation and bias power and an increase in the number of gas species, continuous improvement of the control functions of a plasma etcher has been made to provide accurate nano-scale etching. A set of parameters for the control functions is called a recipe, which is used as input data for the etcher. Etching accuracy can be improved by increasing the number of parameters in the recipe. However, it is difficult to optimize the recipe for obtaining a target etch profile when there are many parameters.

In this work, we present two types of exploration method for recipe optimization using machine learning: one using etching profile data [1] and the other using feature data related to the etching profile [2]. These were developed to assist the development of the etching process and to reduce the time period of the development, respectively.

In the method using the profile data, a recipe is optimized through the repetition of an optimization cycle that consists of learning the relationship between the profiles and the recipes, predicting the recipes to obtain a target profile, performing etching experiments with the predicted recipes, and adding the experimental results to the learning data. In this cycle, kernel ridge regression is used as the learning engine and a Si trench pattern is used to examine the exploration method. By using the predicted recipe, a vertical trench profile was successfully etched, and the profile was further improved by increasing the number of cycles from one to seven.

Next, we developed the recipe exploration method based on machine learning of feature data related to the etching profile in order to accelerate the optimization. A micro/macro cavity method is used to extract the feature data. An approximate region to obtain the vertical profile can be determined in the feature data space because the feature data show the characteristics of ion assist etching and radical etching. The relationship between the feature data and the recipes was learned, and feature data were then explored to obtain the vertical profiles. After the iteration of the exploration, it enabled us to determine the optimum recipes for the vertical profile in just seven times of Si trench etching.

[1] T. Ohmori et al., Proc. of Int. Symp. Dry Process, Tokyo, pp. 9–10 (2017).

[2] H. Nakada et al., Proc. Gaseous Electronics Conf., Pittsburgh (2017).