AVS 65th International Symposium & Exhibition
    Plasma Science and Technology Division Thursday Sessions
       Session PS-ThA

Paper PS-ThA10
Development of the Virtual Metrology Using a Plasma Information Variable (PI-VM) for Monitoring SiO2 Etch Depth

Thursday, October 25, 2018, 5:20 pm, Room 104A

Session: Plasma Diagnostics, Sensors and Controls
Presenter: Yunchang Jang, Seoul National University, Republic of Korea
Authors: Y. Jang, Seoul National University, Republic of Korea
H.-J. Roh, Seoul National University, Republic of Korea
S. Ryu, Seoul National University, Republic of Korea
J.-W. Kwon, Seoul National University, Republic of Korea
G.-H. Kim, Seoul National University, Republic of Korea
Correspondent: Click to Email
Advanced process control (APC) has been attracting attention as a technology to enhance process yield and it requires accurate and reliable virtual metrology (VM). Accuracy of VM is determined by how sensitively the input variables reflect the drift and changes of the process environment. Many previous approaches to improve the performance of VM have been focused on development of the statistical methods to select the valuable input variables from the equipment data and additional sensor data such as optical emission spectroscopy (OES) and plasma impedance monitors (PIM). In this study, the noble variables, named plasma information (PI) variables are introduced, which are obtained by phenomenological analysis and they are added into the VM development. Then we evaluated its contribution to improve the accuracy of VM. It notes that PI variables represents the state of etch plasma so it can be used to monitor the variation of process results in plasma-assisted semiconductor fabrication process. Effect of PI variables on improving VM accuracy has been investigated through following conventional (or standard) VM development procedures as follows; 1. preprocess of input dataset, 2. data exploration, 3. variable selection, 4. training of a model, and 5. Validation of the model. We added PI variables in the steps (i) in-between 2 and 3 steps (called PI-VMSTA) and (ii) in-between 3 and 4 steps (called VMSTA+PI). Each VM model are developed and evaluated by using 50 sets of SiO2 etching depth data, having 20:1 aspect ratio and less than 5 % of variation. PIEEDF, representing variation of electron energy distribution function (EEDF) is obtained from analysis of OES, which is based on the argon excitation kinetics. Pearson’s correlation filter, principal component analysis (PCA), and stepwise variable selection are used for the variable selection methods. Results show that VM models using PIEEDF have better performance than any other conventional VM models because PIEEDF has much higher correlation with output variable than the other equipment and sensor variables. Especially, PI-VMSTA using stepwise variable selection method shows the highest accuracy where PIEEDF provides a basis to select other OES variables. This study shows that a phenomenological-based, statistically tuned VM can be developed by using PI variables as input. It has advantages for management of dataset and selection of control variables in APC application.