Machine Learning Models for Edge Placement Error Based Etch Bias

As the technology node in semiconductor manufacturing continuously shrinks its feature size and boosts the transistor density, etch bias is facing great challenges that require much better control of the edge placement error (EPE). The traditional applications of etch bias either by rule or by model are sometimes of lower precision or too much time consuming. We propose and demonstrate several innovative EPE based machine learning models for etch bias that have successfully achieved satisfying accuracy and time cost for one of the latest advanced tech nodes in industry. In addition, we propose a novel methodology for massive EPE measurement on wafer that is based on automatic image processing. Three types of machine learning models (single neural network, ensemble neural networks, and random forest) and a novel feature vector used for the machine learning have been studied here. A comparison with the commercial etch-model software, Variable Etch Bias (VEB) from Mentor Graphics, has also been taken. As a result, our proposed machine learning models achieved better accuracy within greatly shortened time compared to the VEB model in our test case.