Scatterometry-informed machine learning study to determine bi-directional intercorrelation of adjacent patterning steps

Abstract

The use of machine learning has been well documented in recent years in a wide variety of optical scatterometry applications. Machine learning can either be used in a ‘modeless’ manner to directly correlate measured spectra to reference metrology without an optical model or serve as a complementary technique together with conventional scatterometry modeling to improve the sensitivity of specific parameters. This work presents both modeless and machine learning assisted modeling applications in the gate-all-around nanosheet process flow. Modeless machine learning is utilized to demonstrate reliable prediction of downstream inline measurements where certain geometric attributes may not yet be defined at the step where spectra are collected but may depend on dimensions of the incoming geometric features. A concept is then introduced to compare forward and backward prediction of selected measurement steps and examine possible inconsistencies between the outputs of the associated scatterometry models. This bidirectional machine learning assessment proves to be a simple methodology to reveal the intercorrelation between past and future process steps and helps to identify both model and process stabilities. Finally, an example of machine learning assisted scatterometry modeling is demonstrated and validated by conventional TEM correlation and in the same forward backward intercorrelation matrix.