A Reproducible Approach to Virtual Metrology

Jimmy Pöhlmann, Anna Lopatkina, Paul Jungmann, Claudio Hartmann, Wolfgang Lehner

Abstract: Semiconductor fabrication relies strongly on metrology to monitor process and chip quality. However, physical measurements are expensive and time-consuming. Hence, Machine Learning (ML) models are often used to predict many of those measurements more efficiently, a concept known as Virtual Metrology (ViMet). However, legal constraints limit what can be published in this field, especially regarding the underlying data, without compromising intellectual property. This makes it difficult to reproduce results and to assess an approachs effectiveness based on the current literature alone. To address this, we propose four simple adjustments to the data mining process, which are well-established in other ML fields, to enable reproducibility and improve transparency even when the underlying data cannot be published: automating data preprocessing, automating hyperparameter optimization, using scaled error metrics, and conducting model comparisons. To demonstrate their usage, we conduct a standard data mining task in ViMet, identifying the best modeling approach to predict three metrology measurements.

External IDs:dblp:conf/icdm/PohlmannLJHL25