Deep Learning-Based Virtual Metrology in Multivariate Time Series

In Prognostics and Health Management, virtual metrology is crucial for advanced process control, accounting for the condition of manufacturing machinery. Traditionally, virtual metrology has been tackled using statistical and machine learning approaches, which require extensive domain knowledge and feature engineering. Moreover, the high-dimensional nature of complex industrial systems renders the interpretation of metrology results increasingly difficult. In this work, we introduce PIE-VM, an attention-based multivariate time series regression model incorporating process information for virtual metrology in atomic layer etching. Experimenting on real-world data collected and provided by PSK Inc., a large semiconductor manufacturing equipment company based in South Korea, we empirically demonstrate that our method predicts etch depths more accurately than baseline approaches. Also, we show that our model provides useful information for advanced process control based on its inherent interpretability.