Data Drift Quantification Metrics for Month-on-Month Semiconductor Data

Semiconductor manufacturing process goes through multiple steps that produce a huge volume of wafers and chips. In order to maximize yield, a production line deploys a Quality Prediction model that can detect Critical Dimensions (CD) ahead of time. Often this is achieved using statistical and machine learning approaches. However, the probability distribution of the input training data may drift over time due to various environmental factors in the semiconductor FAB process, causing the machine learning models to fail. This necessitates a drift detection, quantification and mitigation strategy. Although data drift detection study is abundant in literature, there are relatively few established metrics for quantifying data drift. In this paper, we present a novel approach to quantify drift in the input data by adapting the DARE-GRAM method and using the latent vectors from its model training. On a semiconductor dataset with month-on-month drift, we observe a correlation of -0.39654 with RSQ and 0.56735 with RMSE, which are the metrics being used for evaluating our task of Critical Dimension prediction. The proposed solution is potentially impactful for model explainability, domain adaptation and model monitoring and root cause analysis of model performance degradation.