Two-Stage Case-Based Reasoning Framework With Tree-Based Ensemble Methods for Printed Circuit Board Yield Prediction

Abstract

In printed circuit board (PCB) manufacturing, yield is a critical indicator that directly impacts production costs and delivery schedules. For new products, the industry currently relies on empirical predictions by engineers based on historical yields of similar products. However, this approach suffers from limitations in accuracy and consistency due to subjective and non-systematic assessment criteria, issues that are compounded by the complex product structures and multi-stage processes used in the PCB industry. To address these challenges, this study proposes a yield prediction methodology that leverages Case-Based Reasoning (CBR) and tree-based prediction models. This methodology consists of two stages. The first stage is to predict the yield of model units before the new product is put into the first process. In this stage, we utilize SHapley Additive exPlanations (SHAP)–based variable importance to reflect the weight of each variable and determine the optimal number of clusters based on silhouette analysis to search for similar cases. Then, data from retrieved cases is analyzed with a tree-based prediction model to predict yield, which evaluates the relationship between dates and yield to detect whether a statistically significant change has occurred at a particular point in time. The second stage utilizes additional defect information from Automated Optical Inspection (AOI) reports to make more precise yield forecasts on a lot-by-lot basis. This further increases the reliability of the forecast by reflecting the quality variability that occurs during production. Based on these predictions, more accurate material input calculations can be performed, thereby improving delivery compliance rates and minimizing excess inventory.