Feature Selection for Virtual Metrology Modeling: An application to Chemical Mechanical Polishing

This paper focuses on the feature selection problem in a virtual metrology task applied to a chemical mechanical polishing process. One of the main challenges specific to virtual metrology modeling is the relatively wide availability of measurements and traces (features) versus the scarcity of samples (entries), as they are usually costly to obtain. To overcome these challenges, we propose a hybrid feature selection algorithm, called Enhanced Hybrid Feature Selection (EHFS), that combines a filter approach and a genetic algorithm embedding a machine learning model. The filter starts by eliminating noisy and uninformative features. Then, in the wrapper stage, the genetic algorithm is augmented by a solution archive to favor exploration. This added feature avoids the reevaluation of duplicate candidate solutions and consequently decreases the computational time of EHFS.Numerical experiments, conducted on industrial and benchmark datasets, show that the proposed solution approach performs competitively in terms of both solution quality and computational time compared with two existing approaches: the general-purpose Forward Feature Selection (FFS) and virtual metrology-specific Evolutionary Repetitive Backward Elimination (ERBE).