Smart Feature Selection to enable Advanced Virtual Metrology

Abstract

The present dissertation enhances the research in computer science, especially state of the art Machine Learning (ML), in the field of process development in Semiconductor Manufacturing (SM) by the invention of a new Feature Selection (FS) algorithm to discover the most important equipment and context parameters for highest performance of predicting process results in a newly developed advanced Virtual Metrology (VM) system. In complex high-mixture-low-volume SM, chips or rather silicon wafers for numerous products and technologies are manufactured on the same equipment. Process stability and control are key factors for the production of highest quality semiconductors. Advanced Process Control (APC) monitors manufacturing equipment and intervenes in the equipment control if critical states occur. Besides Run-To-Run (R2R) control and Fault Detection and Classification (FDC) new process control development activities focus on VM which predicts metrology results based on productive equipment and context data. More precisely, physical equipment parameters combined with logistical information about the manufactured product are used to predict the process result. The compulsory need for a reliable and most accurate VM system arises to imperatively reduce time and cost expensive physical metrology as well as to increase yield and stability of the manufacturing processes while concurrently minimizing economic expenditures and associated data flow. The four challenges of (1) efficiency of development and deployment of a corporate-wide VM system, (2) scalability of enterprise data storage, data traffic and computational effort, (3) knowledge discovery out of available data for future enhancements and process developments as well as (4) highest accuracy including reliability and reproducibility of the prediction results are so far not successfully mastered at the same time by any other approach. Many ML techniques have already been investigated to build prediction models based on historical data. The outcomes are only partially satisfying in order to achieve the ambitious objectives in terms of highest accuracy resulting in tight control limits which tolerate almost no deviation from the intended process result. For optimization of prediction performance state of the art process engineering requirements lead to three criteria for assessment of the ML algorithm for the VM: outlier detection, model robustness with respect to equipment degradation over time and ever-changing manufacturing processes adapted for further development of products and technologies and finally highest prediction accuracy. It has been shown that simple regression methods fail in terms of prediction accuracy, outlier detection and model robustness while higher-sophisticated regression methods are almost able to constantly achieve these goals. Due to quite similar but still not optimal prediction performance as well as limited computational feasibility in case of numerous input parameters, the choice of superior ML regression methods does not ultimately resolve the problem. Considering the entire cycle of Knowledge Discovery in Databases including Data Mining (DM) another task appears to be crucial: FS. An optimal selection of the decisive parameters and hence reduction of the input space dimension boosts the model performance by omitting redundant as well as spurious information. Various FS algorithms exist to deal with correlated and noisy features, but each of its own is not capable to ensure that the ambitious targets for VM can be achieved in prevalent high-mixture-low-volume SM. The objective of the present doctoral thesis is the development of a smart FS algorithm to enable a by this advanced and also newly developed VM system to comply with all imperative requirements for improved process stability and control. At first, a new Evolutionary Repetitive Backward Elimination (ERBE) FS algorithm is implemented combining the advantages of a Genetic Algorithm (GA) with Leave-One-Out (LOO) Backward Elimination as wrapper for Support Vector Regression (SVR). At second, a new high performance VM system is realized in the productive environment of High Density Plasma (HDP) Chemical Vapor Deposition (CVD) at the Infineon frontend manufacturing site Regensburg. The advanced VM system performs predictions based on three state of the art ML methods (i.e. Neural Network (NN), Decision Tree M5’ (M5’) & SVR) and can be deployed on many other process areas due to its generic approach and the adaptive design of the ERBE FS algorithm. The developed ERBE algorithm for smart FS enhances the new advanced VM system by revealing evidentially the crucial features for multivariate nonlinear regression. Enabling most capable VM turns statistical sampling metrology with typically 10% coverage of process results into a 100% metrological process monitoring and control. Hence, misprocessed wafers can be detected instantly. Subsequent rework or earliest scrap of those wafers result in significantly increased stability of subsequent process steps and thus higher yield. An additional remarkable benefit is the reduction of production cycle time due to the possible saving of time consuming physical metrology resulting in an increase of production volume output up to 10% in case of fab-wide implementation of the new VM system.