Smart implant-layer overlay metrology to enable fab cycle time reduction

摘要

Overlay is one of the most critical design parameters in integrated circuit manufacturing. Maintaining goodoverlay performance during manufacturing is therefore essential in order to obtain high yield and to ensurethat the performance and reliability of the eventual semiconductor device is according to specifications. Forthat reason, optical metrology is nowadays extensively used in any production facility for overlay monitoringand process control. Overlay metrology is typically required after each lithography step for (nearly) every lot.The number of process and lithograpy steps have increased signigoficantly with advancing technology nodes andconsequently there is an increased demand for overlay metrology. Although the benets of overlay metrologyare obvious, the use of metrology should be kept at acceptable levels as it adds cost and increases fab cycletime. Virtual overlay metrology, the replacement of some real overlay measurements with predicted values, is aneu000bective solution for keeping the need for overlay metrology under control.In this work, we develop virtual overlay metrology for a series of nine implant layers. These nine layers areexposed by multiple scanners and an implant layer is not necessarily exposed by a single scanner. We usemachine learning algorithms to build prediction models for the implant-layer overlay. In particular, the modelsare built using neural networks with a set of specically engineered overlay predictors that capture the essenceof the physical concepts of the implant-layer overlay. The inclusion of domain expertise via these engineeredoverlay predictors is essential for achieving a model with high prediction capability.The capability of virtual overlay metrology is evaluated on production data. We show that the overlay predictionperformance is around 0.7 with respect to R2 statistics, and that virtual metrology is able to follow variationsin overlay and to identify outliers. We conclude that the overlay prediction model is suu000eciently accurate for theimplant layers under consideration and we estimate that a metrology reduction of about 70% can be achieved ifvirtual overlay metrology would be enabled in high-volume manufacturing.Virtual metrology can also be used for overlay control. A model can achieve high prediction capability (R2 → 1)only when the overlay levels exceed the baseline scanner performance. Once the root causes of the elevatedoverlay levels are eliminated, the prediction performance will be low again (R2 → 0). Overlay can be controlledby monitoring as many predictors as possible that may have a link with overlay. When virtual overlay metrologystarts to predict overlay levels above the scanner baseline performance, and correlation is found between measuredand predicted overlay (R2 → 1), actions should be triggered towards eliminating or controlling the root causes.