CFM hotspot detection using CNN models

Addressing contamination-free manufacturing (CFM) hotspots that result from lithography and etch defects is an integral part of the semiconductor manufacturing process. During the early phases of technology development, we rely on optical proximity correction (OPC) models to predict lithography and etch defects. However, once the technology reaches a certain level of maturity, new defects are no longer predictable by OPC verification flows due to factors we could not account for while building these models, such as long-range effects, and we must rely on CFM to detect these defects and fix them inside OPC. This iterative process slows down development and increases the overhead cost of new masks and silicon wafer processing. In this work, we propose a reliable framework to detect CFM hotspots and yield detractors for back end of the line (BEOL) levels using in-house convolutional neural networks (CNN) models. These models could be incorporated into our verification flows and executed during tape-out before mask fabrication. This new flow accelerates the technology development process and eliminates extra mask cost and silicon processing by detecting these defects earlier in the process. The Scope of our work is not limited to leveraging machine learning (ML) models, but it extends to the features we use to calibrate our models, and the CFM data preparation techniques deployed to remove random defects. These models could detect silicon issues we could not predict using regular OPC models.