Virtual defect metrology system utilizing photolithography scanner topography maps to quantitatively detect gross defects, perform inline yield prediction, and identify defect-generating tools

BackgroundIn semiconductor high-volume manufacturing fabs, the processing of a large number of wafers daily necessitates stringent defect metrology strategies. Traditional methodologies, constrained by throughput limitations, sample only a small percentage of wafers to minimize efficiency losses, resulting in significant coverage gaps between process steps. This restricted coverage fails to capture all gross defect modes generated across various process tools, subsequently prolonging the timeline for defects’ root cause analysis (RCA).AimOur aim is to introduce an innovative inline virtual metrology (VM) system designed to improve gross defect detection, leveraging wafer topography data from photolithography scanner sensors.ApproachOur proposed system comprises three integral components. The first component is the wafer image generation flow, which converts wafer topography tool sensor data into two-dimensional heat-map images using traditional mathematical modeling techniques. The second component is the defect classification and yield impact estimation flow, which employs a deep learning model to detect and classify defects within these images. Utilizing computer vision techniques, defect morphologies are quantitatively extracted, enabling inline die impact estimation by analyzing the affected pixels and their corresponding wafer coordinates. The third component is the automated RCA flow, which clusters similar defect modes and uses probabilistic analysis and ranking techniques to identify likely defect-generating process tools.ResultsThe inline VM system effectively identifies common gross defects and correlates impacted dies to historical yield data, enabling early inline defect detection. Comprehensive gross defect monitoring is achieved for every wafer and die at each photolithography step, thereby enhancing detection coverage and bridging existing gaps. The automated RCA significantly improves the capability to quickly identify defect sources, facilitating early detection and intervention.ConclusionsBy utilizing VM at each photolithography step, our system ensures exhaustive gross defect monitoring, thus promising substantial advancements in early inline detection and yield improvement. The automated RCA further enhances the capability to swiftly pinpoint defect sources, promoting effective early detection and intervention strategies.