Enhancing multilayer process defect prediction accuracy on an artificial intelligence/machine learning (AI/ML) platform

Artificial Intelligence/Machine Learning (AI/ML) multi-layer hotspot prediction has been proven useful with design houses to capture potential process risky patterns for the most advanced technology modes. Since the design houses only have design target layers from the layouts, all the signature information used in the AI/ML training for hotspots is from the target polygons. Given the fact that most of the hotspots are process-related with single or multiple layers, and the geometric or topological information itself from the target polygons lacks process knowledge, the prediction could result in high false detection pattern rate or miss some real defects. One way to mitigate the problem is to design feature vectors which characterize defect patterns and reflect the process impact. It is a relatively easy and fast way to successfully capture hotspots of interest. To further improve the hotspot detection, the process information may be added directly. In this study, a process-aware layout is generated to include process information in the layout, and it is then used as input to the hotspot prediction. To generate the process-aware layout, a ML model is trained on design target layout and final etch shapes first. The trained model can predict the etch shapes from the design layout. The predicted curve-linear shapes are then converted into rectilinear shapes to create the process-aware layout. Another perspective of reducing the false alarm is to fine tune the AI/ML model. A reinforcement learning model has been developed to reduce the false detection from the supervised hotspot prediction. The false alarm reduction has been successfully demonstrated on a V2-M2 short case. Both of the above-mentioned enhancements have been built on the previously developed featurebased AI/ML platform.