Presentation
CMP-Aware Dummy-Fill Optimization via Generative Modeling and Measurement-Calibrated Prediction
DescriptionConventional rule-based dummy-fill methods satisfy density constraints but do not directly minimize chemical–mechanical polishing (CMP) variation, requiring costly post-layout fixing. We propose a learning-based dummy-fill insertion framework that enables proactive correction at the design stage by coupling a measurement-calibrated (MCAL) CMP predictor with a polygon-based conditional multilayer (PCM) fill generator. Trained on measurement data, the predictor, enhanced with shift correction for systematic offsets, achieves higher accuracy and efficiently identifies regions of high CMP variation. Leveraging PCM sequencing with blockwise causal attention, the generator produces diverse, design-rule-compliant multilayer fills conditioned on local active patterns. Coupling the generator and predictor enables design-space exploration that identifies optimal fills and reveals feature-CMP relationships guiding future fill strategies. The framework achieves substantially lower CMP variation on advanced-node benchmarks, demonstrating its effectiveness for practical deployment in advanced design flows.
Event Type
Work in Progress
TimeMonday, July 276:51pm - 6:52pm PDT
LocationExhibit Hall