Improving the accuracy of Design for Manufacturability (DFM) sign-off checks with machine learning

Industry’s first application of supervised Machine Learning (ML) for improving Design for Manufacturability (DFM) physical verification to target real errors more accurately for designers is presented, which has been implemented leveraging the Calibre platform. Machine learning assisted DFM checks are developed to identify metal-via enclosure design weak points. Using retargeting simulations, rule-based prefiltering localizes geometric weak point configurations. Next, they are extracted from layout clips as image snippets to form density vectors that are then used as feature vectors for ML model training. Finally, a Convolutional Neural Network (CNN) model is developed to predict the post-retargeting, via-metal enclosure weak points given input layout designs. Compared to standard DFM methodologies, machine learning assisted DFM can more precisely capture critical rules. Results show that for six, 22nm layout designs containing a sample size of 861k unique design weak points, ML-DFM outperforms standard DFM physical verification by a ~20% improvement in accuracy, reducing the number of false positives, such that only the most critical violations are identified to aid design fixing, shortening design cycle time.