Predicting variability in nanoscale lithography processes

High performance lithographic hotspot detection using hierarchically refined machine learning

“…In the far right-hand columns of Table 6, our proposed methods demonstrate better performance with superior run-time under real manufacturing conditions. With similar or slightly better hotspot detection rate Hhit of 82%-89%, we show hotspot false-alarm reductions ranging from 2.4X (between [19] and MLK-ANN-GD+LR) to 2300X (between [13] and MLK-SVM-GD+LR). Simulation run-time speed-ups range from 5X (between [15] and MLK-SVM) to 237X (between [13] and MLK-ANN), when calibred in CPU hour per layer per mm 2 unit.…”

“…Later in [13], a support vector machine (SVM) based hotspot detection method is utilized through performing 2D distance transform and histogram extraction on pixel based layout images. Similarly in [14], SVM is employed for hotspot detection through extraction and classification of certain special layout density related metrics.…”

“…Similarly in [14], SVM is employed for hotspot detection through extraction and classification of certain special layout density related metrics. As an improvement over [11,12], [13,14] demonstrate higher detection accuracy and lower classification noise, due to the high fidelity metrics introduced. However, potential issues with such approaches lie in run time and detection coverage, since 2D transforms and density extractions can be expensive to perform, meanwhile detection windows for the layout images are hard to anchor for full chip level detections.…”

“…In [15], critical hotspot signature is proposed and extracted through certain special edge-based metrics. Although such edge-based extractions operate much faster compared with [13,14], its chip level application still faces similar issues such as scanning window coverage, etc.…”

“…In order to be practically employed in modern IC physical design, a successful hotspot detection engine must demonstrate superior speed compared to full lithography simulation (> 100 CPUs running in the order of days) and DRC (tens of CPUs running for a few hours), as well as comparable performance to meet the real design and manufacturing requirements. Unfortunately under such situations, the hotspot evaluation models in [11][12][13][14][15] suffer from severe performance degradation, since real hotspot detection under real manufacturing conditions requires more than a straightforward model, but rather, multi-level detection models with hierarchies of performance refinement.…”