To effectively and efficiently in-line real time monitor the potential multi-layer weak patterns, we quantify the bridge electrical metric between contact and gate electrodes into CD physical metric via big data from the larger field of view (FOV: 8k x 16k with 3 nm pixel equalizes to image main field size 34 um x 34 um @ 3 nm pixel) e-beam quality image contour compared to layout GDS database (D2DB) as shown in Fig. 1. Hadoop-based distributed parallel computing is implemented to improve the performance of big data architectures, Fig. 2. Therefore, the state of art in-line real time catching in-chip potential multi-layer weak patterns can be proven and achieved by following some studying cases [3]. Therefore, manufacturing sources of variations can be partitioned to systematic and random variations by applying statistical techniques based on the big data fundamental infrastructures. After big data handling, the in-chip CD and AA variations are distinguished by their spatial correlation distance. For local variations (LV) there is no correlation, whereas for global variations (GV) the correlation distance is very large [7]-[9]. This is the first time to certificate the validation of spatial distribution from the affordable bias contour big data fundamental infrastructures. And then apply statistical techniques to dig out the variation sources. The GV come from systematic issue, which could be compensated by adaptive LT condition or OPC correction. But LV comes from random issue, which being considered as intrinsic problem such as structure, material, tool capability… etc.
In this paper studying, we can find out the advanced technology node SRAM contact CD local variation (LV) dominates in total variation, about 70%. It often plays significant in-line real time catching WP-DPMO role of the product yield loss, especially for wafer edge is the worst loss within wafer distribution and causes serious reliability concern. The major root cause of variations comes from the PR material induced burr defect (LV), the second one comes from GV enhanced wafer edge short opportunity, which being attributed to three factors, first one factor is wafer edge CD deliberated enlargement for yield improvement as shown in Fig. 10. Second factor is overlaps/AA shifts due to tool capability dealing with incoming wafer’s war page issue and optical periphery layout dependent working pitch issue as shown in Fig. 9 (1)., the last factor comes from wafer edge burr enhanced by wafer edge larger Photo Resistance (PR) spin centrifugal force.
After implementing KPIs such as GV related AA/CD indexes as shown in Fig. 9 (1) and 10, respectively, and LV related burr index as shown in Fig. 11., we can construct the parts per million (PPM) level short probability model via multi-variables regression, canonical correlation analysis and logistic transformation. The model provides prediction of PPM level electrical failure by using in-line real time physical variation analysis. However in order to achieve Total Quality Management (TQM), the adaptive Statistical Process Control (SPC) charts can be implemented to in-line real time catch PPM level product malfunction at manufacturing stage. Applying for early stage monitor likes incoming raw material, Photo Resistance (PR) … etc., the LV related burr KPI SPC charts could be a powerful quality inspection vehicle. To sum up the paper’s contributions, the state of art in-line real time catching in-chip potential multi-layer physical weak patterns can be proven and achieved effectively and efficiently to associate with PPM level product dysfunction.
