This work compared the CD-based and image-assistant approaches for calibrating the OPC models. OPC models were
first developed for 65nm-node memory contact layer and calibrated by contact test patterns with various ellipticities. The
image-assistant model is a hybrid one calibrated by SEM contours and 1D measurement results, while the CD-based
model calibration uses 1D measurement results as the sole data source. The fitting errors, model prediction ability and
OPCed results were compared between these two models. Besides, the challenges on calibrating the edge-detection
algorithm of the CD SEM images to the extracted contours of OPC tool were also discussed. Finally, the layouts
corrected by CD-based and image-assistant models were written on a test mask for wafer-level comparison.
The results displayed that the CD-based model showed smaller error on fitting and interpolation, but image-assistant
model got improvement on extrapolation prediction of array-edge contact, unknown contact pattern and long contacts.
The wafer-level comparison also revealed the image-assistant model outperformed to the CD-based model by smaller
correction error on unexpected patterns.
The low k1-factor challenge in current photolithography has made OPC (Optical Proximity Correction) indispensable for critical patterning layers, and more efforts are needed in the development and calibration of OPC model. One of the key factors that affect the accuracy of wafer result is the accuracy of OPC model, and usually, only a few nanometers' fitting residual of OPC model is tolerable. So, several methods have been reported for improving the accuracy of OPC modeling, but the model fitting becomes more complex as the increase of fitting parameters accordingly.
In this paper, the variable loading kernel to manipulate the behavior of OPC modeling was reported. The variable load kernel can be modified by space domain, and it also can be the combination of many load kernels, such as Kload= a1*Kload1 + a2*Kload2 + ...... + an*Kloadn. By combining of different variable load kernels, the resultant load kernel can be more flexible to manage the model behavior in different line widths and pitches. In the example of OPC fitting residuals of linearity patterns, it is obvious that the different models with different loading kernels yielded different residuals. The use of variable loading kernel achieves the satisfied small residuals for both small and large patterns simultaneously. Accordingly, easier OPC modeling with smaller fitting residual is anticipated by variable load kernel method.