The on-product overlay specification and Advanced Process Control (APC) are getting extremely challenging
particularly after the introduction of multi-patterning applications like Litho-Etch-Litho-Etch (LELE). While the Reticle
Writing Error (RWE) contribution could be marginalized for quite some time in the layer-to-layer overlay budget, it will
become one of the dominating overlay contributors when the intra-layer overlay budget is considered. While most of the
overlay contributors like wafer processing, scanner status, reticle transmission, dose, illumination conditions drop out of
the intra-layer overlay budget, this is certainly not the case for reticle to reticle writing differences.
In this work, we have studied the impact of the RWE on the on-product overlay performance. We show that the RWE
can be characterized by an off-line mask registration tool and the modelled results can be sent as feed-forward
corrections to the ASML TWINSCANTM. By doing so, the overlay control complexity (e.g. send-ahead wafers, APC
settling time) can be reduced significantly. Off-line characterization enables that all reticles virtually become equal after
correction (at least to the level of correction capability of the scanner). This means that all higher order RWE
contributions (currently up to a third order polynomial) can be removed from the fingerprint. We show that out of 50
production reticles (FEOL, 28-nm technology), 30% can be improved on residual level when non-linear feed-forward
corrections are considered as well. The additional benefit of feeding forward linear corrections to the scanner is even
higher: it is anticipated that a large portion of the APC variation might find its origin in the RWE contribution.
In order to send feed-forward corrections to the scanner, we obviously rely on the quality of the off-line RWE
measurements. These measurements are usually provided by a registration tool at the mask shop. To secure the quality,
an independent experimental verification test was developed to check if off-line RWE measurements can be used as
feed-forward corrections to the scanner. The test has been executed on an ASML NXT: 1950i scanner and was designed
such to isolate the reticle writing error contribution. The match between the off-line measurements and the experiment is
striking.
|