Loading effect correction set up by supplementing CD measurement analysis with machine learning

With semiconductor technology approaching and exceeding 10 nm design rules the quality requirements for photomasks are continuously tightening. One of the crucial parameters is improved control of the critical dimension (CD) across the photomask. As long as linearity and through pitch effects are not involved, the quality measure is typically defined as CD uniformity. This parameter is normally measured on repeating structures of same size and shape, which are not necessarily placed in identical environments. Density dependent process effects, also called loading effects (LE), pose a challenge for the required CD control. There are several possible contributors to this kind of error within the mask manufacturing flow, such as etch driven loading effects, fogging effects during 50kV exposure and develop driven loading effects. All of these operate at different working ranges, starting at millimeters going down to only a few 100 μm scale. It is comparably easy to derive models for large scale phenomena like etch loading or fogging effects, in contrast to that it is not as straight forward to find suitable models for very short-range effects. A large amount of CD measurements taken by CD SEM is needed to identify such signals of low magnitude and short scales, which make the setup very resource intensive. Furthermore, this methodology requires artificial designs and test structures which aim to sample only the effect of interest. In this paper we present a strategy which combines CD SEM measurements from dedicated test masks with the results from regular product masks. The aim is the derivation and validation of the loading effect correction range and strength. In the first step the data from test masks is analyzed to set up the basic correction parameters. Following this, the approach is supplemented by product data where we combine mask CD and design data. The clear field distribution of the design is convoluted with respect to a hierarchy of length scales. This data is the input for a support vector machine analysis. Thus, we employ a flat machine learning algorithm. However, the input data has been set up to reflect multiple layers of convolution. This particular approach has been chosen, as each convolution length scale is associated with mask process properties, thus alleviating the burden of interpretation which typically mars the interpretation of models obtained by machine learning approaches.