It is desired to reduce the time required to produce metrology data for calibration of Optical Proximity Correction (OPC) models and also maintain or improve the quality of the data collected with regard to how well that data represents the types of patterns that occur in real circuit designs. Previous work based on clustering in geometry and/or image parameter space has shown some benefit over strictly manual or intuitive selection, but leads to arbitrary pattern exclusion or selection which may not be the best representation of the product. Forming the pattern selection as an optimization problem, which co-optimizes a number of objective functions reflecting modelers’ insight and expertise, has shown to produce models with equivalent quality to the traditional plan of record (POR) set but in a less time.
Assessing pattern printability in new large layouts faces important challenges of runtime and false detection. Lithographic simulation tools and classification techniques do not scale well. We propose a fast pattern detection method that builds jointly a structured overcomplete basis, representing each reference pattern, and a linear predictor of their lithographic difficulty. A pattern from a new design is detected “novel” if its reconstruction error, when coded in the learned basis, is large. This allows a fast detection of unseen clips and a fast prediction of their lithographic difficulty. We show high speedup (1000×) compared to nearest neighbor search, and very high correlation between predicted and calculated lithographic estimate values.
Many chip design and manufacturing applications including design rules development, optical proximity correction
tuning, and source optimization can benefit from rapid estimation of relative difficulty or printability. Simultaneous
source optimization of thousands of clips has been demonstrated recently, but presents performance challenges. We
describe a fast, source independent method to identify patterns which are likely to dominate the solution. In the context
of source optimization the estimator may be used as a filter after clustering, or to influence the selection of representative
cluster elements. A weighted heuristic formula identifies spectral signatures of several factors contributing to difficulty.
Validation methods are described showing improved process window and reduced error counts on 22 nm layout
compared with programmable illuminator sources derived from hand picked patterns, when the formula is used to
influence training clip selection in source optimization. We also show good correlation with fail prediction on a source
produced with hand picked training clips with some level of optical proximity correction tuning.
Joint optimization (JO) of source and mask together is known to produce better SMO solutions than sequential
optimization of the source and the mask. However, large scale JO problems are very difficult to solve because the global
impact of the source variables causes an enormous number of mask variables to be coupled together. This work presents
innovation that minimize this runtime bottleneck. The proposed SMO parallelization algorithm allows separate mask
regions to be processed efficiently across multiple CPUs in a high performance computing (HPC) environment, despite
the fact that a truly joint optimization is being carried out with source variables that interact across the entire mask.
Building on this engine a progressive deletion (PD) method was developed that can directly compute "binding
constructs" for the optimization, i.e. our method can essentially determine the particular feature content which limits the
process window attainable by the optimum source. This method allows us to minimize the uncertainty inherent to
different clustering/ranking methods in seeking an overall optimum source that results from the use of heuristic metrics.
An objective benchmarking of the effectiveness of different pattern sampling methods was performed during postoptimization
analysis. The PD serves as a golden standard for us to develop optimum pattern clustering/ranking
algorithms. With this work, it is shown that it is not necessary to exhaustively optimize the entire mask together with the
source in order to identify these binding clips. If the number of clips to be optimized exceeds the practical limit of the
parallel SMO engine one can starts with a pattern selection step to achieve high clip count compression before SMO.
With this LSSO capability one can address the challenging problem of layout-specific design, or improve the technology
source as cell layouts and sample layouts replace lithography test structures in the development cycle.