Efficient spatial variation modeling via robust dictionary learning

In this paper, we propose a novel spatial variation modeling method based on robust dictionary learning for nanoscale integrated circuits. This method takes advantage of the historical data to efficiently improve the accuracy of wafer-level spatial variation modeling with extremely low measurement cost. Robust regression is adopted by our implementation to reduce the bias posed by outliers. An iterative coordinate descent method is further introduced to solve the dictionary learning problem with consideration of missing data. Our numerical experiments based on industrial measurement data demonstrate that the proposed method achieves up to 70% error reduction over the conventional VP approach without increasing the measurement cost.