Background
Line-end-pull-back (LEPB) is a well-known systematic defect in BEOL metal layers, where a line-end (LE) tip is pulled back from its desired location due to lithography (litho) process effects. Severe LEPB directly affects BEOL connectivity and may lead to partial or total metal-via disconnection.
Aim
LEPB can be characterized through model-based litho simulations but at the cost of high computational resource consumption. This study aims to provide a fast and accurate approximation of computationally expensive litho simulations through regression-based machine learning (ML) modeling.
Approach
LEPB modeling is approached through the LightGBM model. Input features were approached using density pixels, density concentric circle area sampling (CCAS), and geometrical positioning surveying (GPS), which is an edge-based engine that provides a direct one-to-one mapping between model features and geometrical measurements between the LE as a point-of-interest and its surrounding contextual patterns. The importance of LightGBM features by splits was employed to reduce features across the used approaches.
Results
The reduced features of GPS produced almost the same modeling quality (training: RMS = 0.571 nm, δEWD = 0.297 nm, and R2 % = 98.74 % , and testing: RMS = 0.643 nm, δEWD = 0.344 nm, and R2 % = 98.40 % ) with −22.22 % fewer number of features and less feature extraction runtime compared to the full features set of density pixels and density CCAS approaches.
Conclusions
Compared to model-based litho simulations, the obtained calibrated ML models can be used to provide fast, yet accurate predictions of the amounts of pull-back or extensions introduced at LEs near vias, eliminating a major contributor to systematic IC yield loss.
