Ensemble Learning-Fused Solution to the Inverse Problem in Integrated Optical Critical Dimension Metrology

In sub-28-nm nodes, existing optical critical dimension (OCD) metrology tools are challenging to simultaneously and thoroughly meet the metrology requirements of the wafer-level nanostructures in terms of measurement precision, efficiency, and robustness. To tackle the challenge, we propose a Mueller matrix ellipsometry (MME)-based integrated OCD metrology (IM OCD) technique to achieve fast, accurate, and robust in-line metrology of wafer-level nanostructures at advanced nodes. The MME probe is integrated into the integrated circuit (IC) process platform via a polar-coordinate stage, allowing high-throughput measurement of the geometrical parameters of nanostructures with 144 wafers per hour. Meanwhile, we propose an ensemble learning-fused regression (ELFR) approach to improve the solution to the typical inverse problem in nanostructure metrology, effectively addressing the fluctuations in measurement sensitivity and precision caused by the measurement azimuth variations in the IM OCD. The essential of the ELFR approach is the weighted averaging of the respective results from ridge regression (RR) based on the Levenberg-Marquardt algorithm and from bagging-mode neural networks (BMNNs). It improves the metrology results of wafer-level nanostructures to an ultrahigh coefficient of determination better than 0.93, an average deviation, and a maximum deviation less than 1.0 and 2.0 nm, respectively. Characterizations of wafer-level shallow trench isolation (STI) and interconnect-layer nanostructures have demonstrated the feasibility and effectiveness of the proposed MME-based IM OCD method, especially the capability of the ELFR approach to suppress azimuthal-dependent errors.