Efficient surrogate modeling for accelerated generalized ellipsometry in nanostructure metrology

Abstract Generalized ellipsometry (GE) measurement is one of the mainstream techniques for online metrology and characterization of various nanoscale thin films and devices in semiconductor manufacturing. However, its practical application is often limited by the high computational cost of electromagnetic field simulations and iterative inversion processes. To address this challenge, this study investigates the efficient surrogate model for GE to replace the expensive simulation with fast and accurate approximations. Additionally, conventional static sampling strategies require substantial computational resources to generate training datasets. To alleviate this burden, an improved and efficient query-by-committee (QBC) method is proposed for realizing adaptive sampling. In this method, the technique for order preference by similarity to the ideal solution (TOPSIS) is introduced to objectively assign weights to committee members based on their predictive reliability. Both simulation and experimental results demonstrate that the proposed method can achieve high-precision ellipsometry response approximation (correlation coefficient R2 > 0.98) and efficient dimensional inversion, with an average efficiency improvement of 85.91% compared to conventional approaches without surrogate model.