Uncertainty Evaluation and Validation Based on Enhanced Adaptive Bayesian Monte Carlo Dropout Method in TSOM

The complete measurement of optical metrology is crucial in integrated circuits (ICs), including mean values and uncertainties. Through-focus scanning optical microscopy (TSOM) is an optical nondestructive nanoscale measurement IC technique. However, the research on uncertainty evaluation is insufficient in TSOM. The application of Monte Carlo (MC) Dropout in TSOM cannot balance the measurement accuracy and uncertainty evaluation. In this article, we propose an enhanced adaptive Bayesian MC Dropout method with a novel loss function incorporating learnable parameters. This approach simultaneously evaluates data and model uncertainties while optimizing the precision-uncertainty tradeoff through adaptive parameter learning. Then, we propose a verification method to validate the proposed uncertainty evaluation method. Results demonstrate our methods provide uncertainty quantification while preserving measurement accuracy. The proposed methods are also potential candidates for uncertainty evaluation and validation in other model-based measurement techniques.