A multimodal hierarchical learning approach for virtual metrology in semiconductor manufacturing

Abstract

Achieving high-precision wafer yield prediction is a crucial step in improving the quality of semiconductor manufacturing. However, existing methods overlook the multimodal characteristics in wafer fabrication, leading to limitations in prediction accuracy and interpretability. To address the problem, this paper proposes an adaptive modal division and hierarchical learning method for wafer yield prediction. Firstly, Bayesian optimization is employed to adaptively search for the optimal modal division locations in the training samples, categorizing the samples into three distinct yield groups (high, medium, and low) with explicit production relevance. Concurrently, a novel degradation and incremental learning mechanism is designed to address the problem of declining prediction accuracy due to sample imbalance. Subsequently, a classification-regression hierarchical learning architecture is established to separately learn the distribution characteristics of each modality. This involves training classifiers using the labels derived from modal division, followed by distinct regressors for each category to facilitate precise yield predictions. Finally, experimental validations based on simulation and real-world manufacturing data demonstrate that the proposed virtual metrology approach accounting for multimodal characteristics exhibits enhanced performance and robustness.