Explainable AutoML (xAutoML) with Adaptive Modeling for Yield Enhancement in Semiconductor Smart Manufacturing
Weihong Zhai, Qing Kai Han, Lisheng Chen, Xiupeng Shi
Abstract
Yield enhancement is recognized as an ultimate challenge to reducing production costs in semiconductor smart manufacturing (SSM), although there are intrinsic challenges, especially reliable yield diagnosis and prognosis, and understanding the confounding factor in a complex condition, among others. This study proposes a domain-specific explainable automated machine learning (termed xAutoML) to self-learn the optimal models for yield prediction, with an extent of explainability, and also provide insights on key diagnosis factors. The xAutoML integrates the main problem-solving functionalities in an auto-optimization pipeline, and each functionality is designed to fit the domain challenges. Firstly, to capture the key diagnosis factors, knowledge-informed feature extraction and model-agnostic key feature selection are designed. Secondly, combined algorithm selection and hyperparameter tuning with adaptive loss are developed to deliver the optimized classifiers for better defect prediction, as well as adaptively updating according to pattern shift. Moreover, a series of explainable manners are provided throughout the AutoML pipeline, aiming to offer proper information for better understandability and trust. Compared with existing general-purpose AutoML solutions, the proposed xAutoML is domain-specific optimized with targeted countermeasure designs to achieve better performance, auto-optimization, and explainability, which is promising for practical semiconductor yield enhancement, defect diagnosis, and related applications.