Explainable AutoML (xAutoML) with adaptive modeling for yield enhancement in semiconductor smart manufacturing

TL;DR

This work tackles the challenge of achieving high semiconductor yield in smart manufacturing with an explainable, domain-specific AutoML approach. It presents xAutoML, a framework that combines knowledge-informed massive feature extraction, model-agnostic feature selection via CAST, unsupervised anomaly detection, and adaptive, focal-loss-based defect classification, all orchestrated by BOHB optimization to efficiently search configurations. The approach yields 39 key features from over 60,000 engineered candidates, detects anomalies, and achieves a defect-classification accuracy of 92.89% on the SECOM dataset, while maintaining interpretability through explainability tools. The results demonstrate practical potential for yield improvement, robust defect diagnosis, and adaptive, explainable automation in semiconductor manufacturing, with clear pathways for on-site deployment and cognitive extension via AI agents.

Abstract

Enhancing yield is recognized as a paramount driver to reducing production costs in semiconductor smart manufacturing. However, optimizing and ensuring high yield rates is a highly complex and technical challenge, especially while maintaining reliable yield diagnosis and prognosis, and this shall require understanding all the confounding factors in a complex condition. This study proposes a domain-specific explainable automated machine learning technique (termed xAutoML), which autonomously self-learns the optimal models for yield prediction, with an extent of explainability, and also provides insights on key diagnosis factors. The xAutoML incorporates tailored problem-solving functionalities in an auto-optimization pipeline to address the intricacies of semiconductor yield enhancement. Firstly, to capture the key diagnosis factors, knowledge-informed feature extraction coupled with model-agnostic key feature selection is designed. Secondly, combined algorithm selection and hyperparameter tuning with adaptive loss are developed to generate optimized classifiers for better defect prediction, and adaptively evolve in response to shifting data patterns. Moreover, a suite of explainability tools is provided throughout the AutoML pipeline, enhancing user understanding and fostering trust in the automated processes. The proposed xAutoML exhibits superior performance, with domain-specific refined countermeasures, adaptive optimization capabilities, and embedded explainability. Findings exhibit that the proposed xAutoML is a compelling solution for semiconductor yield improvement, defect diagnosis, and related applications.

Figures (6)

• Figure 1: Adaptive xAutoML framework for yield enhancement
• Figure 2: Performance variation trend of CAST(a) and Normalized weight ratio of the current optimal result(b)
• Figure 3: RFE process of CAST(a) and single feature selection method(b)
• Figure 4: Performance change of classification algorithms by different optimizer
• Figure 5: Tuning rules of the maximum performance curves of different optimizers