Fed-SEMI: A Federated Machine Learning Framework for Nano-Scale Defect Classification and Detection in Semiconductor Manufacturing with Decentralized and Private Data

In this research, we introduce a decentralized Federated Learning (FL) framework, as Fed-SEMI, which utilizes the YOLOv8 object detection model to facilitate advanced node semiconductor defect inspection by addressing the challenges associated with conventional “individual” or “centralized” supervised learning. Our objective is to develop a universal defect inspection model by securely leveraging the diversity among multiple datasets, without the need for any exchange of sensitive data. To the best of the authors' knowledge, this framework represents the first implementation of a federated learning approach in the semiconductor manufacturing domain. It utilizes real FAB-originated decentralized and private datasets (heterogenous and non-IID) from multiple participants, sourced from distinct processes and captured using different imaging tools. We investigated two novel weight-aggregation algorithms, namely FedHarmonic and FedArithmetic, against conventional FedAvg, along with an improved data augmentation strategy aimed at generating diverse and realistic training datasets to enhance defect inspection performance. Finally, we benchmarked our proposed framework on four datasets—LS-ADI,LS-AEI, LS-EDR, and HEXCH-DS—under various experimental conditions to demonstrate its capability in ensuring data privacy and security while achieving collaborative learning, scalability, and efficiency.