A concise subspace projection based meta-learning method for fast modeling and monitoring in multi-grade semiconductor process

Abstract

Modern semiconductor industries produce multiple grades of wafers to cater to the diversity of market requirements. Commonly, products with new specifications are manufactured by merely altering the input materials or operations. How to promptly establish an effective monitoring framework for a new grade of semiconductor process is a practical and challenging issue. The difficulty lies in the fact that very limited or even no samples are available for reliable training in a process for the new-grade product, and the new process is not the same as any previous process. An orthotropic subspace projection-based meta-learning method is thus proposed for fast modeling and monitoring with respect to multi-grade processes. With the available historical data sets, the shared common and special features of the meta model are extracted. The common features can be applied directly to processes for products of other grades while the special features are various from each other. The validation data sets are utilized to determine the modified special features and formulate monitoring models for online applications. By comparison, 90% of the fault samples can be detected by the proposed method, much higher than the classical meta-learning method, whose fault detection rate is 83.5%, in the numerical case. For the involved industrial semiconductor process, all the fault samples can be detected, showing obvious advantages over the other related methods.