In the semiconductor industry, process monitoring or fault detection and classification (FDC) has become a critical component of the manufacturing system in order to meet the challenges posed by ever decreasing tolerance of process variation and error. Because process variations often depend on the specific product being produced as well as the specific manufacturing tools being used, to address the dependence of product quality on manufacturing context, currently deployed process monitoring techniques are “threaded” methods [1]. In the last decade, high-mix is becoming the standard manufacturing paradigm in response to rapid and frequent changes in factory plans and process recipes. This trend makes it increasingly difficult for existing FDC techniques to effectively meet manufacturers' requirements [2]. This is because current practice of FDC, i.e., “threaded” FDC, utilizes only part of the historical data, i.e., the ones from the same tool and the same product to build monitoring models. Threaded FDC addresses tool and product variability explicitly, however, in high-mix manufacturing, threaded FDC results in: (1) significant number of models that increases in proportion to the number of product-tool combinations; (2) insufficient data to build models for low-running threads. Although non-threaded (i.e., using data across tools and products) state estimation has drawn some research interest, little attention has been given to non-threaded process monitoring.

In this work, a novel correlation pattern based non-threaded FDC framework is developed to address the challenges posed by high-mix manufacturing. The basic idea of the non-threaded FDC framework is based on the following factor: although normal process data usually show considerable variation across different tools, different products and over time, the correlation patterns (i.e., covariance structures) among different variables are much more stable compared to process mean, due to physical and chemical principles governing the process operation, such as mass and energy balances, transport, kinetics and thermodynamics. Therefore, instead of monitoring the process mean that changes from batch to batch, in the developed approach, we monitor the correlation patterns that are stable from batch to batch to detect and classify process abnormalities.

Because different threads of a same process share a single recipe, they are governed by the same physical and chemical principles, and the data correlation patterns are independent of threads. Therefore, the developed fault detection framework is a true non-threaded fault detection method and only one model is needed for one process/recipe. For similar reasons, the proposed fault detection framework does not suffer from multimodal batch trajectories. In addition, the correlation pattern based monitoring approach does not require data unfolding and there is no need to equalize the unequal batch and step lengths, because data in each step of each batch will generate a correlation pattern matrix whose dimension is determined by the number of the variables instead of step durations.

Both fault detection and classification approaches are developed within the correlation based framework. The effectiveness of the developed fault detection approach is demonstrated using both simulation examples, and industrial data (TI etch data set [3]). Analysis results show the data correlation pattern of multimodal, unequal step length batches are well conserved.

Reference

[1] J. Wang, Q.P. He and T.F. Edgar (2008), “On state estimation of high-mix semiconductor manufacturing using a singular Gauss-Markov model”, submitted to Journal of Process Control

[2] Q.P. He and J. Wang (2007), “Fault detection using K-nearest neighbor rule for semiconductor manufacturing”, IEEE transaction on semiconductor manufacturing, Vol. 20 (4) 345-354.

[3] B.M. Wise, N.B. Gallagher, S.W. Butler, D.D. White JR and G.G. Barna (1999), “A comparison of principal component analysis, multiway principal component analysis, trilinear decomposition and parallel factor analysis for fault detection in a semiconductor etch process”, J. Chemometrics, Vol. 13: 379-396