Run-to-run control and state estimation in high-mix semiconductor manufacturing

https://doi.org/10.1016/j.arcontrol.2007.07.001Get rights and content

Abstract

In the modeling and control of semiconductor manufacturing, the control engineer must be aware of all influences on the performance of each process. Upstream processes may affect the wafer substrate in a manner that alters performance in downstream operations, and the context within which a process is run may fundamentally change the way the process behaves. Incorporating these influences into a control method ultimately leads to better predictability and improved control performance. Control threads are a way of incorporating these effects into the control of a process by partitioning historical data into groups within which the deterministic sources of variation are uniform. However, if there are many products, which require many threads to be defined, there may be insufficient data to model each thread. This multi-product–multi-tool manufacturing environment (“high-mix”) requires advanced methodologies based on state estimation and recursive least squares. Several such approaches are compared in this paper based on simulation models for a high-mix fab.

Introduction

Semiconductor manufacturing inherently involves a series of batch process steps such as deposition, etch, and lithography. For each batch step there is a recipe that stipulates the operating condition to be employed in that piece of equipment (or tool). Recipe modifications from one run to the next are common in semiconductor manufacturing. Typical examples are modifying the reaction time, feed stoichiometry, or equipment temperature. When such modifications are done at the beginning of a run (rather than during a run), the control strategy is called run-to-run (RtR) control. The goal in semiconductor manufacturing is to control qualities such as film thickness or electrical properties that are difficult, if not impossible, to measure in real time in the process environment. Most semiconductor products must be transferred from the processing chamber to a metrology tool (measuring device) before an accurate measurement of the controlled variable can be taken, because not many in situ sensors are available commercially.

Batch run-to-run control can be viewed as implementing a series of set-point changes to the underlying batch process controllers at the end of each run. By analyzing the results of previous batches, the run-to-run controller adjusts the batch recipe in order to reduce variations in product quality. Thus, run-to-run control is equivalent to controlling a sequence of the controlled variable at times k, k + 1, k + 2, …, analogous to a standard discrete control problem.

Run-to-run control is particularly useful to compensate for processes where the controlled variable drifts over time. For example, in a chemical vapor deposition process the reactor walls may become fouled owing to byproduct deposition. This slow drift in the reactor chamber condition requires occasional changes to the batch recipe in order to ensure that the controlled variables remain on-target. Eventually, the reactor chamber must be cleaned to remove the wall deposits, effectively causing a step disturbance to the process outputs when the inputs are held constant. A RtR controller can compensate for a drifting process, but it can also return the process to target after a step disturbance change (Edgar et al., 2000, Moyne et al., 2001).

As run-to-run control has become more widely used throughout the semiconductor industry, it has become apparent that some of its unique manufacturing characteristics are driving the need for enhanced algorithm development. One such trait is the high-mix of products made in a single factory, such as an application specific integrated circuit (ASIC) fab. Not only might there be a great many different products, but as industry requirements change and technology advances, new products are introduced and old ones are phased out. The mix of products is therefore constantly changing. Economic conditions specific to the semiconductor industry are also a factor, because the capital cost as a fraction of the revenue earned in the semiconductor industry is higher than in other types of manufacturing industries. The high cost of process equipment drives manufacturers to maximize the use of their tools, having as little down or idle time as possible. In order to achieve this goal, it is necessary to use whichever tool is available for processing in a given process step, leaving little room for dedication of tools to specific product process streams. Therefore, one lot of a specific product may take a very different processing path through the fab than the next lot of that same product.

Variations in product quality often are functions of the product being produced as well as the manufacturing tools being used, which is termed manufacturing context. Different products behave differently during processing due to factors such as differences in materials used, configuration or layout of devices and interconnects, feature size, and overall chip size. To further complicate matters, seemingly identical tools may process identical wafers differently based on such conditions as the number of lots processed since the last maintenance event, small differences in tool construction, or minor variations in ambient conditions.

Different methods may be employed to treat these variations. Feedforward control (Seborg, Edgar, & Mellichamp, 2004) measures the incoming state of the lot in order to predict its impact on process performance. A typical example would be to measure the initial film thickness for a CMP process that targets a specified post-polish film thickness. The CMP recipe (e.g. polish time) may be adjusted according to the incoming thickness in order to compensate for variation from the deposition process that created the film. Such a method requires an accurate measurement of the incoming state and a predictive model. In addition, every incoming lot must be pre-measured in order to compensate for such disturbances.

It is not always feasible to use feedforward control. In lithography, because the overlay metrology is a measurement of the overlay of two patterns, i.e., the relative position of one pattern with respect to the other, instead of the absolute position of one pattern, the overlay measurement from previous layers is not sufficient for feedforward control due to the lack of absolute position. In addition, conventional application of feedforward control requires 100% sampling of wafer lots, which is economically unfeasible for most operations.

Sources of variation within facilities that manufacture a large number of products are not adequately characterized in terms of their impact on process control. Different products invariably lead to significant differences in the wafer state for a given tool, due to different materials, topography, pattern density, or other such device properties. These differences between products can often lead to variable process performance, even when employing the same tool and process recipe. A fundamental understanding of the differences between products could permit characterizing their effects on process variability, but inadequate process and product characterization often makes this task infeasible.

Section snippets

Run-to-run control structure

The basic premise of run-to-run control is that there is a deterministic component to the variation of a process. This deterministic component is compensated using available process measurements and knowledge of the overall process behavior during operation. Measurements of key quality metrics are made only after a run has been completed. Thus, the run-to-run controller makes recipe adjustments at the beginning of a run based on data from past runs. The standard controller used in run-to-run

Control thread definition

One method of wafer state estimation is to identify groups of lots that have roughly the same incoming process state. With such a method, the actual state is less important than assuring that each lot within a group has the same state. This enables determination of the best recipe settings for each group based on initial data, and then applying those settings to the rest of the lots within the same group. Each group is segregated from the rest of the groups based upon criteria that determine

Manufacturing application of control threads

Definition of the thread criteria for any controller begins with finding those context variables that determine the state of the controlled process. A large number of variables may contribute to the variation within a process, but only so many variables can be used within the thread definition before data poverty degrades control performance. The goal of the threads is to capture most of the variation within the state, which can be accomplished by using those variables that have the greatest

Alternatives to control threads using estimation methods

In the threaded approach, only lots matching a previously defined set of context criteria are used to update the state in a control model similar to Eq. (1). Because many state-of-the-art fabs are operating with multiple process tools and increasingly diversified product mixes, sharing information among different threads becomes necessary to improve the run-to-run control performance. As shown in Fig. 3, the context criteria are “tool” and “product” with two tools and two products, which result

Summary and conclusions

The control thread methodology allows for modeling of distinct sources of deterministic variation without the need to characterize them individually. In such cases where the process or wafer state can be directly measured, such information can be directly incorporated into the control of a process without the use of control threads. In the absence of the ability to directly measure or identify other sources of variation, however, control threads allow for determination of their cumulative

C.A. Bode is a Senior Member of the Technical Staff with the Manufacturing Systems Technology organization at AMD. He is responsible for implementing and continuously updating AMD's world-class control technology at each of AMD's manufacturing sites. He has written and managed the implementation of many control applications, both at AMD and at manufacturing partner sites. Bode joined AMD in 1998. He holds a bachelor's of science degree in chemical engineering from the University of Illinois at

References (18)

  • T.F. Edgar et al.

    Automatic control in microelectronics manufacturing: Practices, challenges, and possibilities

    Automatica

    (2000)
  • Bode, C. A. (2001). Run-to-run control of overlay and linewidth in semiconductor manufacturing, Ph.D. dissertation,...
  • M.W. Braun et al.

    A comparison of supervisory control algorithms for tool/process disturbance tracking

  • S.K. Firth et al.

    Just-in-time adaptive disturbance estimation for run-to-run control of semiconductor processes

    IEEE Transactions on Semiconductor Manufacturing

    (2006)
  • C.K. Hanish

    Run-to-run state estimation in systems with unobservable states

  • S.S. Jang et al.

    Stability and performance analysis of mixed product run-to-run control

  • H.J. Levinson

    Principles of lithography

    (2001)
  • F.L. Lewis

    Optimal estimation

    (1986)
  • M.L. Miller

    Impact of multi-product and process manufacturing on run-to-run control

There are more references available in the full text version of this article.

Cited by (22)

  • Bridging systems theory and data science: A unifying review of dynamic latent variable analytics and process monitoring

    2020, Annual Reviews in Control
    Citation Excerpt :

    Other competing methods include Bayesian methods and particle filtering in Khatibisepehr, Huang, and Khare (2013) and deep learning analytics for inferential sensors and monitoring in Shang, Yang, Huang, and Lyu (2014) and Zhu, Ma, Benton, Romagnoli, and Zhan (2019). Application reviews can be found in Qin, Cherry, Good, Wang, and Harrison (2006) and Bode, Wang, He, and Edgar (2007) for semiconductor manufacturing, Kano and Nakagawa (2008) for steel industries, Yao and Gao (2009) and Ge, Song, and Gao (2013) for multi-way analytics of batch process data, Chiang, Lu, and Castillo (2017) for chemical processes, and Wang, Tsui, and Miao (2018) for mechanical systems health monitoring and maintenance. The decomposition in (5) gives the partition of the measurement space in a principal subspace and a residual subspace.

  • A construction of knowledge rules for reactive planning of job-mix assignment to homogeneous serial batch machines

    2014, International Journal of Production Economics
    Citation Excerpt :

    The use of their model can simultaneously reduce the number of make-to-stock slab designs and increase the proportion of orders covered by those designs. High-mix production also creates complicated problems for process control (Bode et al., 2007). High product variety has negative effects on production efficiency.

View all citing articles on Scopus

C.A. Bode is a Senior Member of the Technical Staff with the Manufacturing Systems Technology organization at AMD. He is responsible for implementing and continuously updating AMD's world-class control technology at each of AMD's manufacturing sites. He has written and managed the implementation of many control applications, both at AMD and at manufacturing partner sites. Bode joined AMD in 1998. He holds a bachelor's of science degree in chemical engineering from the University of Illinois at Urbana-Champaign. He was awarded an M.S.E. degree (1999) and a Ph.D. (2001) in chemical engineering from the University of Texas. He has presented papers at numerous conferences, and has been published in both control and trade journals. Bode is an inventor on 54 U.S. patents.

J. Wang obtained her B.S. degree in chemical engineering from Tsinghua University, Beijing, China in 1994. She received M.S. and Ph.D. degrees in chemical engineering from the University of Texas at Austin in 2001 and 2004, respectively. From 1994 to 1999 she worked as a research scientist at Biochemical Engineering Laboratory at Tsinghua University. From 2002 to 2006, she was a process development engineer and senior process development engineer at Advanced Micro Devices, Inc. Since 2006 she has been with Auburn University as an assistant professor. Her research interests include systems biology, system identification, semiconductor process modeling and control, fault detection and classification, and control performance monitoring. She has 17 papers published in peer-reviewed journals and holds 10 U.S. patents.

Q.P. He received his B.S. degree in chemical engineering from Tsinghua University, Beijing, China, in 1996 and M.S. and Ph.D. degrees in chemical engineering in 2002 and 2005 from the University of Texas, Austin. He is currently an assistant professor at Tuskegee University. His research interests are in the general areas of process modeling, monitoring, optimization and control, with special interests in the modeling and optimization, fault detection and classification of batch processes such as semiconductor manufacturing and pharmaceutical processes. He is also interested in molecular dynamic simulation and Monte Carlo simulation of micro/nanoelectronic and biological systems. He has had over 2 years of experience in semiconductor industry.

T.F. Edgar received his B.S. degree in chemical engineering from the University of Kansas, Lawrence, KS, and the Ph.D. degree from Princeton University, Princeton, NJ. He is Professor of chemical engineering at the University of Texas, Austin, and holds the George T. and Gladys H. Abell Chair in Engineering. His research interests include process modeling, control, and optimization, with over 200 articles and book chapters. Edgar has co-authored the textbooks Optimization of Chemical Processes (New York: McGraw-Hill, 2001) and Process Dynamics and Control (New York: Wiley, 2004). Dr. Edgar received the AIChE Colburn Award in 1980, AIChE Computing in Chemical Engineering Award in 1995, the American Automatic Control Council Education Award in 1992, and the 2005 IFAC Control Engineering Prize. He was the 1997 President of the AIChE.

View full text