Anthony J. Toprac

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

Run-to-run control is the term used for the application of discrete parts manufacturing control as practiced in the semiconductor industry. This paper presents a new algorithm for use in run-to-run control that has been designed to address some of the challenging issues unique to batch-type manufacturing. Just-in-time adaptive disturbance estimation (JADE) uses recursive weighted least squares parameter estimation to identify the contributions to variation that are dependent upon manufacturing context. The strengths and weaknesses of the JADE algorithm are demonstrated in a series of test cases developed to separate the various disturbances and processing issues a control system would be expected to encounter

Plasma etching endpoint detection using multiple wavelengths for small open-area wafers

This article proposes a new approach for etch endpoint detection of small open area wafers. The traditional endpoint detection technique uses a few manually selected wavelengths, which are adequate for large open areas. As the integrated circuit devices continue to shrink in geometry and increase in device density, detecting the endpoint for small open areas presents a serious challenge to process engineers. In this work, a high-resolution optical emission spectroscopy(OES) system is used to provide the necessary sensitivity for detecting subtle endpoint signals. Principal component analysis is used to analyze the OES data and extract key components that capture the endpoint signal. Data analysis from many wafers shows that the endpoint pattern in the principal components is repeatable. Two methods are used to select wavelengths so as to improve reliability and reduce susceptibility to noise. The first method is to remove those spectrum windows that contain no endpoint information. The second method is to use a “sphere” criterion to select the most relevant wavelengths. The final endpoint algorithm using a much-reduced number of wavelengths shows more distinguishable and reliable endpoint features.

A comparison of run-to-run control algorithms

Run-to-run control is the term used for the application of batch process control as practiced in the semiconductor industry. This paper gives a brief introduction to the fundamental parts of a run-to-run control algorithm and surveys several of the popular techniques for the application of this control methodology. In particular, model predictive control is explored as a robust and flexible technique for implementing run-to-run control.

AMD's advanced process control of poly-gate critical dimension

Formation of the MOS-FET polysilicon gate structure is a critical step in integrated circuit manufacturing. Control of poly-gate Critical Dimension (CDs) greatly affects revenue from microprocessor production. Poly-gate CDs correlate strongly to speed. As a result, variation in CD control causes unsaleable slow parts, high revenue fast parts, or scrapped high leakage product from overly fast parts. Controlling to the optimal value CD value, however, it is a difficult task due to the continual drift and step changes that occur in the photolithography and etch tools. As a result of this need, AMDs Fab 25 developed an automated run-to-run controller of poly-gate CDs as part of an Advanced Process Control (APC) initiative. From the perspective of both control and manufacturability, Fab 25s Run-to-Run controller of poly-gate Critical Dimension (CD) has been a critical enabler of our success in manufacturing the K6 product. This paper discusses the architecture, algorithm and results of the poly-gate CD control system.

A Predictive Model for the Chemical Vapor Deposition of Polysilicon in a Cold Wall, Rapid Thermal System

The chemical vapor deposition of polysilicon from thermally activated silane in a cold wall, single-wafer rapid thermal system was studied by experimentation at a variety of low pressure conditions, including very high temperatures. The effect of diluent gas on polysilicon deposition rates was examined using hydrogen, helium, and krypton. A mass-transfer model for the chemical vapor deposition of polysilicon in a cold wall, rapid thermal system was developed. This model was used to produce an empirical rate expression for silicon deposition from silane by regressing kinetic parameters to fit experimental data. The resulting model provided accurate predictions over widely varying conditions in the experimental data.

Lithography overlay controller formulation

Lithography overlay refers to the measurement of the alignment of successive patterns within the manufacture of semiconductor devices. Control of overlay has become of great importance in semiconductor manufacturing, as the tolerance for overlay error is continually shrinking in order to manufacture next-generation semiconductor products. Run-to-run control has become an attractive solution to many control problems within the industry, including overlay. The term run-to-run control refers to any automated procedure whereby recipe settings are updated between successive process runs in order to keep the process under control. The following discussion will present the formulation of such a controller by examining control of overlay. A brief introduction of overlay will be given, highlighting the control challenge overlay presents. A data management methodology that groups like processes together in order to improve controllability, referred to as control threads, will then be presented. Finally, a discussion of linear model predictive control will show its utility in feedback run-to-run control.

Adaptive control of multiple product processes

It is a common practice in todays microelectronics manufacturing facilities to have many different products and processes run on each processing tool. This is caused mainly by the high capital costs associated with the tools and the limited capacity of the facility. A run-to-run controller relies on having a model that is consistent from run to run. When the different processes run on the tool are significantly different, the controller may behave unexpectedly because each change to a new process can appear as a large disturbance. In addition, it may take several successive runs of a given process for the controller to stabilize, but this cannot happen if the processes change too often. Ideally, the controller should be able to determine optimal settings for all processes that must run on the tool, regardless of the order in which they appear. In an adaptive control strategy, an online system identification scheme runs along with the controller and constantly adjusts the model so that it mimics the true behavior of the system. One very difficult task in this situation is determining whether observed errors in the output are due to errors in accounting for tool differences or for product differences. This discussion will outline a scheme for deciding which model parameters are in error and performing the correct model updates.

Modeling of Gas‐Phase Chemistry in the Chemical Vapor Deposition of Polysilicon in a Cold Wall System

The relative contribution of gas-phase chemistry to deposition processes is an important issue both from the standpoint of operation and modeling of these processes. In polysilicon deposition from thermally activated silane in a cold wall rapid thermal chemical vapor deposition (RTCVD) system, the relative contribution of gas-phase chemistry to the overall deposition rate was examined by a mass-balance model. Evaluating the process at conditions examined experimentally, the model indicated that gas-phase reactions may be neglected to good accuracy in predicting polysilicon deposition rate. The model also provided estimates of the level of gas-phase generated SiH[sub 2] associated with deposition on the cold-process chamber walls.

The Effect of Initial State Estimates on Just-in-Time Adaptive Disturbanc Estimation

Run-to-Run control algorithms for high-mix semiconductor processes typically require that the initial product state estimates have sufficient accuracy for satisfactory control. In this paper, we use historical process data and apply single observation just-in-time adaptive disturbance estimation (JADE) to find the initial product state estimates. Single observation JADE with random selection, high-frequency sampling, and exclusion of the earliest data from the average is shown to provide satisfactory initial product state estimates. The effect of initial state estimate accuracy is demonstrated by several simulation and industrial data examples. We also provide a method to estimate relative confidence between individual product state estimates, information that may be used to determine assignment of process error between the tool and product state.

Model-based control of chemical mechanical polishing

Chemical mechanical polishing (CMP) of silicon oxide interlayer dielectric is a critical process in modern multi- layer metal integrated circuit manufacturing. In this process, the rate of planarization of features on a silicon wafer surface changes with age of the polishing pad. This effect creates the need for adjustment of polishing times to compensate for changes in planarization rates. The way that planarization rate varies with polish time must be defined to develop robust control of this process. In this work, a theoretical model for the dependence of planarization rate on polish time was developed. This model was then applied to data from a Westech 472 CMP system and shown to accurately capture the time variation of measured removal rates. A control algorithm using this model was tried using a different CMP tool, the Westech 372, creating a mismatch between the control model and process. Nonetheless, the control model quickly adapted to the new conditions and controlled the process well.