Papers

AVS1996 Session MS-MoM: Sensor-Based Fault Detection and Process Control

Monday, October 14, 1996 8:20 AM in Room 201A

Monday Morning

Start	Invited?	Item
8:20 AM		MS-MoM-1 The Application of Advanced Process Control to Semiconductor Manufacturing in Today's Factories J. Gragg (Motorola, Inc.) Research and development programs such as the Microelectronics Manufacturing Science and Technology (MMST) project have clearly demonstrated that there are many benefits to be realized from the application of advanced process control to semiconductor manufacturing. There are a number of barriers, however, that have impeded the widespread adoption of these practices in the semiconductor industry. Among these are (1) the lack of suitable sensors, (2) the inability to easily modify the control strategies used in existing equipment, and (3) the difficulty of integrating advanced process control hardware and software with existing systems. While solutions to these problems are under development it will probably take several generations of equipment before these solutions become widely available. In the meantime the challenge to the semiconductor industry is to take advantage of these advanced process control methods to improve our manufacturing capabilities within the limitations of our existing manufacturing equipment and systems. This presentation will discuss advanced process control methods that can be applied in today's factories using existing technology, the manufacturing processes to which these methods can be applied, and the benefits resulting from the application of these methods to these processes.
9:00 AM		MS-MoM-3 Neural Network - Plasma Process Implementations in Production E. Rietman (AT&T Bell Laboratories) The talk will be about two plasma processes in production for CMOS chips that are controlled by neural networks. The primary focus will be on implementation issues: hardware modification of the plasma reactor, and software communication with the plasma reactor and a host computer. The neural networks run on a host computer and receive live process signatures. These real-time data are used in computing the etch time for the current wafer, in process and enabling wafer-to-wafer control. In some cases the neural network receives live and archive data in making this computation. The neural network, trained with production data, essentially maps the process signatures to etch time. On-line training considers the desired film thickness for the product being etched. Results for a gate etch process improved the standard deviation of the film thickness by 40%. Early results for via etching indicate an 8% improvement in the standard deviation and an 89% improvement in mean value for target film thickness.
9:20 AM		MS-MoM-4 FDC/MBPC, Current and Near Term Possibilities in SC Manufacturing G. Barna (Texas Instruments) Advanced Process Control (APC), which includes both Fault Detection and Classification (FDC) and Model Based Process Control (MBPC) has been utilized within various TI Wafer Fabs for several years. FDC has been implemented on a variety of tools, specifically looking at process output - such as an endpoint trace from a plasma etcher - anomalies relative to a standard reference signal. MBPC has been utilized for the control of film thickness in both thermal and sputtering reactors. Both these methods have provided a significant impact in Wafer Fab output. Looking at these methods from the viewpoint of a time-line, the near-term possibilities are becoming clearer. FDC will move from the existing univariate signal analysis to the analysis of multiple signals. MBPC will move from the control of a single wafer-state observable to the control of multiple observables, some without direct sensors. Both of these paths depend on an APC architecture to be in place, to acquire, analyze and summarize the information to be used for these more complex APC methods. This presentation will provide a vision for these paths, a definition of the APC architecture required to move along these paths, and the roadblocks that are slowing down the journey.
10:00 AM		MS-MoM-6 Advanced Process Control in Plasma Processing A. Voshchenkov (Lam Research Corporation) By 1999, 0.25\micron\m silicon technology will be in production and the transition from 200mm to 300mm will have begun. This will drive the need for reduction of variances in critical plasma etch characteristics such as critical dimensions, uniformity, selectivity, profile control, and device damage. The traditional methodology of improving performance through tighter specifications on subsystems may be approaching the point of diminishing returns. Present control strategies are based on open loop control of an assumed static process, i.e. no time or processing history dependence. Adaptive process control (APC) will enable automatic compensation of long term process drifts, component degradation, and incoming wafer state variations. Customer acceptance is crucial but no coherent definition of APC has evolved as yet. APC focusses on maintaining stable process results rather than fixed process set-points. An APC sensorized plasma etcher could provide better critical feature size reproducibility, lower defect density, monitoring of chamber effects, less operator intervention, and greater tool availability. Initially manufacturing may not utilize closed-loop control but restrict APC to monitoring the process and outgoing wafer state. On the journey towards APC such model-based methods as automated tool fault detection and fault classification will be developed and accepted in manufacturing first. Lam Research Corpation is evaluating APC capabilities such as model-based sensor integration (SEMATECH J-88), sensor bus architecture, and distributed intelligent subsystems.
10:40 AM		MS-MoM-8 A Production Evaluation of Real-Time Statistical Process Control on a Sub-0.5\micron\m TCP Metal Etch Process E. Boskin (Perceptive Technologies, Inc.); T. Dalton (Digital Semiconductor) The Real-Time Statistical Process Control\super 1\ (RTSPC) algorithm analyzes real-time sensor and actuator data for the detection of equipment faults.A statistical RTSPC model was constructed using data from 10 wafer lots at M2 etch on a TCP metal etch system. The statistical filters in RTSPC detected faults in the data during the model generation process. These faults were caused by a bimodal distribution in several sensor signals. In this case, The RTSPC algorithm did not simply generate a high variance model to describe the data. Rather, RTSPC clearly highlighted the problem.During this study, the reflected RF bias power for the etch tool became too high, resulting in the replacement of the lower RF matching network. Data collected following this indicated that the reflected power decreased to an acceptable level, and that the bimodal behavior disappeared. This example shows that deviations in the time series behavior of the real-time signals can be used to predict equipment failures and required maintenance.The critical dimension (CD) difference (etch - photo) was analyzed for this tool to ascertain the impact of the faulty component on tool results. Real-time sensor and CD data from a second identical etcher was also analyzed; this etcher never displayed the bimodal signal fault. Further, the data showed that the CD delta was 50% larger for the faulty etcher than the second etcher. Finally, after the match was replaced, the two tools produced identical CD deltas during the etch process.\super 1\Lee, Boskin, Liu, Wen, Spanos, "RTSPC: A Software Utility for Real-Time SPC and Tool Data Analysis," IEEE Trans. on Semi. Mfg., 8(1),17,1995.
11:00 AM		MS-MoM-9 Advanced Process Control - How Outside Expertise Can Help D. Spain (Integrated Systems, Inc.) The dramatic improvements in the capabilities of semiconductor fabrication equipment over the past two decades have come not from improved control but from ever more heroic engineering. Arguably, improved control is now where gains in equipment performance can be made most readily, but the semiconductor industry has almost no control engineers. The implementation of control software has traditionally been left to software engineers, who have no formal control training. However, for decades, the government has strongly supported the development of control technology for the aerospace industry, and this expertise is available in the form of third party control companies. Two areas where this expertise has been put to good use are model-based control and embedded control.A model-based approach, where the controller is designed against a mathematical model of the equipment, provides a systematic control design that extracts the maximum performance from semiconductor fabrication processes, where interactions between the various inputs and outputs limits the performance of traditional single-input, single-output control designs.Embedded control, where the calculations required for critical subsystems are performed in local controllers that only need process setpoints from the central computer, has many performance and cost advantages over a totally centralized approach. In a sense, an embedded controller is simply an intelligent "device" that concentrates data from several sensors and uses the available actuators to make an equipment subsystem perform as requested by the central computer. The important difference is that an embedded controller is not a generic device, like a mass flow controller; it is custom tailored to the particular application and equipment.