Extended Abstract
of Presentation at the 1997 AVEM Fall Seminar

Introduction
During this decade the rapid pace of silicon technology will drive critical dimensions to 0.18 µm and to the introduction of 300 mm wafer processing. The challenges of ever increasing complexity must be met with cost effective solutions. Process parameters will require even tighter tolerances due to the reduced size of transistors, small metal interconnect pitch, and increasing aspect ratio of microstructures. This requires improvements in plasma processing capabilities that can be more efficiently achieved by semiconductor processing tools utilizing advanced Fault Detection and Fault Classification (FD/C), and Advanced or Adaptive Process Control (APC) methods.

Technology Drivers
As we enter the 0.25 µm era, silicon technology drives the need for reduction of variances in critical characteristics. For plasma etching of devices, the critical characteristics include critical dimension (CD) control, uniformity, selectivity, profile control, and reduction in device damage (Figure 1). In addition, manufacturability is as significant a driver as is technology. Manufacturability issues include yield, cost of ownership, productivity, contamination control, utilization, and processing tool reliability (Figure 1).

While the device features are getting smaller and the wafers bigger (200 mm to 300 mm transition), the risks are getting greater. The traditional methodology of improving performance through tighter specifications on components and subsystems may be approaching the point of diminishing returns. To stay on the 30% manufacturing cost learning curve, a paradigm shift from fixed process recipe optimization to adaptive process control will evolve with time.

In the current methodology, the presence of random variances drives the development of processes with a wide process window which are less sensitive to uncontrolled variations in the process parameters. This implies that the process results are primarily dependent upon the tool state rather than the process environment. A fixed recipe strategy precludes the ability to compensate for systematic drift and changes. It also precludes the ability to utilize feed-forward information about the incoming wafer state and variations in the material.

Process Equipment Self-Evaluation Techniques
On the way to achieving adaptive or some form of advanced process control, automated equipment self-evaluation techniques such as Fault Detection and Fault Classification are required enablers. Current SPC methods are based on single variables and static models. Such methods are very sensitive to individual sensor drift and errors, and, therefore, must use wide alarm limits to avoid frequent false alarms, thereby lacking sensitivity. Also, prior to full APC implementation, model-based control of subsystems will occur.

Advanced fault detection methods include Real Time Statistical Process Control (RTSPC), a multivariate methodology utilizing a static tool model that has been implemented successfully at Lam Research Corporation (Figure 2). RTSPC allows a wafer-to-wafer processing comparison of in-situ sensor data to a statistically derived tool model for a particular product line. Control and alarm limits can be set at several different time scales and utilize the relationships between multiple signals/parameters (covariance) to achieve higher sensitivity without an increase in false alarms.

Since fault detection methods are based on comparing sensor signals to a modeled state, sensor data must be reliable. Also, some form of local fault detection and self-calibration is needed in "smart" sensors. Fault Detection can be implemented without Fault Classification but the advantages are somewhat limited. The knowledge that something is wrong without indicating what needs to be fixed is insufficient. Fault Classification (FC), or automated diagnosis, is the logical extension of Fault Detection. By the rapid and accurate identification of the "out of normal operating range" components or subsystems, FC minimizes system down time (MTTR) and related costs. It is synergistic with FD and APC methods. A capable Fault Classification system should distinguish between hard faults that require repair or maintenance and those that can be compensated for by a process parameter modification.

Intelligent Sensors, Subsystems, and Processing Tools
An intelligent plasma etch system, for example, consists of "smart" process modules and subsystems with multiple, data rich sensors, a framework for distributed data collection/reduction, data storage and algorithmic analysis, and model-based control. The supporting subsystems need to contain the appropriate set of sensors and actuators which are required to execute, monitor, and control specific system functions, e.g. vacuum/pressure/gas flow, RF power delivery, wafer transport, etc. Intelligent tool design will integrate sensors, data acquisition and storage, with the data reduction and algorithm digital processing capability necessary to support FD, FC, and APC. Local processing of the raw data, e.g. extraction of mean and variance, is required at the sensor or subsystem to convert data into useful information. DSP functions should support flexible, programmable data reduction algorithms. Inclusion of a "standard" digital sensor bus communications network is essential for smart tools.

Development Economics
The benefits of new in-situ sensors and smart tools are primarily operational in nature, i.e. reduction in scrap/higher yields, reduced set-up times and MTTR. The cost-of-ownership impact of these improvements can only be measured by the end user, the chip manufacturer. It is difficult for the tool supplier to determine the "quantitative" value of a new sensor or smart tool capability without participation and support of the chip manufacturer. The value of a particular sensor lies not in the data by itself but in the decisions and actions that are made based upon accurate and timely data. The present paradigm for the development and introduction of new sensors is extremely inefficient. Sensor vendors, etch tool suppliers, and the chip manufacturers would benefit from rapid development and introduction of intelligent processing tools. Obstacles in this food chain arise from timing issues, distribution of costs, and differences in expectations of performance.

Conclusions
While the device features are getting smaller and the wafers bigger (200 mm to 300 mm transition), the risks are getting greater. The traditional methodology of improving performance through tighter specifications on components and subsystems may be approaching the point of diminishing returns as well as driving cost upward. To stay on the 30% manufacturing cost learning curve, a paradigm shift from fixed process recipe optimization to smart tools and adaptive process control will evolve with time.

The availability of "smart", well characterized components and subsystems, is essential to meet time-to-market, future performance, reliability and cost-of-ownership targets. Technologies such as Fault Detection, Fault Classification, and Advanced or Adaptive Process Control are effective and compatible within the framework of intelligent processing systems. The benefits of new in-situ sensors and smart tools are primarily operational in nature, i.e. reduction in scrap/higher yields, reduced set-up times and MTTR, faster start-up and qualification of tools in a fab, etc.

The mix of knowledge and skills required to effectively develop, integrate, qualify, and support smart sensors and process tools, dictates that the sensor vendor, process tool supplier, and the chip manufacturer work as an effective team. Effective teamwork spanning multiple organizations and industry groups will be one of the greater challenges. Rather than the "normal" or often typical competitive supplier-vendor relationship, the efficient and timely introduction of intelligent systems will require a strong partnering, cooperative and interdependent relationship.

Back to Spectrum the AVEM Newsletter