Successful Application of Residual Gas Analysis

Attributing to its high chemical selectivity and sensitivity, residual gas analysis (RGA) has a long history of use for chemical process monitoring and fault detection in semiconductor manufacturing. Since the 1990s, IBM's Microelectronics Division (IMD) had extensively deployed RGA systems on the degas chambers of metal deposition tools in its 200 mm fabs for the purpose of protecting the deposition chambers from incoming contamination such as photoresist. It is well known that excessive amounts of photoresist residue on the wafer caused by incomplete photoresist stripping will cause poor adhesion of the metal film on the wafer and contamination of the sputtering target, eventually leading to foreign materials failure of the metal deposition chamber. The benefits of this in situ degas RGA application lie in two aspects: the saving of the metal targets and tool kits, and the reduction in tool downtime. Shortly after the launch of this RGA application, it became the process of record of the degas process of metal deposition tools at IBM's semiconductor manufacturing facilities.

To integrate the RGA system, the PLC-based vacuum system controller of the RGA system was connected to the pressurized air lines of the pneumatic valves of the degas chamber. The pressure switches were used to translate the air actuation of these valves to digital signals for RGA data framing and automating the RGA application. Similarly, the tool inhibit was accomplished through hard wiring between the vacuum system controller and the digital I/O board of the process tool. This integration method has proven to be extremely robust. However, the nature of this tool-level integration prohibited the RGA system from having access to the abundance of fab-level information such as the product type and the process recipe. For fabs with diverse product lines, this would pose a major integration problem.

Fortunately, there are other options available for sensor integration. One method available in 200 mm fabs is to use a secondary SECS port, which supplies the RGA system with process, equipment and logistic information through SECS messaging. This secondary SECS port represents a significant advancement over air switches and hardwired I/O. However, it also faces many issues of its own. For instance, the secondary SECS port does not support multiple applications at the same time, and does not establish a clear path between the RGA application and the manufacturing execution system (MES). To enhance the capability of the RGA system and the value of the RGA data (or any other sensor data for that matter), a new method for RGA integration is required. This new RGA integration method will allow the RGA application to gain access to the tool-level trace/event/recipe data, as well as the factory-level process and wafer information. IBM's new state-of-the-art 300 mm wafer fab in East Fishkill, N.Y., provided an opportunity to develop and implement this new communications technology. Over the past three years, an advanced RGA system was developed that is used for full production control on the degas chambers of the metal PVD tools at IBM's fully automated 300 mm fab.

Figure 1 shows the major components of the computer integrated manufacturing (CIM) architecture and the RGA system implemented in the 300 mm fab. The RGA sensor is attached to the degas chamber of the metal deposition tool, and the RGA server is located remotely in a server room and communicates to the RGA sensor over the high-speed fab TCP/IP Ethernet network. Central to this architecture is the key component: the advanced process control (APC) third-party interface, which is an XML-based messaging system. The interface defines the contents of XML messages and the message named topics used for system communication. The middleware component, IBM WebSphereMQ, is the messaging system, which handles the sharing of XML documents among different factory components.

1. The computer-integrated manufacturing infrastructure of IBM’s 300 mm fab.

The APC third-party interface is designed to enhance the scalability and portability of the APC applications and thus minimize the effort required to integrate multiple APC applications with the SiView MES system. The RGA application communicates by subscribing to certain MQ named topics and publishing-specific XML documents to MQ named topics. Figure 2 demonstrates the sequences of the XML messages used by the RGA application during a process control job. A summary description of these key messages is given below:

2. The sequence of XML messages during the process of a control job.

• APCRunTimeCapability requests a response from the RGA application before a control job is created. This allows the MES system to query what the application's run-time capabilities are (e.g., ProcessDisposition), and also guarantees the RGA application is ready for process monitoring before each process job starts.
• ControlJobInformation and Event and Trace Data provide the RGA application with the critical ControlJob logistics information and the Event and Trace data required for the control algorithm (Table 1 ). The event data publications are created by SiView based on the S6F11 SECS messages and the event data collection report defined in the SiView Data Collection Server. For S6F1 trace data, the MES provides a common trace data collection plan so that a single S2F23 message is created to enable the trace data collection. Note that the wafer ID is provided along with the tool event data, such as a wafer placed event. The wafer ID provides for wafer tracking and the identification of the product type. The significance of this capability will be discussed further.
• ProcessDisposition allows the RGA application to disposition the process tool and notifies manufacturing of an RGA alarm through the MES e-mail and paging system. The process tool can be stopped within one second after the RGA application determines that contamination levels exceed a pre-defined threshold. The resulting RGA alarm e-mail and text page provide manufacturing detailed information regarding the fault, such as the severity of the alarm and the waferID/substrateID of the alarmed wafer. The manufacturing team relies on this detailed information to determine the possible cause of the RGA alarm and how to disposition the wafers or lot. Additionally, an HTTP link to the RGA raw data file is embedded in the text message if an in-depth RGA spectrum study is required.
• The SPC interface establishes a channel for the RGA data to be published to the MES web-based SPC charts. Through the SPC charts, the process and manufacturing team will be able to access the summarized RGA data and compare it with other sensor or test data. More importantly, the SPC link will allow application of various SPC rules to the RGA data if necessary.

With these communication and control capabilities, the APC third-party interface represents a new sensor/APC integration methodology that is far superior to the previously described solutions.

Figure 1 shows another important feature of the RGA system — the adoption of the centralized structured query language (SQL) database system. The database significantly simplifies the work required to administer a large deployment of RGA systems. The database contains the configuring parameters needed for the RGA application. For example, the RGASystemSetting table (Table 2) determines the sampling valves to be used during the degas chamber baseline monitoring and RGA system calibration. The database also stores the summarized RGA data and logistic information about the wafer and process (Table 1 ). By using a centralized database, one can easily analyze several months of RGA data across different tools (time slides) based on various search criteria such as specific product type and mask level using SQL queries. The convenience and importance that the database system brings to data analysis will become more obvious in the next section of case studies.

The importance of APC to semiconductor manufacturing has been widely recognized since the early 1990s. As the semiconductor industry enters the 300 mm era, traditional offline metrology and process control no longer meet the requirements of manufacturing because of the long mean time to detect (MTTD) of process and equipment faults. The significantly increased capital investment and per-wafer cost of 300 mm manufacturing drives the demand of in situ sensor metrology and fault management.

A key trend in 300 mm manufacturing is to automate the manufacturing process to create a virtual "lights out fab," where manual operation and the chance of human error are minimized. It is for this reason that early questions arose about whether fault detection, such as RGA-based photoresist detection, was necessary in a fully automated fab. Today, we can conclude that such controls are indeed necessary. In the past two years, the degas RGA application has far exceeded the return on investment (ROI) expectation. Its ability to quickly detect various contaminates and process errors proved to be critical to the improvement of product yield, process tool availability, and ultimately the profitability of the 300 mm fab. Although photoresist detection is still the major contributor to ROI, the role of the RGA system has expanded far beyond that particular application, which will be discussed further.

There are numerous reasons why photoresist-contaminated wafers can end up in a PVD tool. It is obvious, even in a true "lights out fab," that some of these reasons will not disappear:

• Improper error recovery after failure at the etch process. The recovered wafer with partially stripped photoresist is forwarded to the liner tools.
• Error at the litho process that generates double coating of photoresist.
• Error at the upstream process that causes non-uniform photoresist coating on the wafer. Photoresist remains where the coating is thick.
• Variation of the photoresist stripping process, especially for new product and processes during fab ramp up.
• Human interaction during error recovery, where wafers are mistakenly forwarded to the liner tool without resist stripping.

Different product types have different outgassing behavior. For instance, wafers using low-k interlayer dielectric materials outgas more than oxide wafers. Because of this, it is critical that the RGA application is able to automatically identify the product type in real time and adjust the control algorithm accordingly. In the 200 mm fab, where the RGA application does not have access to the process logistic information (e.g., the product type), only one alarm limit is allowed for all types of wafers and must be set very high to avoid false alarms on wafers that normally outgas more.

The obvious problem with this approach is that many subtle contaminations or process variations will be missed. The answer to this challenge lies in the APC interface specification implemented in IBM's 300 mm fab. The wafer tracking capability provided by the process tool, coupled with the control job information from the MES, allows the RGA application to implement a control algorithm dependent on product type. This enables expanding the capability of the RGA system well beyond catastrophic resist detection. Figures 3 and 4 show two examples of non-photoresist contamination events detected using the RGA, which are detrimental to yield.

On the left of Figure 3 is a plot of the RGA statistical data that was retrieved from the SQL database. Each data point represents the maximum of the contamination index A of each processed wafer. The contamination index A was created to measure the outgassing of the organic species detected during degas, based on the intensities of several pre-determined masses. When the contamination index violates the pre-defined alarm limit, the alarm will be triggered to inhibit the tool. Clearly, for the first 6000 wafers or so, a good baseline of the contamination index A was established, and no excursions were observed. This, however, changed for the remaining wafers in the plot. Certain wafers exceeded the alarm limit (or approached the alarm limit) across multiple tools. The process team pulled the alarmed wafers for examination, and white stains were found on the backside of the wafers, as shown in the picture on the right side of Figure 3 . The RGA spectrum of all the alarmed wafers share the same characteristics. After further spectral analysis, the nature of the white stains was determined, and they were confirmed as the root cause to the RGA alarms. As a result, a modified wafer backside rinse process was implemented to minimize the occurrence of this contamination.

3. The RGA data indicating wafers with high outgassing due to one type of non-photoresist contamination (left), and a picture showing the corresponding contamination on the backside of the wafer (right).

A second non-photoresist contamination example is illustrated in Figure 4 . The plot on the left is the statistical RGA data retrieved from the central database. Contamination index B was created to detect one specific type of contamination that is corrosive and could potentially impact the product yield. The photo on the right side shows residue near the edge of the wafer front side. The RGA data was used to guide a change in the upstream process in an attempt to eliminate the contamination on the wafers.

4. The RGA data indicating wafers with higher outgassing due to a different type of non-photoresist contamination (left), and the picture showing the corresponding contaminations near the edge of the wafer surface (right).

The RGA system is not only useful in protecting tools from incoming contamination; it is also valuable in process integration error detection. Quite often, upstream process variations will be reflected in the outgassing characteristics of the wafer. The RGA system monitoring the degas process is sensitive to such process variations. The RGA system provides early detection of major process integration errors that would likely proceed without being realized and corrected until test results are available days or even weeks later. Because of its inherently short mean time to detect undesirable process variations, many wafers can be saved from being scrapped.

Figure 5 shows such an example, where a sudden increase in low-level RGA alarms across the metal deposition tools is seen. Initial study of the RGA spectrum excluded the possibility of photoresist. No sign of contamination like the two examples previously discussed were observed during the visual inspection of the wafers. To determine the true cause for these alarms, the data was first evaluated by wafer product type and mask level, and then analyzed in detail. As shown in the left plot of Figure 5, all of the alarms belonged to a single product type and one specific mask level. The right plot of Figure 5 indicates no other products or metal layers produced the same alarms during that period of time. This indicated that a change in process was responsible for the sudden change in the RGA alarms.

5. The RGA data indicating an increase in RGA alarms for one product type due to process change to that product (left), and the RGA data indicating no change in the RGA alarms for a different product type during the same timeframe (right).

Using this information as a guide, the process integration team was able to quickly determine that the root cause of the RGA alarms was related to a process change made 10 days prior to the increase of these RGA alarms. In-depth analysis indicated that the process change caused a pressure increase inside the films of the wafer. When the pressure reached a certain point, the film started to peel off and significantly elevated outgassing characteristics were observed for those wafers. Subsequent microscopic inspection of the wafer confirmed this.

The RGA system has matured tremendously over the past few years. The next level of achievement is to become a mainstream application in semiconductor manufacturing. The future of the RGA lies with progress in three areas. The first is fault classification. Most current RGA applications lack the capability to perform automated fault classification. The end user of the RGA application must often rely on the sensor engineer to determine the cause of an alarm in order to disposition the wafer or lot correctly. This demands 24/7 support from the sensor engineer, which is not always practical. It is highly desirable that the RGA application be able to classify the RGA faults based on the cause and supply this information to the end user. With knowledge of the most likely cause of an RGA alarm, the end user can correctly disposition the wafer or lot. One approach currently being considered for fault classification is applying multivariate analysis techniques to the RGA data. Shown in Figure 6 is a preliminary analysis of an RGA alarm caused by one specific type of contamination. It indicates that multivariate analysis can identify the main contributor (mass), which can be used to classify the RGA alarms.

6. Multivariate analysis results showing the main contributor to the RGA alarm was one specific mass.

The second area of future effort is data mining. There have been an increasing number of reports about using data mining techniques to analyze both tool-level and fab-level data (e.g., sensor data, defect data and final test data) in semiconductor manufacturing. Each fab has its own data warehouse containing vast amounts of data regarding every aspect of the wafer process. Data mining is intended to study and discover the correlation, trends and patterns of this data. It will not only correlate sensor data to the yield results and tool performance, but will also reveal hidden yield-impacting factors. This is extremely valuable to help speed yield ramping.

The third area is RGA-based run-to-run control and endpoint control. Most current RGA applications are implemented for fault detection and process diagnosis. Early research has demonstrated the RGA system can also be used for sensor-based run-to-run control and endpoint control to maintain target metrics, such as film thickness and crystal quality, in the presence of process drift and short-term variability. To realize this in a manufacturing environment requires the application to be robust enough for 24/7 operation and adaptive enough for different processes and product types.