Practical, Real-Time Multivariate FDC

The wide-spread implementation of automated equipment control/advanced process control (AEC/APC) in existing fab environments faces significant obstacles. Even though large amounts of unsynchronized sensor and tool data for process and wafer states are available, this information is usually present as a data "cacophony" that is not well-suited for use in APC protocols. Within a typical fab, the bandwidths and polling rates for these data may vary so widely that the synchronization of data streams becomes extremely difficult. Equipment and metrology data must be synchronized to wafer lots and/or individual wafers to be effectively used in APC.

Even though recent advances in tool integration^1-4 can impose synchronicity on the different data streams within a fab, coherent data sources are insufficient for the effective implementation of APC. The complexity of advanced semiconductor manufacturing processes, when coupled with the many and significant correlations between the variables within and across processes, means that traditional univariate statistical process control (USPC) misses many of the control parameters that are needed for APC.^5,6 More powerful protocols, such as off-line multivariate analysis (MVA) combined with real-time fault detection and classification (FDC), must be used in concert with USPC for effective APC. While not a substitute for SPC limits and alarms, MVA/FDC generates useful information from a relatively small sampling of sensor and tool data. The implementation of MVA/FDC requires a system-level approach to data collection, analysis and process control.

MVA algorithms must have access to synchronized tool, process and wafer-state sensor data within a networked and web-enabled open architecture to effectively implement fab automation using real-time FDC. Such a framework enables the high-speed and real-time information transfer that is a key prerequisite for successful APC.

The TOOLweb integrated suite of products provides a synchronized data stream to an associated suite of off-line MVA and on-line FDC tools. The framework is compliant with all SEMI standards, has sampling capabilities in excess of 100 Hz, and allows complex data to be reduced and published as vectors of values. The primary components of this framework are a data gateway and monitor (SenseLink), a data collection and tool interface (the Blue Box module), and the database and APC host (the TOOLweb server). It is important to note that, in addition to accepting data from TOOLweb-enabled sensors, the system can also access data from many different manufacturers' sensors. Non-compliant sensors are integrated into the framework via the SenseLink module either by use of standard interfaces (such as analog, digital and DeviceNet) or by custom protocol translations for third-party sensors that don't support standard protocols. This aspect of the system is of critical importance for implementation of APC in environments heavily invested in legacy tools and sensors. This framework thus enables the high-speed, multi-user connectivity and data sharing that is a prerequisite for e-diagnostic, APC and AEC purposes.

The Blue Box data interface and multiplexer (Fig. 1 ), trading on this high degree of interconnectivity, synchronizes the sensor and tool recipes with the incoming data streams and equipment-specific semiconductor equipment communications standards from the tool or metrology equipment. It then passes the combined data stream on to the appropriate factory host, APC and FDC tools. Appropriate selections of external sensors and SenseLink modules are integrated to a tool via the Blue Box by using Ethernet and XML-based protocols. Tool and sensor data are collected, and multiple collection plans, each with a different data rate, can run simultaneously. Relevant data is sent to the host in a combined output stream.

1. The Blue Box data interface and multiplexer shows the associated communications protocols and functionalities.

MVA extracts scores and residuals from the extended data set. Off-line MVA is used to create models used in the real-time fault detection engine. The method works well with collinear and missing data. The technique builds both lower-level and whole-batch models, and the lower-level model shows the dynamic behavior of the process. Comparable systems typically do not deal well with collinear or missing data, do not reflect the dynamic nature of the process, and do not reveal buildup trends within the data.

The MVA has a flexible modeling architecture and simple interface that permits very rapid model building. Using "good wafer" data sets, the MVA algorithm builds models of the in-control process using principal component analysis (PCA) and partial least squares (PLS) techniques. PCA and PLS effectively compress large sets of highly correlated data and display the results in relatively simple graphical format via the use of linear combinations of the original multivariate data. PCA generates its models based on variation in the independent variables scores and significantly reduces dimensionality by summarizing the data.^7,8 PLS extends the PCA models to include response variables such as film properties, yields, etc.^9,10 This approach silently and automatically builds models and, by using "local centering," can be used for reliable R2R adaptation of setpoints. New model triggers might rely on fixed variables, such as the number of wafers since the previous model was built, or on process or film variables, such as etch rate or refractive index.

The FDC server accepts these models, and the online fault detection engine computes statistics based on the current model in real time. An overview of the data can be obtained from t1/t2 score scatter plots (Fig. 2a). These plots display outliers and/or data clusters that may indicate "bad" wafers and process characteristics. Although the plots are completely statistical and cannot identify the reasons for the observed clustering and outliers, they do identify areas requiring further investigation. Within the analysis suite, Hotelling's T2 statistic (Fig. 2b) is used to summarize the fit of wafers to the model relative to an alarm limit. Wafers outside the limit are outliers whose measured parameters do not fit the currently accepted model for "good" wafers. The DModX (distance from model plane) statistic (Fig. 2c ) analyzes correlation structure changes and reveals wafer data too far away from the model plane and exceeding a critical limit. Such indicators can reveal abnormal process behaviors not previously seen. Contribution plots that track and summarize the wafer evolution for each process parameter or variable identifier reveal the root cause of the bad wafer(s). The software presents only those variables or VIDs most likely to be root causes. Once the root causes of the fault have been identified, final drill down and defect classification become possible.

2. FDC statistics plots of a t1/t2 score scatter plot that shows data clustering and the presence of an outlier (a); of Hotelling’s T2 statistic that shows the presence of bad wafers outside of critical limits (b); and a DModX plot that shows the presence of an abnormal process (c).

The TOOLweb server collects data from multiple Blue Boxes in parallel and performs real-time MVA for fault detection. It also provides for a fab status overview, long-term (92 days standard) data storage and remote notification on exception. Besides fault detection, other typical applications include fingerprinting, chamber matching and process optimization.

Case 1: Hidden Problems in Large Data Sets

3. Contribution plots for the outlier in Case Study #1, which revealed the variable that caused the process deviation.

The t1/t2 score scatter plot, Hotelling's T2 and DModX plots for this case are those shown in Figure 2. The case involved an etch process that produced non-uniform etch depth and poor-quality wafers (two separate problems). To resolve the problem, the off-line MVA mined the large process and tool data set to isolate the root causes of the process variance. The presence of clusters in the t1/t2 score scatter plots obtained from the data (Fig. 2a) clearly identified two process states or "clusters," as well as an outlier or a wafer that lies significantly outside of standard tool operation. Drilling down through the data indicated that the non-uniformity issue was related to process chamber temperature, and the problem was corrected fairly easily. The Hotelling's T2 and DModX plots identified the poor-quality wafers (Figs. 2b and c). Changes to the setpoints of the identified variable easily resolved the poor wafer quality problem. The outlier contribution plots (Fig. 3 ) revealed the variable that was the cause of the process deviation (argon flow).

Case 2: The Importance of External Sensors

Within a process line, the initial few wafers of every lot that were inserted into the loadlock of a particular deposition tool turned out to be rejected because of poor film quality. The previous step in the process line was a wet clean. The process chamber was equipped with an external RGA sensor. This external RGA sensor was linked into the APC control network via the TOOLweb suite. The Blue Box collected the RGA signals as extended-variable IDs (EVIDs) and synchronized this data with the process tool single-variable IDs (SVIDs). Figure 4a shows the RGA trace for argon in the process chamber vs. the wafer number. It is clearly seen that regular excursions occur in the argon concentration in the gas phase of the deposition tool. Correlation of these data with the recipe for the process tool confirmed a strong relationship between the argon concentration excursions and the initial few wafers into the loadlock. Further analysis of the EVIDs in the system revealed the data shown in Figure 4b. Here, clear RGA evidence for excursions in the concentration of water within the process chamber is evident. Furthermore, the excursions in argon concentration evident in Figure 4a are directly correlated with the excursions in the concentration of water within the process chamber shown in Figure 4b .

4. The RGA sensor data for process chamber argon concentration traces vs. wafer number (a) and process chamber water concentration traces vs. wafer number (b).

The presence of relatively high concentrations of argon in the process chamber was traced to the only available source of the gas, the argon flow to the backside of the wafer. Its presence in the process chamber clearly indicated poor sealing between the wafer and the hot electrostatic chuck for those first few wafers from the loadlock. This leak also carried air and moisture into the process chamber, and these contaminants were determined to be the root cause of the poor film quality. The presence of the integrated RGA sensor and its interconnectivity within the TOOLweb framework permitted the rapid determination of these correlations and showed that wet wafers were preventing good chuck vacuum. The RGA traces indicated that, after a while, the wafers tended to dry out and the leakage stopped. With this data, the problem was rapidly solved and the line quickly returned to in-spec wafer production.

Case 3. Fast Chamber Matching

Within a particular fab, two supposedly identical process chambers yielded distinctly different qualities in the wafers processed through them. The differences were not necessarily outside the specification range of the process, but such systematic differences in quality are nevertheless undesirable.

A t1/t2 score scatter plot of the data taken from the two chambers is shown in Figure 5a. The differences between the two chambers (D and E) are clearly delineated by the data clusters evident in the plot. Figure 5b shows how consistently over time the controlled variables within chambers D and E clustered in two distinct groups while remaining within the formal control limits. Further analysis (Fig. 5c ) quickly and successfully identified the VIDs that contributed to the differences between the two chambers. Once these VIDs were determined, APC feed-forward/feedback control was used to bring the chambers back into matched performance.

5. Statistics plots for Case Study #3 show the scatter plot clustering of the t1/t2 scores (a); the process variables plotted over time of separate clusters for the two chambers (b); and a plot of the contributions of individual variable IDs (c).

The coupling of improvements in fab process hardware sensors, sensor connectivity and communications with new approaches to FDC can overcome the current obstacles to implementing APC and AEC in modern production environments. Many different types of sensors can now be integrated into comprehensive fabwide FDC solutions. The use of open architecture and scalable design allows for practical fab integration without the need to change internal IT infrastructures.

The system described in this report enables high-speed, multi-user connectivity and data sharing for e-diagnostics, APC and AEC. The system's easy access to wafer data facilitates tool fingerprinting, tool optimization and chamber matching by converting large data tables into a few simple MVA plots. The online multivariate FDC and off-line MVA eliminate the false alarms that can be generated using current USPC methods.