Improving Fab Engineering Efficiency With Autonomous Data Analytics

During my earlier career as a process integration engineer, one of my primary responsibilities was to find yield enhancement opportunities by investigating underlying relationships between bin failures and process parameters within the fab. While performing this analysis, there were many impediments to identifying relationships among different data types: sort maps, electrical test maps, parametric data, in-line metrology and defect scans, and more. Discovering correlations and regression results from scatter plots was a primary source of finding these relationships, due to both the relative simplicity of those operations and the ready availability of software widely adopted across the semiconductor industry for performing those calculations.

Unfortunately, adding map data into this mix proved to be a major complicating factor. How could I incorporate this spatial element when much of the fab data does not include it? Inline metrology sampling, electrical test, and bin sort are usually measured in different areas of the wafer. Is it possible to overlay these maps and find meaningful correlations among these critical parameters? Furthermore, who has the time to manually overlay these maps? Can this analysis be done in the background and the results delivered in a standardized format? Fab engineers need tools that can deliver on these needs by discovering correlations in the background among differing data sources, then summarizing and delivering the results.

Time efficiency of analysis is especially critical when investigating quality excursions. During this time, the line may be stopped for the affected products, causing a massive impact on tool utilization and fab shipments. The line can only be restarted after a thorough investigation into:

• Pinpointing and mitigating the root cause of the excursion
• Containing the affected wafers
• Assessing the yield and quality impact of affected wafers
• Creating a disposition plan

A major inefficiency in this process is pulling together data from many diverse sources for analysis. Sort, electrical test, defect, and fab tracking data may be stored in various databases, requiring an engineer to query these various sources separately just to combine the data for analysis before performing the analysis itself. Furthermore, when performing analysis, it is important to understand any intrinsic relationships that exist in the affected product to prioritize some root cause candidates at the outset. Some of this information may be available from product startup, but many foundries do not receive this information. Therefore, it is left to fab device and yield engineers to determine which factors are important as part of the analysis. An effective alternative to this process would perform these functions automatically in the background and report the results to the fab device and yield engineers by:

• Performing commonality analysis to pinpoint the root cause
• Identifying the common signal or signature of the affected wafers
• Correlating the common signal with downstream data to assess the impact
• Reporting the results to the parties responsible so they can concentrate on resolving the problem.

The Decision Support System (DSS) developed by Synopsys delivers on these needs. Using the Synopsys Fab.da data infrastructure, many different data sources (even if they are from multiple applications and suppliers) are efficiently stored and managed in a data lakehouse, allowing DSS to access the data from one application. Using the accumulated data in the lakehouse, DSS analyzes the data in the background at regular intervals, up to a daily basis. These autonomous analyses identify latent relationships among all the stored data types, including:

• Grouping wafers into distinct behavior models for map and trace data
• Commonality (shared wafer lists between data behaviors)
• Regression analysis
• Map similarity (shared area of map behaviors)

Fig. 1: Many data sources are integrated into a data lakehouse.

The results from DSS are delivered to the user as a ranked list, prioritizing higher-correlating, more recent, or particularly distinctive data behaviors that could be of interest to the user based on their desired search terms. Outside the semiconductor industry, a similar approach is adopted by websites such as Priceline.com, where users can easily find the lowest price among different travel combinations of airfare, lodging, and transportation. Similarly, a DSS user can find the models of map and/or trace data that best describe the search conditions.

Among DSS users, particular emphasis has been placed on map similarity to identify relationships between sort, electrical test, and parametric test maps. In a foundry environment, where yield, device, or integration engineers may be siloed from product development, relationships between sort and parametric test parameters may not be documented at initial product ramp. Much of the knowledge of individual device relationships comes from recording correlations that have been identified over the course of a product’s lifetime from excursions, yield enhancement projects, or device targeting. The value of DSS for these users lies in the ability to use the data collected during every minute of production to correlate the map patterns (such as sort bin maps) and create a list of other map patterns (of electrical test parameters) that have significant overlapping areas of the maps.

This kind of analysis involves handling extremely large data volumes. In the case shown in Figure 2, a bullseye bin map signature on one product (VD2420A) was compared with other products to determine if the same signature is found. One other product, VD2421A, was found to have a similar signature with 86% area similarity. Each map cluster contains more than 2,200 wafers, and with DSS having already completed this similarity analysis in the background, it was able to display the results in seconds. DSS users can add these results to a catalog of knowledge about which end-of-line test parameters are interrelated, reducing the amount of manual analysis work that would be required during an investigation.

Fig. 2: DSS identified two distinct products with similar bin signatures.

DSS performs similar behavior analysis of trace data. In one real-world case, a specific chamber was experiencing a gas flow issue that caused a shift in a sheet resistance parameter at the end of the line. Slow ramp of the gas flow affected multiple recipes and products, displaying a similar behavior in the trace data for the flow readback from the mass flow controller. By identifying this same trace data behavior across multiple recipes, the user was able to identify all the ~1600 wafers affected by this issue across four products and recipes purely from the trace data, without having to wait for measurement or end of line data to confirm which wafers were associated with this quality issue.

Fig. 3: Three distinct trace behaviors detected by DSS for gas flow readback. Behavior 3, encompassing 1.1% of all runs, has a slow ramp to setpoint compared to the other behaviors.

DSS can also be used to notify users if there are new behaviors occurring in the data. The Subscription function allows users to receive notifications of any new data models or behaviors, such as a new bin map signature. The user can choose a set of keywords to make a subscription, select which data types are of interest, and receive regular emails with updates on any new models or behaviors containing this set of keywords. The emails come with direct links to the reported DSS results, allowing the user to quickly view and verify the results. This ease of use allows the Subscription function to easily integrate into the daily routines of a wide variety of fab engineers.

Fig. 4: Setting up DSS Subscription for any new correlation, map, or process control behaviors for Bin_8 on product VD2420A.

The rapid increase in the number and complexity of devices in modern semiconductor fabs requires more advanced approaches to data management and access. Fab engineers are hard-pressed to keep up with the deluge of data being collected and need autonomous techniques that take a sophisticated approach to data analysis, yet also provide results in an accessible, easy-to-consume way. The Decision Support System from Synopsys is a major step in improving the efficiency of fab engineers by identifying latent behaviors in a wide variety of data sources, correlating these behaviors with one another, and reporting the results to fab engineering users in an easily accessible way.