INFICON partnered with two FabGuard customers to develop SmartFDC Anomaly Detection algorithms and the corresponding web user interface. This on-going project consists of 3 phases:
On an industry standard PECVD cluster tool, a SmartFDC Anomaly Detection model was configured for all CVD chambers. After training the model using 500 processes to generate a preliminary detection envelope, we noticed that the anomaly scores for Chamber C were consistently high when compared to Chamber B (Figure 1).
A brief investigation into the time series data revealed the anomalies for both chambers were correlated. We found that the start of gas flow on one chamber impacted the gas flow on the other chamber (Figure 2). In general, Chamber C was impacted more Chamber B, and the site is currently investigating the gas delivery for the tool to determine if there is an impact to product.
The detected anomaly is sometimes subtle, but still easily seen in the SmartFDC UI (Figure 3). The variation in anomaly scores is likely due to the data acquisition speed. The sampling frequency may be too slow to consistently observe the full dynamic disruption in Chamber C gas flow when Chamber B's gas flow starts.
An offline version of SmartFDC Anomaly Detection is available to analyze existing time series data. This version of anomaly detection was used with data from an industry standard etch cluster tool as part of a root cause investigation.
First, we quickly trained an anomaly detection model from a set of data before the incident using 100 processes. The resulting model was then applied to the processes under investigation. Normalized anomaly scores were calculated, with zero representing non-anomalous processes and greater than zero representing the relative magnitude of anomalous processes (Figure 4).
Investigation into the time series data revealed issues with the Backside He Flow Rate. Figure 5 shows how traces with both subtle and obvious differences in shape during the process can be detected. For the cases where the anomaly scores are low, it is likely that the wafers are not being seated correctly on the chuck. For the cases where the anomaly scores are high, it looks like the wafers had difficulty maintaining constant contact all the way around the chuck.
SmartFDC Anomaly Detection algorithms have been applied successfully to both run-by-run monitoring and root cause investigations. We look forward to continued development with our industry partners.