Panoptes VM - Highly Accurate Whenever, Wherever

As we discussed in the previous newsletter, #VirtualMetrology (VM) is the technology of predicting post-process metrology variables using process and wafer state information such as sensor data. In order to utilize VM data in high-impact downstream use cases, such as advanced process control for reducing process variability, statistical process control for preventing excursion, or metrology scheduling for saving costs associated with physical metrology, it is critical to have high accuracy comparable to real measurement. Accurate prediction of metrology variables, however, presents a series of technical difficulties.?

(1) Dimensionality of sensor data. For each wafer, a massive amount of data is generated from hundreds of sensors in a process tool, making it difficult to identify informative features for accurate and reliable prediction.?

(2) Scarcity of measurement data. Only a few samples of physical measurement data are available for modeling each process tool or chamber.?

(3) Data drifts and shifts. Sensor and metrology data drift over time and shift as the equipment condition changes, for example, from residue buildup on chamber walls, preventive maintenance, or layer change.

(4) Variety and volume of cases. To achieve high prediction quality, individual models are needed for each recipe, chamber, process, and device, the management of which becomes intractable at a fab scale.?

Conventionally, virtual metrology algorithms focused on modeling the input-output relationship between the sensor (#FDC) data and metrology variables. Panoptes VM takes a holistic approach to capture process and wafer state information through innovative machine learning algorithms and platforms that are developed specifically for semiconductor process data. Among the key ingredients to our approach is our online adaptive learning framework that addresses the problem of data drifts and shifts successfully across a variety of devices, recipes, measurands, and chambers, as presented at ASMC 2023. This framework, combined with several other ideas such as non-stationary feature selection and aggregate modeling across multiple chambers, achieves “high accuracy whenever, wherever.”

The high accuracy of Panoptes VM is demonstrated through its real-world performance at high-volume manufacturing fabs. For example, Panoptes VM is adopted by SK hynix , the second largest memory chip maker in the world, to predict several metrology variables related to thickness and refractive index of deposition processes for memory devices. Earlier results show that when connected to real-time recipe control systems, Panoptes VM reduced process variability by 21.5% on average and improved yield as well.

In the next issue, we will discuss the usability of Panoptes VM - how easy it is for process engineers to set up and maintain VM experiments, how human efforts are minimized in setup and update, and how it is streamlined in connecting with downstream applications.?

Regards,

Mike Young-Han Kim