Metrology is instrumental to all aspects of modern semiconductor manufacturing, especially to advanced process control (APC) and rapid yield learning. Too few measurements result in loss of in-line control and off-line traceability, whereas too many measurements result in waste of money, time, and space. The sampling rate, thus, becomes a fundamental tradeoff, subject to physical and financial constraints in high-volume manufacturing (HVM).
As a solution to break the fundamental tradeoff imposed on physical metrology, virtual metrology (VM) aims to predict post-process metrology variables using process and wafer state information such as fault detection and classification (FDC) and real metrology data. Thanks to both conceptual appeal and immense practical value, numerous attempts, both academic and industrial, have been made to develop VM over the past two decades. Many challenges, including model and data quality, model maintenance, and data and intellectual property (IP) security issues, however, have prevented the broad adoption of VM in HVM fabs.
This talk introduces Panoptes VM, the industry’s first solution that addresses these challenges. The predicted metrology variables from Panoptes VM are highly accurate for a variety of devices and tools with robustness against drifts and shifts in chamber states. With built-in automation, it is straightforward and effortless to set up and maintain hundreds of thousands of models, which are required for fab-scale operations. Although Panoptes VM is based on domain knowledge, the core algorithm is universal and agnostic to underlying customer data, avoiding any data or IP security issues.
The unique impact of Panoptes VM is illustrated through its performance at state-of-the-art HVM fabs, in which Panoptes VM is applied to real-time wafer-to-wafer APC and reduces process variability by 21.5% on average. The talk concludes by exploring other applications of Panoptes VM and presenting a roadmap for addressing the remaining challenges.