RD-11.15-16.2022-Boning

Conference Video|Duration: 31:01
November 16, 2022
View this past event
Please login to view this video.
  • Video details
    Advanced manufacturing, and semiconductor manufacturing, in particular, provides many opportunities but also challenges for machine learning. State-of-the-art factories collect enormous amounts of highly diverse sensor data. Such data can give insight into the health of the equipment and process, enabling rapid anomaly detection. For virtual metrology, sensor data reflecting chamber operation can enable the estimation of resulting wafer states in place of costly ex-situ wafer measurement. For process optimization and control, combinations of empirical learning algorithms and existing physical models or digital twins are of great interest. However, practical application of existing machine learning methods to these needs in semiconductor (and other) manufacturing contexts often encounter important challenges: small data and concept drift. While any one run may have a large amount of sensor data, only a relatively small number of runs are available compared to what is needed for many machine learning methods (e.g., very few or no "bad" runs for training an anomaly detector). In addition, the accuracy of carefully constructed models often "decay" with time, due to subtle drift in equipment, wafer, or fab environment. These can be addressed by calling on or extending a rich set of less common machine learning methods, ranging from density estimation, time series methods, and Bayesian approaches, in addition to deep learning. The future of manufacturing requires machine intelligence, and machine learning research will also benefit from the challenging opportunities and problems that can be found in manufacturing.
Locked Interactive transcript
Please login to view this video.
  • Video details
    Advanced manufacturing, and semiconductor manufacturing, in particular, provides many opportunities but also challenges for machine learning. State-of-the-art factories collect enormous amounts of highly diverse sensor data. Such data can give insight into the health of the equipment and process, enabling rapid anomaly detection. For virtual metrology, sensor data reflecting chamber operation can enable the estimation of resulting wafer states in place of costly ex-situ wafer measurement. For process optimization and control, combinations of empirical learning algorithms and existing physical models or digital twins are of great interest. However, practical application of existing machine learning methods to these needs in semiconductor (and other) manufacturing contexts often encounter important challenges: small data and concept drift. While any one run may have a large amount of sensor data, only a relatively small number of runs are available compared to what is needed for many machine learning methods (e.g., very few or no "bad" runs for training an anomaly detector). In addition, the accuracy of carefully constructed models often "decay" with time, due to subtle drift in equipment, wafer, or fab environment. These can be addressed by calling on or extending a rich set of less common machine learning methods, ranging from density estimation, time series methods, and Bayesian approaches, in addition to deep learning. The future of manufacturing requires machine intelligence, and machine learning research will also benefit from the challenging opportunities and problems that can be found in manufacturing.
Locked Interactive transcript

More Videos From This Event

See all videos