Digital twins are virtual representations of physical systems. The current interest in them is fueled by the convergence of IoT, machine learning and big data technology. As process complexity increases, they are becoming key to efficient operations and high product yields. This presentation focuses on implementation of a semiconductor manufacturing digital twin for yield that detects associations between product quality metrics and up to millions of predictor process parameters - primarily equipment sensor traces and process measurement data.
Moore's law continues not only to drive exponential increases in the performance and storage capacities of integrated circuits, but also in the volumes of data produced by IC manufacturing processes. Process equipment is increasingly better instrumented with sensors, and the number and complexity of the processing steps are growing rapidly. The process complexity necessitates, and the available data volumes enable, a shift to ever more data-driven yield improvement, leveraging the latest big data, machine learning and AI technologies.
There is now a demand for ‘wide-and-big data’ analytic solutions that detect associations between product quality metrics and thousands to millions of process variables (process measurements and raw equipment sensor traces). These cutting edge solutions can support root-cause, clustering, and other analyses at the die level. Further, the results must be available close to "real-time", to enable useful process interventions — for example to identify subtle equipment changes, process shift or drift, or to predict and remedy substandard yield for a lot in the line.
In response to these requirements, hybrid big-data plus in-memory systems are being utilized to address the various new analytic and IT-architecture problems associated with this challenge. They combine large-scale distributed analytics capabilities with comprehensive server- and in-memory-based advanced analytics, and deliver actionable interactive results through intelligent visualizations. One such system involving data pre-processing, sensor trace compression, large-scale feature screening & ranking and detailed modeling and visualization will be described along with an example of its use and performance results.