(136b) Big Data Analytics Applied to Semiconductor Manufacturing

As "big data" requirements and capabilities in data volumes, velocity (rates), veracity (quality), variety (merging) and value (analytics) increase rapidly, semiconductor manufacturers are faced with a need for new approaches to data management and use across the fab [1]. During semiconductor manufacturing a silicon wafer is transformed into a grid of silicon chips in a fabrication facility or "fab". Each chip or "die" has layers of complex circuitry that must be controlled to the nanometer or even molecular level. The semiconductor manufacturing process usually has hundreds of these precision control steps. The dynamic control of these steps in this industry is facilitated with a set of capabilities collectively termed "Advanced Process Control" or APC [2]. Over the past two decades the APC family of capabilities has included dynamic model-based process control, termed "run-to-run" or R2R control [2,3], and Fault Detection and Classification or "FDC" (also referred to as excursion detection and failure mode determination) [2,4]. More recently the APC family of capabilities has been expanded to include predictive capabilities such as Predictive Maintenance (PdM), Virtual Metrology (VM) and even end-of-line yield prediction and yield excursion root cause analysis. The expanded set of APC capabilities represents the primary area where big data solutions are generating significant benefits in semiconductor manufacturing.

Big Data and Semiconductor Manufacturing Advance Process Control

Big data advancements have facilitated improvement in all of the APC capability types. The following are two examples where significant return on investment (ROI) is being achieved.

ï‚· Predictive Maintenance: Predictive Maintenance, PdM, also referred to Predictive Preventative Maintenance, PPM, or a component of Prognostics and Health Management (PHM) is receiving considerable attention in semiconductor manufacturing because it is focused on reducing unscheduled downtime, which is a major source of quality loss and cost in manufacturing [5]. PdM is the technology of utilizing process and equipment state information to predict when a tool or a particular component in a tool might need maintenance, and then utilizing this prediction as information to improve maintenance procedures [2]. This could mean predicting and avoiding unplanned downtimes and/or relaxing un-planned downtime schedules by replacing schedules with predictions. In semiconductor manufacturing some PdM solutions are relatively straight-forward extensions of FDC, however most involve extensive data mining and analysis capabilities to develop models that not only predict the future occurrence of an equipment or component failure, but, as illustrated in Figure1, also indicate the time horizon for remaining useful as well as a confidence and range of the prediction. These PdM prediction events are linked into maintenance, inventory, scheduling and manufacturing execution systems to optimize uptime, cost, productivity and quality [5].

Figure 1: Illustration that PdM solutions require not only the actual prediction, but also a prediction horizon (Time to Failure-TTF) and an indication of prediction quality (provide here as a 95% probability range).

ï‚· Next Generation FDC: Fault Detection (FD) utilizes real-time trace data collection of key equipment and process variables, along with process output quality information and advanced analytics to determine when an excursion has occurred. Fault Classification (FC) then tries to relate the excursion to a cause or classification. FDC is particularly challenging in semiconductor manufacturing because the processes complex and dynamic to the point that purely empirical solutions generally are ineffective. In other words equipment and process subject matter expertise (SME) must be incorporated into the development and maintenance of these models. As a result the FD model development process is usually highly manual and univariate leading to tens of thousands of models and millions of limits to monitor and manage across the fab. With the advent of big data capabilities the model and limits building and management can be semi-automated and optimized leveraging big data analytics advancements while retaining the ability to incorporate SME in solutions. New FD enhancements that are proving effective include semi-automated trace partitioning, feature selection, and merging in of process quality data to support feature ranking and limits optimization through receive-operating-characteristic (ROM) analysis. Interestingly many of the same algorithms leveraged for PdM can be applied here especially for ranking and optimization. The primary benefits over traditional FD are significantly reduced time for model setup and maintenance, and reduced false and missed alarms [6].

Industry Requirements for Big Data Analytics

The choice of analytics being leveraged to provide enhanced APC capabilities in semiconductor manufacturing reflects the complexity of semiconductor manufacturing processes in terms of the need for precision on 3 analytics response, the dynamics of the processes, and the need for incorporation of (human) SME into the analysis. As an example, typical multivariate supervised techniques such as Partial Least Squares (PLS), and Support Vector Regression (SVR) are often leveraged for predictive capabilities, however clustering techniques are also needed to help separate out the prediction event types, model adaptation techniques are needed to track the process dynamics, phenomenological model elements are often needed to improve empirically-tuned model accuracy, and the entire model building and maintenance process must be structured to support SME involvement. Additionally predictions must include an indication of prediction accuracy so that decision making can be optimized based on an understanding of the potential for false or missed alarms and the associated financial impact. All of these techniques benefit from the increased data volume and improved data quality for model building, however volume and quality of data to support robust models is still the primary impediment to the widespread success of prediction capabilities in the industry. Thus while many other manufacturing environments might turn to more general big data techniques, such as deep learning, for generalized analysis solutions, the semiconductor industry is currently focused on more guided techniques that leverage an understanding of the process, equipment and data management domain knowledge.

Summary

As the semiconductor industry advances into the era of smart manufacturing, specific requirements must be met to address the challenges of increasing manufacturing complexity. As shown in Figure 2, the big data infrastructure will enable a number of capabilities from manufacturing operations optimization, through supply chain network integration, and the leveraging of the simulation space and the concept of a "digital twin" for continuous fab optimization [7]. APC capabilities continue to be improved in order to optimize individual processing with respect to maximized quality and throughput and minimized cost. These capabilities include control, diagnostics, and more recently, predictive solutions. They leverage big data opportunities in analytics that allow for the incorporation of process, equipment and data expertise across the entire solution development and maintenance life cycle. It is expected that this approach will continue as the industry evolves from reactive to predictive to prescriptive computational analysis.

Figure 2: The Fab of the future requires integration of equipment, expertise, manufacturing capability and computational data analysis

References:

[1] J. Moyne, J. Samantaray and M. Armacost "Big Data Capabilities Applied to Semiconductor Manufacturing Advanced Process Control," IEEE Transactions on Semiconductor Manufacturing, Vol. 29, No. 4, November 2016, pp. 283-291.

[2] 2015 International Technology Roadmap for Semiconductors (ITRS): Factory Integration Chapter. Available: www.itrs2.net.

[3] Run-to-Run Control in Semiconductor Manufacturing, J. Moyne, E. Del Castillo, and A. Hurwitz, CRC Press, November 2000.

[4] T. Lee and C. O. Kim, "Statistical Comparison of Fault Detection Models for Semiconductor Manufacturing Processes," in IEEE Transactions on Semiconductor Manufacturing, vol. 28, no. 1, pp. 80-91, Feb. 2015.

[5] J. Iskandar, J. Moyne, K. Subrahmanyam, P. Hawkins and M. Armacost "Predictive Maintenance in Semiconductor Manufacturing: Moving to Fab-Wide Solutions," Proceedings of the 26th Annual Advanced Semiconductor Manufacturing Conference (ASMC 2015), Saratoga Springs, New York, May 2015.

[6] J. Moyne, B. Schulze, J. Iskandar and M. Armacost "Next Generation Advanced Process Control: Leveraging Big Data and Prediction," Proceedings of the 26th Annual Advanced Semiconductor Manufacturing Conference (ASMC 2016), Saratoga Springs, New York, May 2016.

[7] J. Moyne, "Emerging trends in IC manufacturing analytics and decision making," APC Conference XXVIII, Phoenix, AZ, October 2016. Available: www.apcconference.com.