Adopting Predictive Maintenance On Fab Tools

Predictive maintenance cuts equipment downtime while boosting fab efficiency


Predictive maintenance, based on more and better sensor data from semiconductor manufacturing equipment, can reduce downtime in the fab and ultimately cut costs compared with regularly scheduled maintenance. But implementing this approach is non-trivial, and it can be disruptive to well-honed processes and flows.

Not performing maintenance quickly enough can result in damage to wafers or the process tool itself. Where just-in-time maintenance strategies have been implemented, the result typically is higher overall equipment efficiency (OEE), longer time between maintenance activities, and prevention of yield excursions due to equipment components failure.

Nevertheless, adoption of this approach has been slower than expected. There are still gaps in the communications standards, and fabs remain wary of costly downtime due to a change in technology, methodology, or process. These changes also require up-front investments in analytics and training, so buy-in by upper management is critical, and often based on a well-documented return on investment.

Real-time and preventive maintenance strategies
Industry 4.0, or smart manufacturing, has been focused on optimizing yield and quality by leveraging manufacturing-generated data to improve factory performance and increase production agility. Applying analytics to manufacturing equipment maintenance is a new addition, and one that holds real promise.

To preclude unexpected equipment malfunctions that can impact fab capacity, tool owners follow preventive maintenance schedules. With a more connected factory control and automation systems and the increased use of sensors — vibration, sound, image — it’s now possible to proactively predict or detect failing components or process tools.

A predictive approach basically forecasts impending failure in equipment. That differs from a real-time approach, which is designed to detect the first instance of a failure. Both are valid approaches to improve the ROI significantly over scheduled maintenance routines.

To implement a predictive or real-time maintenance approach requires engineers to comprehend the available measurement data, determine which parameters are most relevant and implement the improved maintenance strategy. Also, to justify the investment to their management, engineers need to present the benefits in terms of fab capacity metrics, product yield, and quality. The more accurate the economic data, the more convincing the ROI justification can be.

“One of the levels of expertise that is missing is the expertise necessary to understand the factory operations from a financial standpoint,” said Mike McIntyre, director of software product management at Onto Innovation. “The finances of operations — in terms of the cost of a process, how much it is to maintain, and the value of the material being put through that process — is yet another vector that must be considered in determining how best to optimize the factory.”

Still, with a judicious use of predictive maintenance, engineers can make better informed decisions about when and how to maintain fab equipment. To do so requires an engineering investment, and this investment is can impede adoption rates.

Tool maintenance practices, standards
Understanding current maintenance practices provides a starting point for understanding what is possible and the potential barriers for an ideal integrated solution for predictive or real-time maintenance.

“In our conversations with customers, we find that most have some sort of fixed schedule preventive maintenance system in place. In other words, they replace a piece of hardware at some predetermined interval, whether or not the hardware is in need of being replaced,” said Wes Smith, CEO of Galaxy Semiconductor. “The time interval is usually based on some analysis of historical patterns, but there’s a healthy dose of educated guessing from the experienced maintenance teams. Sometimes, product metrology serves as the trigger for maintenance. When a product measurement starts to drift, often a maintenance event will be scheduled based on the experience of the local teams.”

Sensors in fab and sub-fab equipment run the gamut of environmental, electrical and physical parameters that can be measured including temperature, gas flow rates, chemical, optical, or vibration sensors. Sensor data typically is collected and communicated using existing industry standards.

“There are standards for equipment connectivity, data collection, and fab operations. These feed into a consistent methodology for predictive maintenance solutions,” said Anjaneya Thakar, director of product marketing at Synopsys. “However, I do not know of any specific standards for predictive maintenance.”

Others agree. “Almost all sensor data can be obtained from the equipment or sensors through CIM interfaces or drivers that adhere to SEMI standards communication protocols (HSMS, SECS, GEM, EDA/Interface-A, OPC, PLC, MQTT, GRPC), or even custom equipment drivers,” said Jon Holt, volume manufacturing solutions worldwide fab applications solutions manager at PDF Solutions. “However, there are no real detailed standards for predictive maintenance strategies and execution that I am aware of. Indirectly, there is the SEMI E10 standard (Specification for Definition and Measurement of Equipment Reliability, Availability, Maintainability, and Utilization); SEMI E58 (Automated Reliability, Availability, and Maintainability Standard: Concepts, Behavior, and Services); and ISA-95 Level 3 (Factory Automation).

Industry gaps
While existing standards provide access to equipment information, there remains much more work to do, particularly in connecting the sub-fab equipment data with the fab equipment data.

“We lack the appropriate trace data from many critical tool components, said Boyd Finlay, process engineer at GlobalFoundries. “We require ‘sensor frequency response’ to measure ‘complete’ Nyquist signals for measurement and control. One very basic example is that we do not see true ‘mechanical drift’ on any valves unless we deploy third-party sensors. This approach takes way too long on huge fleets and is costly. As a result, we are not ready for PdM (predictive maintenance).”

The problem relates to communication protocols that seamlessly integrate, an area where there currently is no standard.

“Main tool and the chamber is where the processing happens,” Finlay said. “Ancillary tools describe the sub-systems of these tools, which are typically located elsewhere in the cleanroom or in the sub-fab — for example, pump, abatement, chillers. The communication bus between ancillary tools and main tool/chambers has not been standardized. We do not get appropriate level trace data for key failure modes on the ancillary kit, and it is not integrated through the main tool/chambers for Host Equipment Interface signal contextualization, i.e. process start/stop, recipe step 1,2,3, sensor A,B,C, etc.”

Predictive maintenance examples
For the easiest foray into a predictive or real-time maintenance, engineering teams can look for shifts in behavior of a single parameter. Consider, for example, the vibration in a robotic arm as it moves a wafer from one location to another. Sensor data can be used to set the baseline vibration level, and trend analysis supplies actionable information that guides maintenance decisions.

“Robot handling is a very critical part of the process, which impacts the quality of the processed wafer. The robot arm contains a platform, controller, arm, drive, end effector, sensor and motors, shaft, bearing and other hardware,” said Vidya Vijay, senior program manager CyberOptics. A robot arm has at least nine parts that could need replacing, and anomalous vibration could indicate failure in a specific part.

Electrical measurements are common in fab equipment and include power, current, and contact resistance.

“In plating applications, measurement of the contact resistance of the pins that enable electroplating of wafers can be recorded when the ring is new, and used after so many lots of processed wafers,” said Vijay. “This can provide a clear picture for when the contact ring has to be cleaned or replaced. As the fingers get coated with the plating chemicals, resistance increases. Resistance can be measured using a WaferSense auto resistance sensor (ARS).”

Vacuum systems are pervasive throughout wafer fabrication factories in deposition, etching, and ion implantation tools. Detecting vacuum leaks in chambers, and impending vacuum pump failure is of the utmost interest to tool owners, process engineers, and technicians.

“One area where predictive maintenance is more prevalent is in vacuum pumps.  We’ve spoken to representatives of more than one vacuum hardware provider who have described to us their methods for predictive maintenance,” said Galaxy’s Smith. “A typical configuration involves an array of microphones and a sophisticated signal processing algorithm. In some cases, these are augmented by electrical sensors measuring the current draw required to maintain the expected vacuum level.”

Suppliers of equipment in which vacuums are essential have been actively working to provide predictive maintenance solutions for their customers. Two papers from the 2022 Advanced Semiconductor Manufacturing Conference provided success stories and they emphasized the ROI with a shift away from scheduled to preventative maintenance.

Edwards Vacuum authors shared their strategy for a predictive maintenance practices for cryogenic pumps.[1] These pumps support the environment in process chambers, and a faulty pump causes contamination those chambers. If a pump fails unexpectedly, costs of the unplanned event are significant compared to a planned service event. Specifically, the authors pointed to:

  • Repair costs: An unplanned down event almost always results in more parts that need to be replaced.
  • Unplanned down time: Unplanned down time always takes much longer to recover from, e.g., due to required chamber cleaning.
  • Tool requalification: This may take considerable time and include additional metrology costs.
  • Wafer cost: Advanced-node wafers each may cost many thousands of dollars.

The Edwards team described using multiple variables based on a data set from more than 100 sensors, with both static and time-series data. To flag an impeding failure, either a rule-based or statistical-based algorithm can be used, and these are augmented with data review by subject matter experts. Both analytical approaches made use of multiple variables.

PSK engineers shared their real-time detection technique for vacuum leaks in dry-strip machines.[2] This equipment removes photoresist from the wafer using a remote plasma source and an O2/N2 mix. The motivation for detecting leaks in real-time is to improve productivity, since conventional methods to check for leaks require the strip tool to be taken out of production mode, directly impacting OEE. This method also changes the process gas composition, which can have undesired processing results. The engineering team chose the following parameters relevant to vacuum leaks — pumping/venting time, angle position of auto pressure control (APC) valve, and chamber base pressures. With a methodical data collection more than 500 runs, they specifically looked at the APC angle relationship to chamber temperature and successfully demonstrated real-time capability to detect leaks from chosen parameters.

Adoption rate of predictive maintenance
“Use of analytics for maintenance is behind using these techniques for yield, performance, and metrology. In this area, there is still much in the way of exploration going on,” said Onto’s McIntyre. To go from exploring to adoption, he further emphasized that the factory teams need to see “hard and factual savings being documented and published in peer-reviewed forums.”

Getting off the traditional scheduled maintenance and implementing preventive maintenance requires management commitment and engineering investment. To be fully enabled, engineers need access to the sensor data, flexibility in data formats for integration purposes, and new standards for communication between various third-party sensors and the systems that monitor such equipment data in the sub-fab and fab.

Unfortunately, providing data from some equipment or sensor’s proprietary format does not foster data integration across a wide variety of equipment and sensor vendors. A fab will have equipment from many vendors as well as invest in third party sensors.

“We provide flexible options for data analysis. We have the entire dataset available in our proprietary software format, as well as in the simplest CSV formats,” said CyberOptics’ Vijay. “We also provide the options of directly integrating the sensor output into the tool GUI for automated maintenance routines.”

GlobalFoundries’ Finlay pointed to gaps in communication between ancillary tools and main process tools. “These gaps perhaps slow a fab maintenance/controls maturation by an estimated five to eight years,” he said. “We could ramp to mature yields and higher net manufacturing times with these issues addressed because it helps to measure the ‘required’ parameters for a Reliability Centered Maintenance strategy.”

To increase the adoption of predictive maintenance PDF Solutions’ Holt highlighted the following areas:

  • Workforce re-education on the capabilities;
  • Industry 4.0 infrastructure adoption and deployment (ability to train and deploy models either on the tools or at the edge);
  • OEM buy-in that does not invalidate warranties or support contracts;
  • Widespread sharing of success stories.

Then there’s connecting the need for maintenance with the necessary information to perform the maintenance.

“Predictive maintenance provides a way to identify potential failure before it occurs. It also provides information about the probability of the potential failure occurrence within a certain timeframe. That by itself is a huge value, however it is only part of the story,” said Yishai Barak, director of service lifecycle management at Siemens Digital Industries Software. “Once you are aware of the potential problem, how do you ensure that your action plan for resolution is done in the most efficient way? Services lifecycle management provides an answer by treating service as part of product lifecycle management and integrating with predictive maintenance solutions. While predictive maintenance assists you with the identification part, service lifecycle management assists you with the resolution part.”

To improve maintenance practices in wafer fabs, engineering teams can use sensor data taken inside and outside of process equipment. Applying analytical decisions to scheduled maintenance has been demonstrated. In this shift to predictive/real-time techniques for maintenance, equipment vendors are actively working on such approaches to help their customers get the most out of fab equipment.

However, as noted by GlobalFoundries’ Finlay, adoption is being slowed by gaps in communications standards that are needed to enable an integrated predictive methodology in a large factory. But with demonstrable ROI, one can expect more adoption on an equipment or process step basis.

“Predictive maintenance is a solution that must be applied to the right problem(s) so that it generates enough ROI to drive the desired behavioral changes,” said Don Ong, head of innovation at Advantest. “It involves having the skills and abilities to identify the right problem or opportunity, and then to execute using the appropriate technology.”


  1. E. Collart, A. Longley, D. Gordon, J. Nordquist and P. Matthews, “Predictive Maintenance Practices for Cryogenic Pumps in Semiconductor Manufacturing,” 2022 33rd Annual SEMI Advanced Semiconductor Manufacturing Conference (ASMC), 2022, pp. 1-6, doi: 10.1109/ASMC54647.2022.9792482.
  2. J. Jeong, T. Ha, H. Ji and S. J. Yoon, “Real-time Vacuum Leak Detection Technology to Calculate Vacuum Leak Parameters for Dry Stripping : EO: Equipment Optimization,” 2022 33rd Annual SEMI Advanced Semiconductor Manufacturing Conference (ASMC), 2022, pp. 1-4, doi: 10.1109/ASMC54647.2022.9792477.

Related Stories
Too Much Fab And Test Data, Low Utilization
For now, growth of data collected has outstripped engineers’ ability to analyze it all.

Much Smarter Manufacturing
How AI, ubiquitous connectivity, and sensors everywhere are reshaping manufacturing of chips, and nearly everything else.

IIoT And Predictive Maintenance
Users are realizing ROI through predictive maintenance with new technology.

Leave a Reply

(Note: This name will be displayed publicly)