Behind the scenes, advanced equipment control and robotics are enabling higher levels of repeatability, tool uptime, and yield.
Key Takeaways:
Robotics and AI are taking on more tasks in fabs today, particularly to reduce unscheduled downtime and ensure more good wafers are produced with existing capacity.
Both leading and legacy fabs are boosting operational uptime using collaborative robots and autonomous vehicles equipped with sensors to replace many of the mundane tasks performed by humans in the past. Automation ensures repeatability, reserving human intervention for problem-solving activity.
At the same time, process engineers now have the computing capability at their fingertips to better tie together equipment state, process context, and wafer results, one of the pivotal ways they keep variation under control despite fabricating devices with nanometer-scale critical dimensions and tolerances.
“The strongest quality-control strategies connect material data, equipment data, and wafer results so that subtle sources of variation can be detected before they become a yield problem,” said Jon Holt, worldwide fab applications solutions manager at PDF Solutions. “A material may meet its incoming specification yet behave differently because of delivery-line condition, temperature history, residence time, chamber condition, or interaction with another material. Modern equipment-control and data-collection systems help preserve this context, enforce operating limits, identify drift, and prevent processing when critical parameters fall outside an approved range.”
Rising manufacturing costs, global expansion and labor shortages all contribute to greater implementation of automation, robotics and AI.
“There are three trends driving big advances in equipment intelligence,” said Russell Dover, managing director of equipment intelligence product development at Lam Research. “One is the increasing cost of manufacturing and the critical need to get as much productivity and ROI from existing investments. When I started at a fab in the UK over 30 years ago, a fab cost around $20 million. Now that barely covers a set of equipment, with a gigascale fab costing around $20 billion. Second, fabs are booming in new geographic locations, so companies want to replicate the capabilities of a fab in one region in a similar fab in a completely different location, creating interesting fleet-management challenges. Third, our significant talent gap is fueling a need for more automation, AI, machine learning, and robotics — not to replace people, but to embrace robotics to consistently perform repetitive or tedious tasks while taking better advantage of human ingenuity and the speed advantages that ML and AI can afford.”
Robots and cobots improve precision, ensure safety
Intel Foundry uses Boston Dynamics’ quadruped robot to provide real-time monitoring of equipment conditions, largely behind-the-scenes in the subfab (see figure 1). There, hot motors, pumps or electrical equipment may give off a thermal signature due to an impending overheating situation. Failing motors tend to vibrate. Leaking air or specialty gases can make high-frequency sounds that acoustic imaging systems can differentiate from background noise. Such robots also can perform routine, simple tasks that require weaving through tight spaces to record readings on digital or analog gauges. Intel Foundry’s outfitted robot, called Chip, play an important role in maintaining equipment health in multiple facilities without requiring a technician or engineer to physically visit a location just to record and report back routine readings to a host controller.
Fig. 1: Intel Foundry equips the Boston Dynamics’ quadruped robot to continually perform visual, thermal, and acoustic inspections in its subfabs, ultrapure water, and air cooling/distribution facilities. Photo: Semiconductor Engineering
So far, Intel says its robots can perform 16 different preventive inspection tasks using visual, thermal, and acoustic inspections, in addition to lidar and RealSense depth cameras. This approach improves the quality of data collected during inspections by streamlining the process for technicians and decreasing safety risks.
“For example, during a routine inspection of boiler pumps, Chip detected hot water unexpectedly moving through the line of a powered-down pump,” explained Joe Robinson, IoT & Robotics Architect at Intel Foundry in a recent article. [1] “The robot alerted area technicians about the anomaly, enabling them to safely isolate the pump for diagnosis. Technicians found a non-functioning gasket in the pump. Together, this proactive repair by humans and a robot prevented possible hot water leakage and additional equipment damage.”
In practical fab terms, Intel Foundry is using Chip robots to perform:
Safety drives new automation solutions
Robotics and automation are especially important after emergencies. For example, fab personnel in earthquake-prone Taiwan have found a better way to recover from a disaster. Robots now perform safety inspections, which would otherwise expose personnel to gas leaks, toxic chemicals, or collapsing objects.
UMC’s fab facility, smart manufacturing and IT teams together developed an Autonomous Mobile inspection Robot (AMR) to safely replace manual inspections. Formerly, each day two technicians would spend 1 hour or more to complete routine inspections, which are now performed automatically.
In a similar manner to Intel’s approach, UMC’s AMR uses ultrasonic acoustic cameras to detect high-frequency sound waves to pinpoint gas or liquid leaks or detect sound wave oscillations caused by vibrating equipment. The robot’s AI precisely locates potential sources of anomalies and triggers early intervention by the preventive maintenance team to address equipment problems and prevent equipment damage.
Fig. 2: The Autonomous Mobile Inspection Robot (AMR) helps preventive maintenance teams identify thermal signatures, gas leaks, liquid puddles, and records readings on gauges. Source: UMC
The AMRs also use thermal sensors to detect hidden signs of equipment overheating as well as liquid leakage, common hazards in fabs and subfabs. The thermal imaging system can detect liquid residues on the ground that are difficult to see with the naked eye, identify leaks in a timely manner, and clearly indicate potentially overheating parts so engineering personnel can quickly address problem areas.
How equipment control reduces wafer variation
Process variation is becoming increasingly critical with every new device node. “As device dimensions continue to shrink and process windows tighten, nearly every process becomes more sensitive to variation, but patterning, deposition, etch, CMP, and thermal processes are especially critical,” said Alan Weber, vice president of new product innovations at PDF Solutions. “Small differences in chamber conditions, material delivery, temperature, pressure, plasma behavior, endpoint detection, or equipment state can translate into meaningful changes in critical dimensions, film properties, overlay, uniformity, and ultimately device performance and yield.”
Engineers are going after multiple sources of variation. “Advanced device architectures require more selective processes, thinner and more conformal films, and tighter control of interfaces between materials. This makes the condition and history of the equipment increasingly important,” said Patrick Pannese, vice president of strategy and business development at PDF Solutions. “Variation may arise not only between tools, but also between chambers, recipes, maintenance cycles, consumable lifetimes, and even successive wafers within the same lot. Consequently, equipment health monitoring, high-frequency data collection, chamber matching, run-to-run control, and correlation of equipment data with wafer results are becoming essential parts of process control.”
Returning process tools to useful production status faster
Generally, whether it is at the process tool or subfab equipment level, fabs are increasingly interested in moving from unscheduled downtime to predictive maintenance. “Any time you can use machine learning and analytics to predict that something is failing, you can move from what would become an unscheduled event into a scheduled event that is also very attractive from a fab operational standpoint,” said Lam Research’s Dover.
For example, Lam Research offers both plasma-based etching equipment and deposition equipment that requires a repeatable chamber condition before process wafer runs. “We use the optical emission spectroscopy (OES) data from a plasma to indicate when a chamber is ready to process wafers. For instance, right after a chamber has been cleaned during maintenance, it is not prepared to run wafers,” said Dover. “First, it needs to be ‘conditioned,’ or uniformly coated with a film. Most conditioning is based on a learned limit, running the plasma in the tool for 25, 50, 100, or 200 RF hours while running dummy wafers until the surfaces are uniformly coated. This does give you a known condition, and it works.”
“But if you look at the OES spectrum, you can start to determine when you actually achieve that condition,” he said. “This is important for two reasons. One, you get that chamber back into production sooner, so we improve the tool availability and get more wafers out. The second is that the life of a chamber between maintenance cycles is a function of the total RF hours. If you’re burning 200 to 300 hours just doing seasoning, that’s 200 to 300 hours that gets subtracted from useful production because it goes toward the next maintenance cycle.”
Tool-to-tool matching
Within a given fab, process engineers routinely perform tool-to-tool matching to ensure wafers processed by different tools using the same materials and the same recipe produce the same results. “The primary goal of tool matching (and its first cousin, chamber matching) is to avoid the dedication of specific production equipment and its subsystems to certain critical products or layers just because they come out better on these ‘golden tools,'” explained Steve Zamek, director of product management at PDF Solutions. “This approach not only complicates factory scheduling, but also reduces overall capacity of the fab.”
Both are undesirable side effects. “Run-to-run technology can be applied to capture the current tool state so that this can be factored into subsequent runs using feedforward models that combine material state and tool state to then adjust recipe parameters to achieve the desired result,” Zamek said. “A more robust approach uses so-called fingerprinting technology to characterize the underlying mechanisms in the equipment that can affect the process result (power generation, flow control, pressure control, etc.) and ensure that these mechanisms in all tools that need to be in a “matched set” are brought within tight tolerance of the reference tool or specification. The assumption underlying this approach is that if all individual components of the equipment are behaving consistently (i.e., they are ‘matched’), then the tool in aggregate should behave consistently, as well.”
With global expansion of fab activity, which is forecast to double the size of the semiconductor industry within five years, there is pent-up need to apply “matching” procedures from one fab to another, taking quality control to even higher levels.
Many process tools, such as deposition, etching, and annealing systems, contain multiple identical chambers under vacuum to accelerate throughput. For these systems, chamber-to-chamber matching is important because a problem on every third wafer or every fifth wafer in a lot can be missed, impacting yield down the line. A macro defect inspection tool equipped with randomization features can detect subtle differences between wafers using randomization to provide insight about where process variation originated.
“In one example, a slight difference in signal intensity showed a difference in topography of only about 1,000 angstroms of BPSG (boron phosphosilicate glass), but it occurred only on one out of every third wafer,” said Errol Akomer, applications director at Microtronic. “Using a randomization strategy where wafer position is shuffled within lots, problems that occur on particular wafers can be traced back to their source more quickly.” [3]
Here come the cobots
The semiconductor industry has been using x-axis robots for wafer handling inside process tools, die inside pick-and-place tools, as well as automation for moving wafer lots (FOUPs) between tools in the fab for decades. What’s different about cobots is that they are not fully tied to one piece of equipment and can be reconfigured to manage new situations. For instance, the same cobot can remove screws from a chamber opening and perform chamber cleaning using different end effectors and recipes (see figure 2).
Fig. 3: The collaborative robot (cobot) performs maintenance procedures using different end effectors to loosen screws and clean the chamber in a safer, more repeatable manner than a human can. Source: Lam Research
One of the key areas of focus for cobots is reducing unscheduled downtime. ‘Most unscheduled downtime comes from mistakes that happened during preventive maintenance,” said Lam Research’s Dover. “We recently introduced our Dextro cobot with multiple end effectors, which can clean process chambers using a dry ice spray, for instance, with a higher level of precision and repeatability than a human performing the same task, while ensuring a higher level of safety. For example, even when the cobot is driving in screws, a person has a hard time applying exactly the same amount of torque each time, but this is not an issue for a cobot. And semiconductor process tools are not designed ergonomically, so technicians have to adapt to the tool, not the other way around. There are also savings from the standpoint of PPE (personal protective equipment) and the time needed to bring in and train technicians.”
When applied across hundreds of tools, the result can be substantial bottom-line cost savings. “Our customers not only document better outcomes in terms of the preventive maintenance procedures being done correctly, but the reduction in variability is showing up as improved Cpk and marginal yield, which can amount to thousands to millions of dollars,” said Dover. The greater repeatability also improves tool-to-tool performance matching and equipment uptime.
Dreams of the “lights out” fab
Ever since the development of the front-opening unified pod (FOUP), chipmakers have talked about the dream of fully autonomous fabs where on-site engineers and technicians are needed only to work behind the scenes. For some, this dream includes the use of humanoid robots.
“A number of key customers have visions of autonomous fabs or lights-out fabs, sort of a “Star Trek Next Generation” view with humanoid robots that just go around and do the work,” said Dover. “Whether that is pure science fiction, or whether the current leaps and bounds in AI technology will actually get us there, I don’t know. But I do know that people are aspiring to have humanoid robots doing preventive maintenance.”
It’s important to note that these robots have to interoperate with SEMI standards, host communication, and host control. “Compliance with SEMI standards would increase the building space by 30% versus non-SEMI compliance because you need safe paths for people, so the cost factor is way bigger than people are assuming,” said Dover. Also, there are other incompatibilities. “For instance, a humanoid robot operates for a limited time on a battery pack, and some of these maintenance routines take 12 hours. There are significant power issues to be overcome.”
So as humanoid robots mature, many fabs are reaping the benefits of inspection robots and cobots to help keep equipment in the fab, subfab, ultrapure water, and other supporting facilities up-and-running.
References
Related Articles
Increasing Roles For Robotics In Fabs
AI and robotics are taking on bigger, more complex, and increasingly autonomous tasks, but integration with existing equipment and processes remains a formidable challenge.
Tool Matching Getting Tougher Across Test & Metrology
New techniques are needed to keep pace with ever-shrinking features and tolerances.
New Frontiers In Fault Detection And Classification
AI is enabling timelier and more accurate data, but AI-based command and control has yet to appear.
Leave a Reply