Better Analytics Needed For Assembly

Many defects occur after manufacturing, but not all of them are being caught.


Package equipment sensors, newer inspection techniques, and analytics enable quality and yield improvement, but all of those will require a bigger investment on the part of assembly houses.

That’s easier said than done. Assembly operations long have operated on thin profit margins because their tasks were considered easy to manage. Much has changed over the past several years, however. The role of assembly houses has grown significantly as device scaling becomes more expensive and power/performance improvements diminish, making advanced packaging significantly more attractive. In addition, there is a growing emphasis on more reliability in many markets, raising the need for more testing, inspection and analytics of these multi-chip packages. That requires more sophisticated and expensive equipment, as well as new methodologies. The combination of all of these factors has put a squeeze on smaller assembly houses, which do not have the resources of the larger OSATs.

“Their budgets are a fraction of what they are on the front-end side,” said Subodh Kulkarni, president and CEO of CyberOptics. “The mindset to inspect and measure is a lot lower in the backend area than on the frontend side. All the methodical disciplines established in the frontend and don’t necessarily exist in the backend.”

There is widespread agreement that improvement is needed for everything from wire bond to 2.5 and 3D packages. An increasing number of inspection steps, more data from analyzing equipment, and assessing of process variables will bring the packaging manufacturing process closer to the kind of attention to detail that has long been heralded in the fab. But this can only happen with a mindset change of OSATs and IDM assembly operation management.

While assembly processes have exceedingly high yields, not all defects can be detected by electrical test and these result in field failures. There are several reasons for this. First, larger die and multi-chip products result in higher and denser connectivity, along with diversity in package sizes along all three axes.

“Assembly processes have narrower process windows than 30 years ago,” noted Scott Jewler, COO at SVXR. “The pitch has reduced significantly, and as the pitch reduces the size of the interconnect also reduces. With higher interconnect counts there’s more functional integration.”

That leads to other challenges. “We have seen increasing complexity in the IC units manufactured over the last decade,” said Olivier Dupont, product marketing manager in KLA’s ICOS Division. “Devices are now getting up to 120 x 120 millimeters for artificial intelligence applications, and these large packages come with specific packaging inspection challenges. For these, as well as smaller advanced packages, we now inspect on top, bottom and side at much higher resolution. A second trend is the attention to package thickness. Mobile device makers require thin packages, and this drives the need for process control as the tolerance in mobile systems is minimal.”

Others agree. “The variations and complexity of advanced packaging continues to increase, and the size is decreasing,” noted CyberOptics’ Kulkarni. “It is key to ensure defects can be detected and critical parameters can be measured accurately, and at a speed that would enable 100% inspection and metrology. All it takes is one bad connection and the whole package fails. Having a robust, 100% inspection and metrology process in place is critical to ensure long-term reliability and yields.”

What’s different
Quality and reliability always have been a concern. But as chip lifecycles lengthen, and as advanced-node chips are used in safety-critical applications, the focus on reliability is increasing up and down the supply chain.

Nowhere is this more evident than in automotive. Anticipating the dramatic increase of ICs in vehicles — an estimated 6,000 to 8,000 devices, according to recent projections —automakers are demanding no more than 10 defective parts per billion. Even wire bond units need to meet this quality goal.

Now it’s creeping into other markets, as well. In his recent SEMICON West presentation, Jewler reminded listeners that the failure rate for the 2005 XBOX 360 at launch was at least 10%. In contrast, when Samsung delayed last year’s release of the Galaxy Fold, it was due to less than 100 failures in pre-release. The reason is that each defect can cost millions of dollars.

Most of the attention is focused on the actual manufacturing and materials used in manufacturing, but defects can creep in during the assembly process, as well. Even consumer IC device makers have increased their expectations for assembly factories.

“You’ve got somebody in the supply chain making assembly parts. They are absolutely critical to the whole thing,” said Dave Huntley, business development director at PDF Solutions. “The big fabless customers want their suppliers to be able to ramp up fast and stay ramped up and respond to problems really quickly. Without the necessary process controls in place, they simply can’t do it.”

To address the anticipated needs of assembly factories, inspection equipment suppliers, assembly equipment suppliers, and data analytic companies have been actively developing technologies and discussing the deployment of these technologies with factories to address these higher expectations.

Solder balls and wire bonds
Fundamentally, it comes down to guaranteeing connectivity throughout the product’s life. Making connections between a die and package substrate depends on the packaging technology. There are two basic steps — die attach/alignment and bonding. While failure modes are well-known, electrical test does not detect them all.

Fig. 1: Wire bond failure modes. Source: Semiengineering/Susan Rambo

Fig. 2: Flip chip solder ball failure modes. Source: SVXR

The interconnects within a package are electromechanical in nature, and certain interconnect defects manifest as reliability failures due to environment and use conditions.

“With flip-chip a lot of our customers want to find contact non-wets (typically called cold solder joints),” said Jewler. “A solder ball makes contact with the substrate pad but didn’t fully wet to the substrate. These defects occur at a few hundred parts per million (ppm). Yet they are the most frightening because they pass electrical test. Then in the field it goes through some thermal cycles or gets bent and it fails.”

Not all of this can be caught using existing methods. “We did some studies with a customer where we measured resistance through various poorly formed solder joints, and it’s really hard to correlate,” he noted. “So it’s difficult to find these with an electrical test.”

Still, much more is riding on getting this right than in the past. “For flip-chip packages you want to be able to detect reliability issues with the solder ball connections,” said Craig Hillman, director of product management for new and emerging technologies at ANSYS. “You can electrically check at time zero before final test, yet the tests are fairly simple — DC go/no-go. If you have a weak connection that passes electrical test, you may not have a sufficient mechanical/metallurgical connection, which makes it vulnerable to failure due to vibration and temperature cycling.”

This is particularly important in safety-critical applications. “Automotive products also experience a larger range of operating conditions, especially temperature,” said Tim Skunes, vice president of R&D at CyberOptics. “Bump-based connections in advanced packages — between chips, substrates, and the outside world — are a source of reliability failures, especially when subjected to repeated thermal stress over extended lifetimes. It puts a premium on bump inspection and metrology.”

Electrical test is very useful for signal pins. Not all pins are used for signals, though. About 30% to 40% of bonds are for power.

“We need more tests for the chip-package interface,” said Andy Heinig, Department Head of Efficient Electronics at Fraunhofer IIS’ Engineering of Adaptive Systems Division. “There are currently a lot of power pins between the chip and the package. For older nodes with higher supply voltages, there are now issues if some of the pins are missing due to assembly problems. But with a decreased supply voltage, every pin is necessary for current functionality.”

Equipment sensors and process variables
To understand assembly process control, engineers sample parts to check quality. This limits the ability to detect the cause of those problems, however. Most experts believe innovation is required in inspection methods and in process variable collection.

Unlike in wafer manufacturing, assembly manufacturers do not typically look at data from the actual equipment performing key process steps. This is partially because the mindset has not been on process control, which requires investment in hardware, software, and engineering analysis. It’s also partly due to the fact that not all data is shared across the supply chain.

This is starting to change. Even wire bond manufacturers have an interest in doing so. While it’s an older technology, about 75% of today’s packages use wire bonding, according to TechSearch International. All customers need the improved quality.

PDF’s Huntley provided an example from the die attach process. “Die attach machines apply pressure to achieve a specific height,” he said. “If the pressure varies from the norm, that indicates something was wrong. Maybe there was a particle underneath the package, or the epoxy dispense prior to applying pressure had an issue, or maybe the epoxy was partially set. This can be a source of yield loss at final test failure, or even worse, field failure.”

The larger OSATs have begun to collect data by installing sensors on tools. “We work with some of the biggest packaging companies,” said Jon Holt, senior director of fab applications at PDF Solutions. “Variables for data collection include deformation, oven temperature, atmospheric pressure and wire pull. They find correlations between this data and failures.”

This enables the data analysis to support fault detection and classification (FDC). But it gets complicated because the number of process and equipment variables increases with newer packaging technology. Consider, for example, that most flip chip bonding uses a mass reflow process. After placement it runs through a reflow oven.

“The surface tension on the solder kind of self-centers the die,” said Jewler. “While you can get failures from warpage between substrate and die due to coefficient temperature differences, the process variables consist of placement accuracy and temperature profile.”

However, the advanced package processes that are used to support high-bandwidth memory and high-end graphics processors are moving to thermal compression bonding, which is an in-situ bonding process.

“You’re actually placing the die, then applying heat and force at the same time to form the bond,” Jewler said. “You no longer get the benefit of surface tension self-aligning the die because you are holding the die in a constant position. This results in more process variables — planarity of the die to the substrate, tilt, and x-y alignment. In addition, the tool must sense when the solder melts and pull the force way down. This requires an extremely sensitive control for the position in the Z axis and the force that’s being applied to form the bonds. The process does have a lot of opportunity to utilize feedback control using the subsurface inspection technology that SXVR provides. We are now developing on the tool the ability to check and measure shift in x and y axes, rotation, and tilt. This data could be monitored real-time. The manufacturer could use that data to tighten their process, and continuously monitor the variables as they output product.”

Despite the fact that wire bond is a mature technology, a failure is hard to predict. PDF’s Huntley noted that wire bonding equipment vendors also are developing equipment monitoring capabilities for the voltage and current waveforms on the power supply for each bond. These waveforms can be analyzed to detect anomalies that will nonetheless pass final test (see related video).

If there are 200 packages per strip and each package has 100 wire bonds and each wire has 2 bonds, that results in 40,000 waveforms for just 200 packages. It is a significant challenge to collect this data and analyze it in a meaningful time period prior to final test and shipment to the customer. This requires real-time data collection coupled with big data analytics.

Seeing defects
Key to providing feedback for assembly process steps is finding defects close to their formation. Electrical test occurs at several steps after alignment and bonding, as noted earlier, but does not detect all package-induced defects.

That’s where image inspection comes in. Inspection, which is used for the die, the package, and the handling process, typically has been done at the end of assembly or final test.

“Package inspection typically occurs two times. First, after the assembly of final packages, before final test, and, a second time, once package testing is completed prior to shipping out the devices,” said Dupont. “Inspection of bare dies can also be performed before package assembling. The ICOS F160 inspects bare dies after dicing to check for cracks before sending the dies to subsequent assembly process steps.”

Inspection is essential for for advanced packaging, but obviously not everything can be inspected after assembly. “For mid-end/advanced packaging applications, inspection and measurements can be conducted for a wide range of applications including gold bumps, solder balls and bumps, wafer bumps, copper pillars, among others for critical packaging features, including bump height, co-planarity, diameter and shape, relative location, and variety of other measurements,” said CyberOptics’ Skunes.

Until recently, optical inspection or X-ray inspection were used to assess flip-chip bond quality. Using techniques found commonly in failure analysis labs, these are time-consuming and not always comprehensive in bond coverage. On top of that, customers now are demanding higher yield and product quality.

“This leads manufacturers to require 100% inspection for wafer-level and advanced packaging, rather than sampling strategies, which puts a premium on inspection speed/throughput,” said Skunes. “Slow processes have led to these sampling strategies. But 100% inspection and metrology, with high resolution and high accuracy, can be conducted today — two to three times faster than alternative solutions. This Multiple-Reflection Suppression (MRS) sensor technology can quickly and effectively measure shiny and mirror-like surfaces – critical for highly accurate measurements.”

Speed is improving for all of these tools. “SVXR tools uses a single top-down X-ray image, and it does about 3,000mm² per minute,” said Jewler. “That’s quick. Yet it also requires a higher dynamic range to precisely detect small changes in the thickness of the solder joints. On today’s large devices we’re talking about hundreds of thousands of solder joints, and the images are not really of any use to anybody. To make the technology deployable in the field we invested in computer vision algorithms and machine learning. We developed the capability to capture images and process the images and collect data around certain features of the solder bumps. This enables outlier detection.”

The use of machine learning on images seems inevitable.

“Packaging is becoming very complicated, and electrical testing is not going to be the answer,” said Doug Elder, general manager of OptimalPlus, an NI company. “Machine vision can be used to capture images of assembly-related items. One can now digitize these images, run an ML algorithm on them to determine what is a good image and a bad image, and use this in-line and in real-time for various machine vision-related steps in your back-end processes.”

With a 100% inspection capability there’s higher coverage. Yet inspection throughout the assembly factories has not been deployed everywhere in the process.

“It really depends on the criticality of the process step,” said KLA’s Dupont. “For the most critical, the requirement may be 100% inspection. But for other steps it may be sufficient to check the quality of the process with a sample of devices.”

With faster imaging capability enabling 100% inspection to support production speeds this allows more inspection steps to be placed prior to final inspections. Dupont noted that there’s motivation to inspect before final test especially for more complex packaging technology. Incoming die/wafer inspection has become more common. For flip-chip inspecting right after bonding where the defects are formed can provide a lot of value.

Customers have shifted their expectations on assembly factory quality and reliability. With package failures representing about 50% of the field failures, assembly factories will need to pick up their game. This only can happen with an investment in monitoring of manufacturing processes and product.

Suppliers of equipment, sensors, inspection capabilities, and analytics all have a role to play. They are stepping up their efforts with increased equipment data collection and faster inspection to enable 100% product monitoring, as well as data analytics for analyzing the big data that now can be collected.

“We see that there is a big potential for developing, and improving the quality of the process by adding more inspection steps to enable more process control, but there is a balance between cost and usage by the customer,” said Dupont.

Still, continuous monitoring represents a major mindset change for assembly factory general managers, who make the decisions on factory improvements. Existing equipment cannot support the increased monitoring. It requires upgrades. Likewise, data management systems will need improvements as factories shift from minimal data collection to collecting data on all parts for all the connections, inspecting at multiple insertion points, and collecting data from assembly equipment.

“For 20 years the value or revenue contribution from assembly has been trending up,” said PDF’s Huntley. “With heterogeneous devices requiring multiple die-lets, then putting them together in assembly, it becomes even a bigger part of the value proposition for the final device. It just makes sense to spend more money in assembly steps.”

Related Stories

Using Fab Sensors To Reduce Auto Defects

The Need For Traceability In Auto Chips

Advanced Packaging Makes Testing More Complex

The Next Advanced Packages

Panel Fan-out Ramps, Challenges Remain

The Next Advanced Packages

Planning For Panel-Level Fan-out

Leave a Reply

(Note: This name will be displayed publicly)