Machine vision aids IC wire bonding

Machine vision systems are crucial in today's microlelectronics assembly operations, such as wire bonding.

As ICs continue to become more sophisticated and minimum feature sizes keep shrinking, the need for higher accuracy in fine pitch interconnect applications has become more important than ever.

Machines simply have less room for error, and all subsystems in the machine need to work at a higher degree of accuracy.

Machine vision systems, which play an important role in enabling precise wire bonding by automatically locating proper bond points and the position of lead tips, face some extraordinary challenges with respect to today's fine pitch devices.

For example, a vision system must be able to locate bond pads on a die with extremely high accuracy even though there is often less visual information to work with.

It is not uncommon, for example, to see pad pitch sizes as small as 50um on some of the devices used in today's telecomms products.

Even if a vision system offers the necessary accuracy to handle such miniscule geometries, however, it must be robust enough to withstand a number of variations.

For example, if a die rotates in a waffle pack, the vision system must recognise the die even though its orientation varies from what the system was trained to find.

A vision system also typically needs to contend with non-linear lighting variations.

Depending on how light hits the surface of a die, the vision system may see a die pattern, or an entirely washed out scene due to reflection.

Reflections can also affect the appearance of the bond pads, which the vision system uses as reference points to measure from.

Finally, probing can cause scuff-like marks on the bond pads, which may in turn create a visually confusing scene for a vision system.

To better ensure pattern recognition robustness of its new 3088iP wire bonder, ESEC, a leading global provider of chip assembly equipment, has equipped the new machine with a Cognex vision system.

The primary goal of using the system, according to ESEC project leading Andre Muff, was to improve pattern recognition robustness.

"In our older Wire Bonders, we had been using an in-house developed vision system", he explains.

"While the system had worked well for several years, it is no longer up to date to meet future demands.

In fact, we stopped development of the wire bonder PRS system at a point where we had reached certain limits.

For example, the system had trouble locating die that had rotated more than 7 degrees".

Having successfully used in-house vision system in previous wire bonding equipment, ESEC chose to equip the 3088iP with a Cognex MVS-82400/VME embedded vision system, which features an on-board MMX processor and a suite of specialised vision software tools.

One of the primary advantages of the system, according to Muff, is that many of its software tools are based on PatMax geometric pattern matching technology, which he claims provides greater robustness and accuracy than the previous system.

"With PatMax, we now have support for full range rotation of a chip, and it even allows us to locate patterns reliably when a chip is tilted on the lead frame".

Traditional vision technologies have relied upon a technique known as grayscale normalised correlation for pattern matching, which makes a grid-like comparison of an image of an actual pattern acquired by the vision system with a reference image of the same pattern stored in PC memory.

The x-y position at which the two images best match up is calculated, and from this, the pattern's location is determined.

Though effective in certain applications, this approach limits a vision system's ability to accurately locate patterns that vary in angle, optical scale, or contrast.

Add in other problems such as non-linear lighting variations, reflections across the surface of a device, and stray marks caused by probing, and correlation methods typically fail altogether.

Geometric pattern matching technology, in contrast, uses geometric information in place of pixel grid-based correlation.

For example, it may interpret a square as four line segments and a football as two arcs.

It does this by applying a three-step geometric measurement process to a pattern.

It first identifies and isolates the key individual features within the pattern and measures characteristics such as shape, dimensions, angle, arcs, and shading.

Then, it corresponds the spatial relationships between the key features of the reference image to the runtime image, encompassing both distance and relative angle.

Finally, by analysing the geometric information from both the features and spatial relationships, the pattern's x, y and theta positional coordinates can be determined.

In doing so, geometric pattern matching enables ESEC's 3088iP to locate bond and lead patterns despite changes in die angle or optical size, or the effects of stray lighting and probe mark clutter.

Additionally, it can provide location accuracy to 1/40th of a pixel despite these problems, whereas normalised correlation approaches offer only about 1/4 to 1/10 pixel accuracy.

"This type of accuracy is especially important when dealing with ultra fine pad pitches of many of today's components", says Muff.

The high accuracy also helped ESEC improve the throughput of the new Wire Bonder.

According to him, the machine has a new optical system that enables customers to capture more lead frames and pad sites in a single field of view.

"In order to improve our rate of units per hour we are now using a larger field of view.

The Cognex system has the computational power to locate and process more leads and pads in a single field of view, which means that we have reduced the number of pictures we need to take to cover an entire device".

Muff states that units per hour rates have improved up to 40% for low lead count devices, and up to 12% for those with high lead counts.

In addition to improving the robustness and throughput of the 3088iP, the vision system ties in with ESEC's "process portability" concept, which basically means that the same process can be used on the same type of machine.

In machine vision terms, this means that an operator can train reference images on one machine and use the resulting models on all other wire bonders without having to retrain on each one individually.

"The reference images the vision system is trained on become part of the overall recipe", explains Muff.

"And being able to teach a recipe on one wire bonder and use it on others makes things much easier to handle and manage for our customers.

They are able to distribute recipes either offline by using a floppy disk or online by using the host".

Back to top