Welcome to the machine

Artificial intelligence can improve the processes in semiconductor manufacturing – and open up entirely new chips in the future.

LED production is a highly technological process. And one that produces a mass of data. What if machines could independently gain knowledge from all this? At OSRAM in Regensburg, data scientist Dr. Maike Stern is teaching machines how to see and understand production images. Her work serves as the basis for using machine learning in chip production. Experts agree that only the use of artificial intelligence will ensure competitive production in Europe for the long term.

The starting point for Maike Stern’s work was a set of images of semiconductor wafers containing thousands of chips. But how does a computer learn what these images show and how does it interpret the content?

Artificial image interpreters

Deep learning algorithms are used here, similar to the ones used in Google image searches. Deep learning is an area of machine learning based on a continuous “learning process”. “Many of the algorithms we work with were developed at Google, Facebook or Baidu,” explains Maike Stern. “These companies have huge datasets of images at their disposal. We have to adapt and train our algorithms using much smaller datasets – sometimes as small as 200 images.

For example, Stern has trained an artificial neural network to identify defective chips in photoluminescence images based on the measurements of completed wafers. “A neural network specifically for image processing – a fully convolutional network – learns to recognize rules in the data on the basis of concrete examples. It learns filters specifically designed for the dataset so that it can identify certain patterns and then apply them to new cases. By learning many such filters, a wide variety of patterns can be distinguished which allow conclusions to be drawn about the LED’s properties,” adds Stern. The filters are also concatenated, and the filtered images are successively compressed, enabling the network to recognize even complex patterns in the images.

Finally, the image information is extrapolated back to the original image dimensions in several stages. A prediction can be therefore made for each individual pixel in the image. This means that every faulty chip is marked in the output image.

Versatility

Since this neural network learns a large number of special filters each time it is used, it can also be applied to datasets that contain many special cases. This is precisely where classic image processing methods fail. A version of the algorithm Stern has developed is currently being used to check the quality of special wafer types. It looks for anomalies in the measurements of the finished wafers which would indicate degeneration of the crystal structure. To do that manually would be extremely time-consuming. Automating this analysis accelerates the process, and at lower cost.

The Data Science Team in Regensburg is now using the architecture of the algorithm in other areas, such as the analysis of solder images. This is because the thermodynamic processes involved in soldering harbor many sources of faults, such as air pockets as the solder cools down. Detecting these flaws used to involve experts evaluating X-ray images. Now much of this laborious task can be automated and thus accelerated.

But algorithms can do much more than merely automate and accelerate processes – they can make certain products possible in the first place, or even find completely new solutions.

For example, Dr. Hans Lindberg, a colleague of Maike Stern, uses algorithms to check the quality of ever smaller LED chips: “Data analysis and image processing algorithms allow the measurement of chip types that cannot be measured using classic methods. Our tool combines and evaluates data from pre-processes to identify defects. This can significantly reduce the time and effort spent on measurement.” The previous opto-electrical method in which each individual chip was contacted and measured with a needle is to a large extent no longer needed. In view of the fact that the time spent on measurement increases quadratically as a function of miniaturization, this is a key improvement.

Smart algorithms

The use of artificial intelligence in chip design goes a step further. LEDs can have an incredibly

large number of different designs in terms of material composition and layer thickness ( ~10²⁹ ). Naively testing all conceivable designs to find the best solution would take about five billion times as long as the age of our universe is. Reinforcement learning, a subsection of machine learning, allows computers to systematically search for optima in this vast number of theoretically possible designs. However, accurate understanding of the algorithm and the problem is essential in order to point the algorithm in the direction of development.

By successively checking chip designs the algorithm, which was designed with the assistance of optics experts, independently learns a strategy for finding the best possible chip structure in as few steps as possible. The algorithm receives a “reward” if the physical properties of the design improve. If they get worse, this leads to a “punishment”. The reward is like the grades you get in school. The algorithm gets better grades, or “values”, if it works in the intended direction of development. This way, the algorithm develops designs that lead to the highest possible rewards. It uses its experience but also tries to discover new types of LED design, namely chip designs, that even experienced engineers would not have produced.

Will AI take over?

“Of course, in actual practice there is still further scope for improvements in data quality. We are still a long way from our vision of the “golden route” – the optimum path that a wafer automatically finds through all the stages in its creation process. But with machine learning we are now laying the foundations for competitive quality,” explains Stern. In her job at OSRAM, she is taking a dual approach: being at the forefront of research and at the same time ensuring that machine learning is applied in production.

But what will engineers be doing in the future? Their expertise will still be needed – but their jobs will change. They will be working with algorithms and specifying what they want to know. Artificial intelligence will then find the smartest way to the chip.