Artificial intelligence in semiconductor production at OSRAM

The OSRAM Podcast: Episode #7 with Dr. Maike Stern & Dr. Jörg Schäfer about artificial neural networks


Welcome to the Photonstudio, the OSRAM Podcast. My name is Karin Steinmetzer and I would like to welcome you today to a very exciting topic. With me in Regensburg at OSRAM Opto Semiconductors are data scientist Dr. Maike Stern and production manager Dr. Jörg Schäfer.

I would like to find out from them what artificial intelligence can achieve in semiconductor manufacturing. Because a lot of data is generated during the production of an LED. What would happen if machines were able to extract information from this data independently? I'd like to find out what projects are currently in progress in production and whether self-learning machines could one day take over the work of our engineers.

Karin: Before we dive into data science with Maike, I have the first question for Jörg: why do we need artificial intelligence in our semiconductor production?

Jörg: There are very complex process sequences in semiconductor and chip manufacturing. Understanding the interaction of these processes in their entirety is a great challenge. An artificial intelligence that takes up these interrelationships and develops solutions would be an enormous relief in everyday production. These are data volumes and interrelationships that are very difficult for humans to understand. Because of the amount of data and because of the interdependencies. I believe that AI would be a great benefit for us in everyday production.

Karin: The first step towards making everyday production easier with methods such as artificial intelligence was taken by Maike. You taught computers to interpret images of wafers, i.e. the semiconductor wafers on which the chips are located. How does a computer learn to understand images? It not only lacks expert knowledge, but also the human visual experience and the ability to classify image content.

Maike: This can be explained best with an example. And not by means of an LED wafer, because no one will probably have that in mind now, but by means of dogs and cats. If we teach a computer how to distinguish between dog and cat images, you can imagine how we solve this problem ourselves in our heads. When we think of a cat, it has completely different characteristics than a dog, for example. That means the shape of the ears is different, the muzzle of the animal is different and maybe also the color of the fur. And this is exactly what we take advantage of when we teach the computer to distinguish these pictures. If we now think of a classic algorithm, we would use different filters to highlight certain features in the image. And if we now use different filters, we can emphasize very different features in the images, for example the different ear shapes of the animals. And if we then collect these features, we can open a case. Then we have more evidence that there is a dog in the picture or that I have a cat in the picture. And so computers learn to recognize pictures. That means we give them a task, use different algorithms to collect features and then we get a classification of the image.

Karin: What does this mean transferred to images of wafers?

Maike: Here we have LED chips that are okay and we have other LED chips that are probably defective. And these two classes also have different characteristics. We extract these characteristics from the images again to get a classification.

Karin: The goal is that the algorithm can also interpret new images independently. How does it work that it learns to apply this to new images?

Maike: If we want the algorithm to be able to recognize new images, we assume that the features we have learned are transferable to new images. And we now assume that new defects have similar characteristics. This means that the algorithm also transfers its knowledge to new examples. But this also explains why these deep learning or machine learning algorithms need so much data. Because the more data I give the algorithm, the better it knows the data and the better it can then generalize.

Karin: I would like to pick up the keyword data and ask Jörg something. Machine learning applications, i.e. artificial intelligence, need mainly three things: algorithms, enormous computing capacity and a lot of data in good quality. Do we have the latter in our production and if not, how do we get it?

Jörg: The amount of data available to us is growing. The systems and technologies we use are increasingly equipped with sensors. This means that the amount of data generated is constantly increasing, which enables us to use this artificial intelligence more effectively. Where we certainly still have a lot of work to do is on data quality. Various projects are currently underway in our production to ensure that the data quality is available in sufficient quantity.

Karin: Maike, could you perhaps add a few words from your own perspective? From the perspective of the person who has to process this data in intelligent algorithms.

Maike: I can only agree, there is an incredible amount going on in our company in terms of data quality and quantity. But it is also the case that these algorithms are very data-hungry. But we can get a lot out of the algorithm even with only a few data. You always think that with artificial intelligence the programmer doesn't have to put in any work at all. But the work is just shifted. While I have to think a lot about how I present the data to the algorithm, I also have to think very carefully about how I build the algorithm. Because there is an incredible amount of knowledge in the structure of the algorithm. This is also shown by the fact that there are very different algorithms. Algorithms for speech processing look completely different from algorithms for image processing. If we now have data sets that are rather small, we can use the knowledge we have gained from working with the experts to better design the algorithm to solve the problem with more prior knowledge. In this way we can compensate for these small data sets to a certain extent.

Karin: So back to Jörg: where do you see the main fields of application for machine learning?

Jörg: There is a concrete example that is already being worked on. This is our value chain from epitaxy to the finished product. Here we have many process steps that determine and influence the wavelength, brightness and color of the product. To represent and understand these interrelationships across the entire value chain, an incredible amount of data is required. I see a chance to make significant progress with artificial intelligence. Not only to understand what results are available today, but also to get suggestions on how this could be better synchronized. For me, this is a concrete example that is already being worked on. Surely it will take a while before you get results there.

Karin: Jörg, Maike looks a bit critically, is that more of a distant future?

Jörg: For me, artificial intelligence is perspective from a production point of view anyway, but we also have to tackle these perspective topics. In addition to this topic, there are also other topics such as improving logistical processes and generating an understanding of the process. But this is not only about having the AI do the data processing, but also about generating suggestions. In an idealized world, these are then submitted to the engineer, who can look at them and decide whether to accept the proposal or not. Also, preventive maintenance is a topic that is very interesting. It will certainly not be ready for discussion in the next 1-2 years, but if we don't start moving towards it now, we won't be ready in 4-5 years. But by then we must be ready to improve our competitiveness.

Karin: Maike, are you already working on it?

Maike: We are definitely already working on it. There are colleagues in the company at very different places who are developing very sophisticated algorithms. And such a system, as Jörg just described, consists of many different algorithms with different tasks.

Karin: Perhaps a little more concrete about the possibilities. What do intelligent algorithms make possible in our LED production that we cannot achieve with conventional methods? Where can we find completely new solutions?

Maike: To find completely new solutions, we use a methodology called Reinforcement Learning. This is an algorithm that learns to find its way in the world by itself. We are currently using this algorithm to come up with new LED chip designs. This means that when we design a chip, there are different materials that we apply to each other in different thicknesses. On the one hand, we can vary how many layers we want to apply, with which material and how thick these layers should be. This gives us an infinite number of possibilities how this chip will look like in the end. Reinforcement Learning gives us the possibility, in contrast to classical or genetic algorithms, to scan the search space more intelligently, so that we can find a better solution much faster. Furthermore, it is classical for these algorithms to look at unknown areas. This allows us to discover chip designs that we would not have thought of before.

Karin: Is artificial intelligence then always the best solution or are there also cases in which we still make progress with classical methods? Where does artificial intelligence really make sense?

Maike: In our company there are many examples where it makes much more sense not to use machine learning methods. The nice thing is that a lot of programmers in my group work with classical methods and have a tremendous amount of knowledge. So we can see if our algorithms are really better. Basically, it always makes sense to use machine learning or deep learning methods if the data contains a lot of variations and we can't get any further with classical methods.

Karin: First of all, thanks to both of you for the insight into where AI can change our production. Finally, a question for Jörg: does it mean that one day machines will be so independent that our colleagues in production will be redundant?

Jörg: Clear answer: No! Over the last centuries we have seen again and again that a job description is shifting. The contents of certain activities are shifting and artificial intelligence is exactly the same for me. The activities of engineers, for example, will shift. They will be partially facilitated, but this will result in new tasks and problems that have to be handled by humans. The computer supports and helps. Humans will not become superfluous!

Karin: Many thanks to both of you for taking the time for this exciting conversation!

Maike: Thank you!

Jörg: Thank you, too!

As always, you can listen to the latest episode of our Photonstudio on Soundcloud, iTunes, Spotify and Google Podcast. If you want to know more about the use of Machine Learning at OSRAM, I recommend the online version of our innovation magazine ON at www.osram-group.com/innovation. You will also find many other exciting articles from the world of photonics. See you at the next episode of the Photonstudio!

Photonics is THE key technology in every OSRAM solution