semiconductor shortage

Harnessing machine learning to ease semiconductor shortage

Since spring 2020, there has been a global shortage of semiconductors. With demand currently far outpacing supply, the need to maximise throughput and minimise waste has never been more acute. Of course, improvements in these areas also have positive effects on sustainability. Elisa IndustrIQ is at the forefront of the drive to boost semiconductor manufacturing, enlisting machine learning to spot manufacturing deviations to improve yield and prevent standstills. This article introduces two examples.

Worldwide semiconductor shortage continuing

The worldwide semiconductor shortage that emerged in 2020 due to the global coronavirus pandemic looks set to continue into 2021. Demand in the auto industry is still huge, not least because the industry is in the middle of a phase shift towards electric vehicles, increasing demand for chips even further. The shocks to production caused by the global pandemic are also making themselves felt as shortages of some consumer goods.

This means it is more important than ever for semiconductor manufacturers to minimise waste and maximise throughput in their fabrication plants.

Semiconductor manufacturing requires extreme precision

Manufacturing semiconductors are extremely expensive and complex. The fabrication plants (“fabs”) cost billions to construct, featuring “clean room” environmental controls that remove every speck of dust as well as damping to eliminate all vibrations. The manufacturing processes can involve hundreds of discrete process steps repeated dozens of times to create chips on a silicon wafer. It often takes as long as three months to “bake” a chip, and after that, it can be months more before it finds its way into the end product. Flaws or defects at any stage in the process can result in an unacceptable waste of resources, time, and money.

Modern semiconductor designs use 3D geometry, with hundreds of layers grown one by one on a semiconductor wafer. Each new layer needs to be laid on a surface that is as smooth and uniform as possible, and the tolerances involved are almost impossibly strict. A method called chemical mechanical processing smooths the surface of each layer, removing material to even out any irregular topography to keep the wafer as flat as possible. Deviations of more than a few nanometres can result in a faulty batch of wafers being produced. (For comparison, a human red blood cell is about 7,000 nanometres wide, and a human hair is close to 100,000 nanometres!)

In a typical semiconductor fab running a mature process, such as 110-nanometre production processes, the yields are very high – around 98%, with only about 2% of a wafer going to scrap. But for less-mature processes, such as 7 nm processes or the latest 2 nm processes, the yields can be as low as 50%, with around 10% going to scrap, and it can take years until these new processes are properly mature. In financial terms, these losses were equivalent to a global financial impact of around $110 billion every year even before the coronavirus pandemic began. The financial impact has only grown since then.

One of the key methods of monitoring the process is the use of “wafer maps,” which are images that illustrate how each layer on the wafer builds up over time and how smooth the surface of the layer is. These are created from precise measurements taken at specific points on the surface of the wafer. These measurements are then converted into an image that helps people to visualise the topography of the wafer and how much the surface deviates from the ideal. Naturally, wafer maps are unique to each specific chip design, which means there is always a considerable effort of analysis involved for each design.

Elisa’s solution

Elisa IndustrIQ spotted the opportunity to step in and provide semiconductor manufacturers with a solution to this problem. Some years ago, Elisa had already begun focusing on how it could utilise artificial intelligence and machine learning to improve its own telecommunication operations, particularly through automation and predictive maintenance. Elisa has now set its sights on taking what it has learned and bringing those advantages to a range of other industries. To accomplish this, Elisa has been taking on leading industry-specific expertise in the field, such as with the acquisition of camLine, a specialist in manufacturing control software solutions.

In the case of semiconductor manufacturing, Elisa’s experts are applying machine learning techniques to create a system that can not only extrapolate from any set of measurement points to create wafer maps comparable across different measurement site geometries but can also help in using historical process data to find known root causes and identify corrective actions.

Advanced analytics in semicon

 

The system transforms the raw data into a handful of parameters – known as a “feature vector” in the world of machine learning – that can describe the surface of the wafer regardless of the specific measurement points used.

Nils Knoblauch, Product Manager at camLine, explains: “The system utilises a dataset of wafer map measurement points to build two models. One is trained only with known good wafers and is used to compute an ‘anomaly score’ to detect future wafers with unusual shapes. This anomaly score improves on typical methods of detecting deviations since it can distinguish convex shapes from concave ones, for example.”

Advanced analytics graph

“The second model is trained with the complete set of wafers from the past – including the anomalous ones – to build up a knowledge base about the process history,” says Knoblauch. “If a future wafer is detected as having an anomalous shape, the system searches this knowledgebase for the most similar wafers and presents a list of the most wafers that are most similar in shape to the current one. The user can then look up how the problem was solved in the past to return to normal processing conditions.”

Benefits of the solution

This ability means that the Elisa IndustrIQ wafer profile ML solution helps semiconductor foundries to retrieve existing process experience faster and resolve “out of control” events more quickly. And the time saved in doing this immediately increases the throughput of the fab because the processing tool that produced the anomalous wafer is usually at a standstill until the problem is resolved. There is too much value at risk to continue production when there is even the slightest indication of a persistent problem that could also spoil the next production batch. The use of cutting-edge ML techniques to boost throughput will help to ease the current semiconductor shortage.

The system can be trained with a relatively small dataset. Typically, deep learning models would require information about thousands of manufactured wafers per design to learn effectively. But the Elisa IndustrIQ system requires a dataset of only a few hundred wafers from any mix of product designs to be adequately trained, greatly expanding its usefulness.

camLine’s flagship product – LineWorks SPACE – is now integrated into the Elisa IndustrIQ wafer profile ML solution, allowing engineers to have easy, real-time detection of anomalous wafers and to review these anomalous wafers online.

wafer map analysis

Through careful analysis of wafer maps along with other process data, the integrated solution may also be able to spot potential quality issues as they develop, allowing manufacturers to take action to remedy the problem before any faulty wafers are produced, either by repairing or replacing the machinery in question or by identifying an alternative production path.

Using this solution can reduce wastage in semiconductor fabs, as well as increase tool utilisation, and improve the yield of every wafer manufactured. The effect on the bottom line is truly significant, and it can help ease bottlenecks in production.

Looking to the future

This solution is already available, and even more exciting applications are already on the horizon. One of the future opportunities lies in statistical process control (SPC), a quality control method familiar to all manufacturing industries. It uses statistical analysis of data produced in the manufacturing process to produce “control charts” that help engineers spot products that fall outside acceptable boundaries of quality. It has several advantages over older quality control methods, such as inspection, in that it can help spot problems early and even prevent them, rather than simply reacting to problems after they have occurred.

However, in a manufacturing process with as many complex steps as semiconductor manufacturing, SPC has severe limitations, and the control charts it produces can’t tell you everything you need to know about the process. While vast amounts of process data are produced in semiconductor manufacturing processes, the complexity of the processes means that SPC is of only limited usefulness in identifying issues in the process.

A wafer can be anomalous yet still to fall within the upper and lower control bounds of SPC. SPC wouldn’t usually pick up a wafer like this, yet it could have detrimental effects later in the production process after more layers have been baked onto the wafer. Another problem is that even if SPC identifies an anomalous wafer, it often cannot help resolve the issue, as the root cause of the anomaly may lie many stages earlier in the production process.

That’s where we see an opportunity. Elisa IndustrIQ can leverage expertise in machine learning and apply it to statistical process control – an approach we have dubbed SPC 4.0.

We can use existing control infrastructure and process data to train our model to spot anomalous wafers that fall within the control limits and identify the correlations between anomalous wafers now and the factors much earlier in the manufacturing that resulted in the problem.

Combined with state-of-the-art neural network technology, this solution even has the potential to create an equipment health index to help manufacturers maintain a real-time understanding of the condition of all of their production equipment. It could help identify when conditions exist in specific machinery that may later lead to an anomalous wafer. The system would send an alert that allows the manufacturer to route the wafer through a production path using specific machines to prevent the anomaly from occurring. It could also flag up machines with a poor health index for preventive maintenance to ensure that production remains as efficient and uninterrupted as possible.

These are just a couple of examples of the applications for the industry of advanced artificial intelligence and machine learning techniques. Elisa’s experts are constantly discovering more and more ways of applying advanced techniques to help clients in all sorts of industries overcome their challenges.