Case Study: Semiconductor Manufacturing.

Welcome to my blog series on examining how machine learning (ML) can supplement the analytical capabilities of humans in various vertical applications. For this post, we’ll be examining how ML can benefit manufacturing, and what better use case to look at than the most advanced manufacturing environment in the world, a semiconductor fab?

A quick summation of the semiconductor manufacturing process may be in order. The individual semiconductors or chips you’ve seen on motherboards in PCs or other electronic devices are created in large, extremely expensive and complicated factories called foundries or fabs. From start to finish, the process to create a chip takes weeks; just how many weeks depends on the size and sophistication of a specific chip. For example, semiconductors made using today’s most advanced manufacturing processes can take up to 15 weeks to manufacture.

The process begins with a large ingot of a particular material (most commonly silicon) that is sliced into very thin layers known as wafers. A wafer then moves through a series of complicated chemical and photo lithographic processes designed to “print” the millions of individual transistors that make up a single chip onto the wafer. Each step in this process is referred to as a layer. A wafer may go through up to 11 different layer processes before it is complete. The end result is a wafer that has hundreds of copies of the same chip printed on it. After some Q&A testing, the wafer is then cut very precisely into individual chips which are put into packages (small plastic or ceramic boxes that surround each chip. The chip package is what you actually see when looking at chips on a motherboard). This is a highly simplified explanation of the process, but GLOBALFOUNDRIES, a world leader in semiconductor manufacturing, has an excellent video here that explains it in more detail.

Because the transistors that make up semiconductor chips are so small (millions of them could fit on the head of a pin), it is vital that the manufacturing process be conducted in very clean and controlled environments filled with highly sensitive (not to mention very expensive) manufacturing equipment. Just one mote of dust getting into the semiconductor production process can destroy a wafer or damage a piece of equipment, resulting in thousands of man hours and hundreds of thousands to millions of dollars lost. To avoid this, semiconductor manufacturing lines are housed in clean rooms where precise control is kept over environmental pollutant levels, temperature, humidity and vibration to ensure that each wafer’s exposure to potential damage is minimized. Furthermore, wafers are often monitored via video as they move along the production line in the hope that problems are identified quickly so flawed wafers can be spotted and removed from the manufacturing process before more time and resources are wasted on them.

This is where ML can really pay dividends for semiconductor manufacturing. Following one wafer through weeks of manufacturing generates lots of video; a modern fab can generate upwards of two terabytes of video data every day. But by analyzing that video and other sensor data captured from the manufacturing line, semiconductor fabs could move beyond simply finding and replacing a broken or contaminated machine and repairing to minimize down time. They could actually predict when a machine is likely to fail and take steps to fix it without stopping production; a process known as predictive maintenance.

Unlike other manufacturers, semiconductor fabs aren’t aren’t trying to collect as much data as possible to make their analysis more accurate in predicting breakdowns. After decades of monitoring their processes to improve efficiencies, most large fabs already have plenty of data available for analysis. Their problem is one of scale. How do you combine decades of legacy manufacturing data with the terabytes of new data the line generates every day and still be able to analyze it in real time? Only a computer equipped with the necessary machine learning algorithms, written by humans leveraging their semiconductor manufacturing expertise to anticipate production line issues, has the processing speed required to sort through databases as large as a semiconductor fab’s and identify machines that will require fixing. And only a computer capable of performing millions of calculations a second can the apply that analysis in real time, determine the proper course of action and then issue a command to fix a machine without the need for human intervention.