DAVID PARK, Vice President of Marketing, Lynceus
The global semiconductor industry has experienced a robust period of growth during the past several years with demand fueled by the repercussions of the COVID-19 pandemic. Semiconductor demand was already strong due to the continuing trend of semiconductors being used in everyday devices from household appliances to cars to consumer electronics. COVID drove demand levels even higher for products that needed semiconductors. Still, the pandemic also resulted in the global semiconductor manufacturing ecosystem to slow down to a crawl for months before returning to normal operations. And the industry has been trying to catch up ever since.
The semiconductor industry has historically been a cyclical business with good times and tough times, and as a result, the industry has become more guarded about overbuilding capacity. As the pandemic unfolded and semiconductor demand ramped up at an unprecedented pace, there wasn’t (and still isn’t) enough capacity to manufacture enough chips to meet the demand. Many foundries and IDMs (integrated device manufacturers) are building new manufacturing plants (fabs), but those will take years to come online, and all existing, mature fabs are already running 24/7. The industry needs to look for additional ways to increase capacity in the near term. One way to achieve this goal is to apply machine learning (ML) to current manufacturing processes.
Industry 4.0 has been a key focus for manufacturers for the past several years, with organizations collecting and storing massive amounts of manufacturing and product data with the goal of analyzing that data to improve key performance indicators (KPIs) such as yield and quality Machine learning is an excellent technology for leveraging all collected tool and product data. It can do so without requiring additional infrastructure that can slow down adoption or drive up the cost of implementation.
ML models can be very beneficial for a semiconductor foundry or IDM by providing real-time predictions of the critical measurements resulting from an individual process step. Process engineers and technicians spend a lot of time monitoring process equipment to ensure they are operating correctly and then sending a percentage of the product to physical metrology tools to validate the process step results. Historically, it is not practical to assess every single wafer at every single process step since the operation of metrology tools costs time and money. The fab must also be confident that the metrology sampling rate can identify excursions or anomalies in their manufacturing process that could impact product yield or quality. Machine learning can change this dynamic by simultaneously reducing the physical metrology sampling rate (saving money and time) while providing complete coverage for every single manufactured part (validating that the process step was performed as expected).
Machine learning can accomplish this by applying “virtual metrology” to any process step, such as etch, lithography or CMP. Assuming the ML algorithm (or “model”) is well-designed and implemented, it can run inline and in real-time with the factory floor’s process equipment. Leveraging the data that is already collected by the fab’s MES (manufacturing execution system) and SPC (statistical process control) systems, ML models have all the data they need readily available to calculate and predict the results for any metrology measurement on the product being manufactured. By focusing on predicting the results of individual process steps, ML models do not need and are not reliant on a complex, big data analytics infrastructure. Properly architected ML models can be agnostic and complementary to any existing environment within a fab, making them relatively fast and easy to implement. The benefit of implementing virtual metrology in this manner is a significant reduction in sampling with physical metrology combined with 100% visibility into the quality of the process step because all the product is now being validated through virtual metrology.
In addition, since the ML model predicts the metrology result for 100% of the product, it is also possible to recognize excursions faster since the ML model runs inline and in real-time with the process step. In the rare cases of a major excursion, the ability to rapidly identify those excursions instead of waiting hours or even days before detection can significantly reduce costs by avoiding the need to scrap bad wafers and the unnecessary manufacturing time and materials spent on a product that was produced prior to the excursion detection. The final benefit of implementing ML models for virtual metrology is reducing in the time, and effort engineers spend in monitoring equipment and finding the root cause of anomalies.
Click here to read the full article in Semiconductor Digest magazine.