The Business Problem
The manufacturing process of memory chips involves around 1,500 steps that need to be performed in sterile conditions to avoid specks of dust from damaging the wafers. However, damages occur nonetheless. The quality issues that arise, scratches, holes etc., are often microscopic and near-invisible to the human eye!
The manufacturing environment is home to a multitude of machines, pipes and parts. These wear out, break down or start dripping. Detecting these issues in an early stage is crucial. Engineers are usually responsible for maintenance. However, even the most highly skilled engineer can miss early indicators that something is wrong.
Inherently, the process of manufacturing memory chips has a lot of potential for error. Relying on human vigilance to identify quality issues and mechanical problems was costing Micron Technology a lot of money, on average $250,000 per hour of downtime (Micron Technology).
This particular business problem is well-suited for AI solutions. The problems are clearly defined, measurable and there is enough in-house data to use Machine Learning (ML) on multiple fronts with good accuracy. The solutions also work with smaller volumes of data, but the accuracy of the ML algorithm will not be as good. As more data is gathered, accuracy will improve.
Another major memory chip manufacturer (Intel) also implemented Machine Vision and Machine Learning algorithms in its wafer production process. A whitepaper on their approach states the following interesting conclusion:
“Similar technology can be used in many different industries-wherever machines capture images, regardless of the original use for those images.”
The ML algorithms are designed to detect anomalies in an earlier stage, with higher precision and frequency than its human counterparts. However, and this cannot be stressed enough, humans are still needed to interpret and act upon the alarms given by the system.
Micron Technology implemented Machine Vision technology into its photolithographic cameras as they etch the circuitry into the wafers. The technology scans for frequently occurring flaws and alerts the engineers whenever a flaw has been detected. Depending on the type of flaw, it takes somewhere between 15 seconds and 15 minutes before the alert is given.
The company’s Auto-Defect-Classification system (ADC) solves the problem of classifying every defect manually. The system makes use of deep learning (Opperman, 2019) to sort and categorise millions of flaws. A whitepaper by Intel explains the ADC system more in-depth.
To further drive the effectiveness and accuracy of AI, Micron implemented thermal imaging to monitor its manufacturing process. “Heat maps” of the factory environment under normal working conditions are overlayed onto a ‘digital twin’, essentially a digital copy of the factory environment. This map then provides a baseline to compare real-time infrared imagery of the factory. If the system spots an anomaly, i.e. irregular temperatures compared to the digital twin, the system sounds alarm.
Possibly the most surprising of the trio, an AI solution has been created to identify unusual noises within the manufacturing process. Similar to your car producing odd sounds, a machine making an unusual sound often indicates trouble. The AI system at Micron has been trained to spot irregularities in sound frequencies by converting sound to visual datapoints. To capture the sounds of individual machinery in a loud environment, audial sensors are placed close to machinery or pumps. Categorisation of sounds and potential causes is done by the engineers.
Benefits & Impact
Firstly, Micron Technology’s AI solutions have increased manufacturing efficiency and accuracy noticeably. Secondly, Worker safety has improved (Workers come into contact with extreme temperatures and harmful substances less frequently). Thirdly, AI solutions have freed up valuable time for the company’s engineers to focus their efforts elsewhere. Lastly, the implementation of AI solutions in the manufacturing process has spread out to other processes in the business, such as product demand forecasts, Increasing the accuracy by 10 to 20 %.
- 10% increase in manufacturing output
- 35% less quality issues
- 25% faster time to yield maturity
- Avoided millions of USD through early detection of machine breakdowns and quality issues
- Freed up time for engineers to focus their efforts elsewhere
- Increased worker safety
- Paves the way for AI solutions in other business processes
Accessibility and Requirements
The driving force behind the three AI solutions implemented at Micron Technology is data, lots and lots of data. The company collects petabytes of manufacturing data from over 8.000 sources and over 500 servers worldwide. This data is send to two environments of the open-source software program “Apache Hadoop” for data mining. Hadoop is designed for parallel processing of large data-sets, meaning that multiple datasets can be processed at the same time.
For Machine Vision over 2.000.000 images are stored in the Hadoop environment. For Acoustic Listening, Micron sends the relevant data to a GPU system to handle the massive workload of the complex machine learning algorithm in a swift manner. GPU’s can continue to accelerate applications by dividing tasks among many processers, allowing it to process the vast amount of data that is poured into the system.
Requirements for AI solutions
- Sufficient images and data on flaws, ‘normal’ sound frequencies and temperatures
- Deep learning algorithm (for classification of flaws for Machine Vision)
- A digital twin (for Thermal Imaging)
- Infrared cameras (for Thermal Imaging)
- Audial sensors (for Acoustic Listening)
- Apache Hadoop for data mining
- GPU systems
- Machine Learning algorithms
More AI Case Studies:
AI case study 5: Workflow automation with TradeCloud One
The Business Problem In this business case, AI workflow automation will be exemplified using the example of a TradeCloud client. The client is a manufacturer of machines for the sorting of eggs and [...]
AI case study 4: Machine Learning in the Manufacturing Process
The Business Problem The manufacturing process of memory chips involves around 1,500 steps that need to be performed in sterile conditions to avoid specks of dust from damaging the wafers. However, damages occur [...]
AI case study 3: Cost-saving AI in Manufacturing Logistics
The Business Problem In the world of logistics, empty fleet management cannot be overlooked. For container shipping, the Boston Consulting Group estimates that up to 8% of a shipping line's operating costs relate [...]
AI case study 2: Efficient inventory management using Artificial Intelligence
The Business Problem A major challenge for manufacturing companies is to know what, when, where and how much stock should be ordered and kept. SME’s traditionally calculate this manually using Excel, Google Sheets [...]
AI case study 1: Demand Forecasting using Artificial Intelligence
The Business Problem One of the biggest challenges for business executives today is demand volatility in relation to demand forecasting. Whereas data availability continues to increase, customer purchase patterns are becoming increasingly complex, and [...]
Alexsoft. (2019, November 11). Demand Forecasting Methods: Using Machine Learning and Predictive Analytics to See the Future of Sales. Retrieved April 6, 2021, from Alexsoft
Armstrong, M. M. (2020, December 4). Cheat sheet: What is Digital Twin? Retrieved April 21, 2021, from ibm.com
BCG. (2015, November 17). Think Outside Your Boxes: Solving the Global Container-Repositioning Puzzle . Retrieved from Boston Consulting Group
Best Practice AI. (n.d.). Danone reduces forecast error and lost sales by 20 and 30 percent respectively and achieve a 10 point ROI improvement in promotions with machine learning. Retrieved April 6, 2021, from Bestpractice
Brosset, P., Patsko, S., Khadikar, A., Thieullent, A., Buvat, J., Khemka, Y., & Jain, A. (n.d.). Scaling AI in Manufacturing Operations. Retrieved April 06, 2021, from Capgemini
Caulfield, B. (2019, December 16). What’s the Difference Between a CPU and a GPU? Retrieved April 22, 2021, from nvidia.com
Dilmegani, C. (2021, January 7). Demand forecasting in the age of AI & machine learning . Retrieved April 6, 2021, from AImultiple
Europawire. (2020, April 15). Shipment planning and inventory management improved with AI and big data on Kuehne + Nagel’s new version of SeaExplorer. Retrieved from Europawire
Hyeongchan, K. (2020, December 16). Detecting Sounds with Deep Learning. Retrieved April 21, 2021, from towardsdatascience.com
Kvartalnyi, N. (2021, May 11). 6 TIPS OF HOW TO IMPROVE INVENTORY MANAGEMENT USING ARTIFICIAL INTELLIGENCE. Retrieved from Inoxoft
Majumdar, D. (n.d.). Case Study-How SynergyLabs AI solutions Brought Efficiency in warehouse Inventory management. Retrieved April 19, 2021, from Synlabs
Micron Technology. (2021). Case Study: Micron Uses Data and Artificial Intelligence to See, Hear and Feel. Retrieved April 21, 2021, from sg.micron.com
Opperman, A. (2019, November 19). What is Deep Learning and How does it work? Retrieved April 21, 2021, from towardsdatascience.com
Serheichuk, N. (2020, December 15). Inventory management automation: How you can benefit from it. Retrieved from N-ix
Subramanyam, J. (2019, July 8). Prediction Models: Traditional versus Machine Learning. Retrieved May 20, 2021, from Gartner.com
Supply Chain 247. (2017, March 28). Coca-Cola Leverages AI for Inventory Management. Retrieved from SupplyChain247
Supply Chain Dive. (2018, August 17). Two-thirds of companies consider Excel a supply chain system. Retrieved from Supply Chain Dive
Symphony Retail. (n.d.). demand forecasting ai. Retrieved April 06, 2021, from symphonyretail
TradeGecko. (2019, December 4). What is inventory management? Retrieved from tradegecko
Tuv, E., Murat, G., Enis, P., & Lee, D. H. (2018, November). Faster, More Accurate Defect Classification Using Machine Vision. Retrieved April 21, 2021, from Intel.com