Detection and prevention of assembly defects, by machine learning algorithms, in semiconductor industry for automotive

Years of experience within semiconductor manufacturing facilities have led to optimize processes to serve both quality and cost. The solution to achieve next generational levels requires a new approach: this one is fitting with implementation of advanced analytics and machine learning algorithms. Applied to manufacturing data which corresponds with a real big data context, these algorithms can provide insights and automate responses to detect, prevent and ultimately eliminate the most severe failure modes. The project described in this paper targets a wafer sawing process. Various challenges that are raised in such a project are of different natures. A first one is the need for a high level of technical expertise in the manufacturing process of focus: this is essential to define the meaningful dataset that represents comprehensively the desired output of the process. Another component is the data collection aspect: many data have to be collected, stored and parsed, and some small signals found will become the leading indicator to an upcoming process degradation and capability of capturing them is essential. Another key data is traceability of the processed material. Additionally, ensuring an informatic technology architecture to support collection, storage, parsing and computation of the datasets is a significant challenge. Lastly, project success is related to the data scientist expertise to build adequate machine learning algorithms. Optimization of the models can take several iterations with back and forth communication between data scientists and process technical experts. This paper describes issues revealed, some solutions found, and future expectations.