Applying Hierarchical Machine Learning Forecast to Manufacturing Process Sequences Topic/Category: Factory of the Future OR Emerging Technologies
It has been demonstrated it is possible to combine design, process, and metrology information to create machine learning models that accurately predict the behavior of individual products in the manufacturing line. In this work, we extend the application of these single process models to a full production sequence. Instead of training a large machine learning model to describe the full production sequence, we sequentially train the models in a way that prior models can be used to impute the sparse manufacturing data of previous process steps. This approach uses process and context-aware methods to effectively augment real-world facility data that is otherwise expensive to collect. By doing so, it is possible to identify which process steps have the most impact. While the primary work was validated in the final electrical test after a Back-End-Of-Line process, the methods used have been demonstrated to work in packaging operations. Two of the challenges to apply machine learning methods to manufacturing processes, are the complexity and heterogeneous nature of each individual unit operation, as well as the lack of complete characterization for every single item being produced due to throughput and cost constraints. In this work we demonstrate the benefits of using a hierarchical approach to product manufacturing in which individual unit operations are modelled individually, but results of previous operations inform and forecast subsequent manufacturing processes. To demonstrate this methodology, we have selected a portion of a Semiconductor manufacturing process, composed by over 40 distinct unit operations.