Predictive Quality of In-Fabrication Products in Smart Manufacturing Using Graph-Based Deep Learning

Graph neural networks are a very powerful way to learn about relationships between entities in graphs. With the rise of IoT devices in manufacturing, more data is being collected to minimise the waste of both valuable resources and time for fabrication. In this paper, we introduce a methodology for predictive quality of in-fabrication products using graph neural networks. Data is collected from a live-working semiconductor wafer fabrication facility and used to produce heterogeneous graphs that represent the fabrication timeline of a wafer. The model uses the graph attention network architecture to classify whether a timeline is scrap or non-scrap. It uses historical graph-level labelled data and achieves an F1-score of 0.928, compared to baselines models of a LSTM and a Homogeneous Graph Attention Network with scores of 0.424 and 0.786 respectively. It gives a foundational framework for future anomaly detection with semiconductor fabrication, allowing real-world data to be analysed with graph-based deep learning tools to provide interpretation and accessible graph-based results.