The work here has been published during the 2023 version of the annual Winter Simulation Conference. You can read the abstract below, to read the full paper please fill in the information in the form on this page.
In partnership with GlobalFoundries we have significantly advanced Processing Time (PT) and machine availability prediction in fabrication plants, utilizing an attention based neural network. This model is integrated into an automated Machine Learning Operations (MLOps) pipeline consisting of data collection, preprocessing, training and deployment. The data is augmented with features such as chamber usage and process sequences. Compared to the current model, which calculates average processing times over a predefined context, our approach has reduced the Mean Absolute Error (MAE) of PT predictions by 43% to 80% across the crucial areas: Etch, Diffusion, and Deposition. The model also produces high quality predictions for the remaining tools. The model is in the process of being implemented in the fabrication process (FAB) to improve scheduling, dispatching, and improve crucial Key Performance Indicators (KPIs) such as cycle time and throughput.