In collaboration with GlobalFoundries, we’re excited to share the latest advancements in Processing Time (PT) prediction for semiconductor fabrication plants. Our breakthrough, powered by minds.ai Maestro platform and cutting-edge neural networks, was recently presented at the WinterSim 2023 conference and detailed in our published paper.
The Challenge
Semiconductor manufacturing relies on precise timing at each stage to optimize efficiency. Processing Time (PT) – the duration a tool takes to process a job – is a critical metric influencing key fabrication Key Performance Indicators (KPIs) like cycle time and throughput.
Our Solution
Leveraging an attention-based neural network within an automated Machine Learning Operations (MLOps) pipeline, we’ve addressed the complexities and temporal dependencies of fabrication processes. Data, enriched with features like chamber usage and process sequences, goes through seamless stages of collection, preprocessing, training, and deployment.
Key Results
Compared to the current model used by GlobalFoundries, our approach has shown a remarkable 43% to 80% reduction in Mean Absolute Error (MAE) for Processing Time predictions across crucial areas like Etch, Diffusion, and Deposition. The model also excels in providing high-quality predictions for other tools, surpassing the limitations of average processing time calculations over a predefined context. Through integration with Dispatching logic, utilization improvements of 3-5% are attained. By incorporating the Processing Time, and also Tool Availability, predictions with scheduling and Automated Material Handling Systems (AMHS) allows for even greater improvements in throughput and cycle time.
Moving Forward
We’re committed to expanding this groundbreaking solution to more manufacturing facilities and customers. If you’re keen to explore our PT prediction method or other semiconductor manufacturing solutions, feel free to reach out. Together, let’s propel the efficiency of semiconductor fabrication to new heights.
The initial results – as presented during the WinterSim 2023 Conference.
