Modeling the manufacturing performance of semiconductor equipment with deep and recurrent neural networks

Abstract

Clustered photolithography tool often has a bottleneck due to its complex internal structure and is by far one of the most expensive equipment tools. Therefore, it is very important to accurately and efficiently estimate the cycle time, lot residency time and throughput time, which are directly related to the production yield. Physics-based models, revealed by previous studies, are inaccurate, computationally complex, or accurate, but have low fidelities under certain settings. Therefore, deep and recurrent neural networks are used to make an accurate and robust model regardless of the input condition, e.g., when lot sizes or buffer sizes change in this paper. We additionally compare the accuracy of recurrent neural network when test input gets farther from train input condition.