Transfer Learning Model with Simulation and Experimental Data for Tool Virtualization in Poly-Si Etching

We investigate a model construction method for the digital twins (DTs) of semiconductor manufacturing tools (i.e., virtualized manufacturing tools) to maintain and correct manufacturing tools for nanoscale fabrication. Plasma etching is a non-linear phenomenon with a multi-input/output, and simulator calculations span extremely large spatiotemporal orders. Therefore, constructing a general-purpose plasma simulator and obtaining a DT using only the simulator is difficult. However, using machine learning to construct a DT model requires a large amount of time and experimental plasma etching data for training. Therefore, we utilize a transfer learning (TL) model that learns data from the simulation tools and experimental data obtained from the actual tools. We develop a DT model for plasma etching tools using the TL model and evaluate using a prediction task of the etching rate (ER) obtained from basic recipes. The TL model with the largest amount of simulation data showed the highest ER prediction accuracy.