Multiscale modeling and neural network model based control of a plasma etch process

Use of Optical Emission Spectroscopy Data for Fault Detection of Mass Flow Controller in Plasma Etch Equipment

To minimize wafer yield losses by misprocessing during semiconductor manufacturing, faster and more accurate fault detection during the plasma process are desired to increase production yields. Process faults can be caused by abnormal equipment conditions, and the performance drifts of the parts or components of complicated semiconductor fabrication equipment are some of the most unnoticed factors that eventually change the plasma conditions. In this work, we propose improved stability and accuracy of process fault detection using optical emission spectroscopy (OES) data. Under a controlled experimental setup of arbitrarily induced fault scenarios, the extended isolation forest (EIF) approach was used to detect anomalies in OES data compared with the conventional isolation forest method in terms of accuracy and speed. We also used the OES data to generate features related to electron temperature and found that using the electron temperature features together with equipment status variable identification data (SVID) and OES data improved the prediction accuracy of process/equipment fault detection by a maximum of 0.84%.

Multiscale Modeling and Recurrent Neural Network Based Optimization of a Plasma Etch Process

In this article, we focus on the development of a multiscale modeling and recurrent neural network (RNN) based optimization framework of a plasma etch process on a three-dimensional substrate with uniform thickness using the inductive coupled plasma (ICP). Specifically, the gas flow and chemical reactions of plasma are simulated by a macroscopic fluid model. In addition, the etch process on the substrate is simulated by a kinetic Monte Carlo (kMC) model. While long time horizon optimization cannot be completed due to the computational complexity of the simulation models, RNN models are applied to approximate the fluid model and kMC model. The training data of RNN models are generated by open-loop simulations of the fluid model and the kMC model. Additionally, the stochastic characteristic of the kMC model is presented by a probability function. The well-trained RNN models and the probability function are then implemented in computing an open-loop optimization problem, in which a moving optimization method is applied to overcome the error accumulation problem when using RNN models. The optimization goal is to achieve the desired average etching depth and average bottom roughness within the least amount of time. The simulation results show that our prediction model is accurate enough and the optimization objectives can be completed well.