Machine learning-based run-to-run control of a spatial thermal atomic layer etching reactor

In response to the next technological revolution, atomic layer processes have emerged to produce high -performing, thin-film semiconductor materials. To overcome the long purging times required for conventional atomic layer processes, spatial atomic layer processes have been recognized for their ability to reduce processing times; however, they lack characterization and control. This research aims to construct two novel run-to-run (R2R) control systems using a machine learning model with an artificial neural network (ANN) and an exponentially weighted moving average (EWMA) method for the spatial thermal atomic layer etching (SALE) of aluminum oxide thin films. The two R2R controllers are used in conjunction with a multiscale computational fluid dynamics model of a SALE process with various disturbances to test their effectiveness. Closed-loop simulation results demonstrate that the ANN-based R2R control system reduces etching per cycle variability, maintains the process output within a small region around the setpoint, and outperforms the traditional EWMA-based R2R control system in efficiency.