Integration of on-line machine learning-based endpoint control and run-to-run control for an atomic layer etching process

Abstract

Control methods for Atomic Layer Etching (ALE) processes are constantly evolving due to the increasing level of precision needed to manufacture next-gen semiconductor devices. This work presents a novel, real-time Endpoint-based (EP) control approach for an Al₂O₃ ALE process in a discrete feed reactor. The proposed method dynamically adjusts the process time of both ALE half-cycles to ensure an optimal process outcome. The EP controller uses a machine learning-based transformer to take in variable-length, time-series pressure profiles to identify when the ALE process is complete. However, this model requires a large amount of process data to ensure that it will perform well even when under a variety of kinetic and pressure disturbances that mimic common issues in a real-world manufacturing environment. Thus, this work uses a multiscale modeling method that integrates a macroscopic Computational Fluid Dynamics (CFD) and a mesoscopic kinetic Monte Carlo (kMC) simulation to generate process data and test the proposed controllers. After testing the performance of the EP controller on individual runs, various combinations of ex-situ Run-to-Run (R2R) and EP controllers are examined in order to determine the strongest control strategy in a manufacturing environment. The final results show that the EP controller is highly accurate when trained on conditions that are representative of its implementation environment. Compared to traditional EWMA controllers, it has significantly fewer misprocesses, which enhances the overall control performance and efficiency of the ALE process.