Reinforcement Learning based Intelligent Semiconductor Manufacturing Applied to Laser Annealing

Abstract

The recent shift in the paradigm of the industrial revolution, i.e., Industry 4.0, has forced industries to reemphasize manufacturing standards to improve the manufacturing and production systems. In this work, instead of more popular supervised and unsupervised learning, we employ reinforcement learning (RL) to determine the process parameters to minimize the sheet resistance of the device. We have used seven independent variables used in fabricating the devices such as dopant ions (I), dose (D), energy (E), wavelength (W), repetition rate (Rep), temperature (T), and power (P). The trained RL agent will identify the combinations of independent variables parameters that minimize the sheet resistance. Advantage actor-critic (A2C) and deep Q-network (DQN) agents have been used with batch sizes of 2, 5, and 10, with exploration rates of 0 and 0.2 for every batch size. Both agents with a batch size of 2 and an exploration rate of 0.2 have given the best result. We also show that compared to the Taguchi method, RL can potentially provide faster locating of optimal experimental conditions.