A Machine-Learning-based Model for Emerging Memories Featuring Multiple States

Machine learning approaches have been studied for nanoelectronics device modeling, including emerging memories. In this work, a recurrent network is developed to cover the multi-state memory. A multi-state envelope hysteresis function (MSEHF) is constructed, for which a recurrent module is designed to reproduce. The sequential data for the network training is designated by combining different modes for all the state transitions. With a feedforward module for the nonlinear states, the obtained composite recurrent neural network is shown to capture key multi-state memory properties. While the proposed approach is also compatible with two-state memories, the machine learning model is applicable in circuit simulations with the desired high accuracy and correct functions.