Multi-output deep learning model for simultaneous prediction of figure of merits (I on, G m, and V th) of gallium nitride high electron mobility transistors

This work reports on the development of a multi-output deep learning (DL) model for simultaneous prediction of the figure of merits (Ion, Gm, and Vth) of a gallium nitride (GaN) based high electron mobility transistors (HEMTs) for various epitaxial structures. To generate an initial data set, 2160 GaN HEMTs have also been simulated by an experimentally validated simulation methodology. A generative adversarial network (GAN) has been also introduced in semiconductor device modeling to augment the training data set. The generated data set by GAN is found to be in good agreement with the initial data set with a Frechet Inception Distance score of 0.151. The final data set has seven dimensions, i.e., aluminum gallium nitride (AlGaN) thickness (tAlGaN), aluminum content in AlGaN, doping in AlGaN, type of doping in AlGaN, Ion, Gm, and Vth, where the first four are inputs and the last three are the outputs of the DL model. The DL model is developed with the possibility of reducing unnecessary use of technology computer-aided design simulations for similar types of problems as such simulations require huge computational resources, expertise, and development time to obtain output. Mean squared error and R-squared values for the predicted Ion, Gm, and Vth are 59.69, 4.28, and 0.09, and 0.99, 0.99, and 0.97, respectively.