Defect Diffusion-Informed Recurrent Neural Network for Investigating Dielectric Relaxation

Dielectric relaxation (DR) plays a critical role in understanding interactions between defects and dipoles in materials applied a time-varying electric field. Based on the observations during the DR experiment, intrinsic quantities of materials are extracted typically by exponential-like decay functions. However, these traditional methods involve the rigid constraint of mathematical models and the ill-posedness of inverse calculation, and thus lead to an ambiguous physical meaning of parameters and even failing in parameter estimations from noisy observations. In this paper, a forward method with the learnable current of defects is proposed to extract the physical quantities of DR like time constant, activation energy, defect number, and dielectric loss, where the defect current is learned by a recurrent neural network (RNN) with a cell of shallow multilayer perceptron (SMP) for solving the defect diffusion-controlled chemical reaction equation. After presenting the defect diffusion-informed RNN framework, the efficient architectures are discussed by utilizing TensorFlow-based implementations; and then, validation studies compared with the well-known analytical solution are performed; finally, the robustness and the generalization ability are demonstrated according to the experimental DR data in semiconductor devices. The suggested methodology provides a better DR analysis for discovering hidden physics in emerging materials and devices.