An ANN-based surrogate model for wave propagation in uncertain media. Issue 2 (4th March 2023)

• Record Type: Journal Article
• Authors: Cheng, Xi; Shao, Wei; Wang, Kai; Wang, Bing-Zhong
• Is Part Of: Waves in random and complex media. Volume 33:Issue 2(2023)
• Page Start: 428
• Page End: 441
• Publication Date: 2023-03-04
• Subjects: Artificial neural network (ANN) -- ground penetrating radar (GPR) -- overfitting -- selecting activation function -- surrogate model; Electromagnetic waves -- Periodicals; Electromagnetic waves -- Scattering -- Periodicals; 531.1133
• DOI: 10.1080/17455030.2021.1880669
• Languages: English
• ISSNs: 1745-5030

Abstract

ABSTRACT: The parameter values of complex media can be uncertain in reality and uncertainty quantification (UQ) in the simulation of wave propagation is essential. In this paper, an efficient surrogate model based on the artificial neural network (ANN) is developed to imitate the concerned ground penetrating radar (GPR) calculation. A two-dimensional GPR system is simulated with the auxiliary differential equation (ADE) finite-difference time-domain (FDTD), and the soil is considered as a nonmagnetic medium with the frequency-dependent dielectric permittivity represented by a two-term Debye model. A solid metallic object and a piece of dry granite are buried in the dispersive soil whose parameter values are considered uncertain. The framework of the surrogate model is presented for GPR system modeling. Different popular activation functions are applied to the proposed surrogate model and their results are analyzed. Compared with other activation functions, the activation function of exponential linear unit has a better performance in the training process. Moreover, the dropout method is used to reduce overfitting to a large extent. At last, the well trained surrogate model is applied for UQ to GPR calculation instead of running one thousand full-wave GPR simulations, and the results show its accuracy and efficiency.