Field-material coupled neural network: A novel prior-free and data-free inverse problem solver for extracting complex dielectric constant in terahertz band

TL;DR

This paper introduces a physics-informed neural network that directly extracts complex dielectric constants from terahertz measurements without prior data or models, improving material characterization accuracy.

Contribution

The proposed FMCNN uniquely couples field and material neural networks through Maxwell's equations, enabling prior-free and data-free inversion for dielectric constant extraction.

Findings

Good agreement with traditional models over broad frequency range

Effective in high-frequency THz band above 0.2 THz

Provides a physics-consistent, data-efficient inversion method

Abstract

Accurate extraction of the complex dielectric constant in the terahertz (THz) band is essential for material characterization and non-destructive evaluation yet remains challenging due to the ill-posed nature of electromagnetic inverse problems and the limited availability of reliable reference data. In this work, a field-material couple neural network (FMCNN) is proposed to retrieve the complex dielectric constant directly from THz measurements. The FMCNN consists of a field neural network and a material neural network that are strongly coupled through the frequency-domain Maxwell equations in the form of a Helmholtz equation, with the governing physics enforced by partial differential equation (PDE) and boundary condition constraints. This formulation enables prior-free and data-free inversion, requiring only measured test data as input. The extracted dielectric constants are validated by comparison with results from a one-dimensional normal-incidence model and the Drude-Lorentz model, showing good agreement over a broad frequency range, particularly above 0.2 THz. These results demonstrate that the FMCNN provides a physics-consistent and data-efficient approach for material parameter extraction in the THz band, offering an alternative to conventional model-based methods.