Predicting plasmonic pole representations of complex dielectric spectra using convolutional neural networks

Abstract

Extracting spectral properties such as energy position, broadening, and spectral weight from dielectric spectra is critical for interpreting collective electronic excitations in many-body physics. Conventional approaches typically rely on nonlinear fitting procedures that can be computationally demanding and highly sensitive to initialization and fitting choices. In this work, we develop a convolutional neural-network framework for the direct inversion of dielectric spectra into their underlying plasmonic pole structures using a multipole-Padé representation. Rather than training on a constrained database of spectra associated with a specific set of materials, the network is trained entirely on synthetic spectra generated from the analytical multipole-Padé expression with randomized parameters. This enables the network to learn the general mapping between the spectra and their features in an unbiased way, without requiring large, material-specific datasets derived from real materials. We demonstrate that the synthetically trained network generalizes to complex first-principles and experimental dielectric spectra, extracting the underlying pole parameters with high accuracy in a single forward pass. This approach provides an efficient and robust alternative to conventional nonlinear fitting, enabling high-throughput, automated analysis of dielectric spectra across diverse materials and applications.