Speaker
Description
Numerical modelling plays a crucial role in revealing the behaviour of light and matter interactions at the nanoscale, exploiting computational schemes such as the Discontinuous Galerkin Time-Domain (DGTD) method. Given the computational complexity associated to this task, we study reduced-order modelling (ROM) due to the pressing need for fast surrogate models capable of handling physically and geometrically parametrized electromagnetic problems. Traditional ROM techniques like Proper Orthogonal Decomposition (POD) and Greedy algorithm have already been investigated in the literature, along with their inherent limitations in effectively capturing nonlinear phenomena. Here, we exploit a deep learning-based ROM strategy showing promising potential, the Graph Convolutional Autoencoder (GCA) method, serving as a nonlinear extension of POD compression, harnessing the power of Graph Neural Networks to induce a geometric bias in the learning process when dealing with complex and unstructured meshes. We propose a fully data-driven nonlinear ROM extending the GCA method tailored for time-domain electromagnetics, fully exploiting the training dataset of high-order (in space and time) DGTD snapshots, and the spectral quantities of interest. We explore the use of advanced latent-space propagators to better capture the high-frequency behaviour of electromagnetic problems. This direction aims to unlock faster and more accurate surrogate models for time-domain electromagnetics, especially for applications requiring complex geometries, in the context of inverse design and optimization.