Events

A recent report by the US Department of Energy defines the area of scientific machine learning as “a core component of artificial intelligence (AI) and a computational technology that can be trained, with scientific data, to augment or automate human skills”, which has “the potential to transform science and energy research”. Physics-informed neural networks have been introduced in the scientific machine learning literature, as a promising means of solving partial differential equations of computational physics. However, the cost of training these networks is significant and the question of whether they can be competitive to conventional partial differential equation solvers is still open. In this presentation, we discuss the physics-informed neural network-based solution of Maxwell’s equations in the time-domain. We focus on the potential of relevant solvers to be competitive in terms of execution time with the popular Finite-Difference Time-Domain (FDTD) method, including training time in the comparison. To this end, we present a physics-informed deep operator network (PI-DON) for the solution of Maxwell’s equations in the time-domain. The PI-DON integrates a Deep Curl Operator (DCO) that is trained to approximate the discrete curl operator, combined with an unsupervised, physics-informed training process used by the network to simulate specific electromagnetic structures. The PI-DON demonstrates strong generalizability to structures that include small material and geometric variations with respect to the ones used during its unsupervised training. We exploit this feature to dramatically accelerate simulations of large frequency-selective surfaces and metasurfaces, as well as uncertainty quantification analyses on 3-D microwave circuits. Co-sponsored by: University of Utah ECE Department Speaker(s): Dr. Costas Sarris, Virtual: https://events.vtools.ieee.org/m/534364