Solving Inverse Problems in Magnetism with Physics-Informed Neural Networks

by Mr. Mykhailo Flaks

University of Mainz

at Quantum optics seminar

Wed, 13 Nov 2024, 16:00
Zoom Only

Abstract

This is a student seminar.

Zoom link: https://us02web.zoom.us/j/81690221170?pwd=PNxJDaH3oOKybbjQupe8OS9y1Ay3Mj.1

Abstract:
In this talk, I will present a physics-informed neural network (PINN) implementation for reconstructing magnetization and current distributions from experimentally recorded magnetic field maps. The method is specifically designed for the analysis of magnetic field maps obtained with single-spin magnetometry using the nitrogen-vacancy (NV) center in diamond. To solve for the source distribution form magnetic field maps, one needs to solve an ill-posed inverse problem, which is plagued by numerical artifacts due to noise in the measurements and incompleteness of the measurement data. I will review the standard techniques for solving inverse problems in magnetism and discuss their associated complexities. Then, I will demonstrate how implementing PINNs enables us to determine source distributions from noisy and incomplete experimental data. By leveraging the adaptable structure of neural networks, I will show how we can integrate extra constraints and prior knowledge into the reconstruction process, potentially allowing for improved reconstructions in specific cases.

Created on 09-11-2024 by Folman, Ron (folman)
Updaded on 09-11-2024 by Folman, Ron (folman)