Effective dipole extraction from noisy magnetic field image data using deep convolutional neural networks

Abstract

Quantum sensors such as Nitrogen-Vacancy (NV) centers in diamond image magnetic fields with high spatial resolution, making them a powerful tool in characterizing a variety of magnetic samples. However, extracting net magnetic moment information from noisy magnetic field image data is an ill-posed inverse problem. We address this challenge with an image-based machine learning approach. We use a Residual Neural Network (ResNet) framework to obtain vector net magnetic moments of an effective dipole from noisy data. We train our model with synthetically generated magnetic sources superimposed on measured lab noise images. Our model estimates the net magnetic moment to within 64% at Signal-to-Noise Ratio (SNR) 0.1, and 16.5% at SNR 0.4 and above. We recover the declination angle to within 26.3 degrees at SNR 0.1 and above. Inclination angle is recovered to within 8.32 degrees at SNR 0.1 and above. These results demonstrate a robust solution for characterization of magnetic sources in noisy regimes, where traditional fitting techniques are insufficient.