GP010-06
Recovery of net magnetic moments from complex magnetic field maps using convolutional neural networks

Tuesday, 15 December 2020: 21:07
Virtual
Raisa Trubko1,2, Xingyu Li3, Roger R Fu4, Brendan J. Meade5, Michael Volk5, Kimberly Hess6 and Zoe Levitt7, (1)Harvard University, Earth and Planetary Sciences, Cambridge, United States, (2)Worcester Polytechnic Institute, Physics Department, Worcester, United States, (3)University of Science and Technology of China, School of Earth and Space Sciences, Hefei, Anhui, China, (4)Harvard University, Department of Earth and Planetary Sciences, Cambridge, MA, United States, (5)Harvard University, Cambridge, MA, United States, (6)Harvard University, Earth and Planetary Sciences, Cambridge, MA, United States, (7)Massachusetts Institute of Technology, Cambridge, MA, United States
Recorded Presentation
Abstract:
The development and advancement of new instrumentation, such as the Quantum Diamond Microscope (QDM), has created a need for new data analysis methods. This necessity stems from the micrometer spatial resolution of the QDM, which may result in complex, non-dipolar magnetic field maps. Because information about the direction and intensity of ancient magnetic fields is encoded in the net magnetic moment of paleomagnetic samples, developing techniques to quantify the net moment from QDM maps is critical for its applications in the Earth sciences.

Extracting net magnetic moments from complex magnetic field maps is a difficult inverse problem. Currently, net moment analysis of QDM magnetic field maps uses a combination of numerical upward continuation and dipole fitting. Although capable of <5% error, this method requires significant source free space to be mapped around the sample, making it labor intensive and impractical for some samples. Inversions using higher degree and order spectral fits have fewer constraints but are computationally costly.

Here, we demonstrate that machine learning can be a useful tool for analyzing complex, non-dipolar magnetic field maps in cases where other currently available analysis techniques are insufficient. More specifically, we train three convolutional neural networks to each report a component of the vector magnetic moment from an input magnetic field map as a classification problem and then calculate the net magnetic moment and direction. To train the neural networks, we generate a large synthetic data set that spans the full range of net moment magnitudes and directions. We also characterize the dipolarity of the synthetic data set to permit comparison to real datasets. To apply the algorithm to a wide range of net moment magnitudes, we develop a normalization pre-conditioning procedure based on the root mean square (RMS) of the input magnetic field map. The net moment results of our algorithm are currently within ~5% in magnitude and ~5˚ in direction, matching best results from traditional algorithms described above but with less stringent requirements on measurement parameters.

<< Previous Abstract | Next Abstract >>