Machine Learning models for Moment Magnetometry applied to High-Resolution Magnetic Maps

Paleomagnetic measurements of natural remanent magnetization were – until recently -- performed on large, millimeter to centimeter-sized rock samples using classical rock magnetometers. Developments of magnetic imaging (i.e., quantum diamond microscope (QDM) and scanning SQUID microscope) allow for the measurement of individual, weakly magnetic (10^-16 Am²) particle clusters within a sample. However, this increased spatial resolution adds to the complexity of the magnetic signal and requires novel approaches in analyzing the data.

Commonly, the magnetic moment of sources within a sample are analyzed by dipole inversion, meaning fitting a dipole signal to a cropped part of a magnetic map. While this gives accurate results for sources that are well separated and dipole-like in character, recovering net magnetic moments from complex maps proved difficult. 

Machine learning (ML) has been used to study a vast number of problems and has become a part of our daily lives. Here, we present the first ML model to predict net magnetic moments from magnetic maps. Our ML models consist of three simple convolutional neural networks for source declination, inclination, moment magnitude, each of which has been trained on a large (500k) synthetic map dataset. The models are tested against several datasets, varying in size, source distribution and resolution. Our models can quickly and accurately recover the magnetic moment, even for very complex (i.e., low dipolarity) maps. The accuracy of the models is < 8˚ for inclination and declination and < 10 % for moment. Surprisingly, we find little correlation in the prediction accuracy to the dipolarity of the map. These results show that ML is a promising alternative to dipole inversions, especially for maps with low dipolarity parameters, where traditional dipole inversions can fail.