The Implementation of Neural Networks for Phaseless Parametric Inversion

Abstract

We present a machine learning work flow for the parametric inversion of grain bin measurements in which a neural network is trained solely on synthetic data for a unique bin geometry. This neural network can subsequently be used to rapidly obtain 4 inversion parameters (grain height, cone angle, and bulk real and imaginary permittivity of the grain) from uncalibrated, experimental data. We have previously shown that these 4 parameters can be used to calibrate experimental data and serve as prior information for full-data inversion. Our results show that a densely connected neural network that supports multifrequency data can better predict the cone angle of grain, and perform almost as well on grain height predictions, as the single-frequency simplex inversion method previously described. These findings suggest that neural networks trained on synthetic data may be a useful tool in the inversion of experimental data, providing prior information and a method for calibration.