Testing of a machine learning method into geophysical inversion

Ground-based radars, shown above, are a source of observations for which geophysical inversion is used. Credit: Comprehensive Nuclear-Test-Ban Treaty, CC BY 2.0

A common problem in geology is the need to obtain an invisible physical structure based on limited observations. For example, a radar observation that enters the ground tries to assess the structure of the ground without measurements on the ground. This class of problems is called inversion, in which the approximate physical model is corrected repeatedly until it is consistent with the observations.

The choice of models that act as Bayesian can have a strong impact on the inversion results. And because the models are generally more complex than in the physical world, the process can also lead to an overly simplified solution. To combat this problem, it is common to expand the theoretical model with well-known real-world examples, such as evidence collected from exits or wells. This combination can lead to a number of model changes to ensure the actual diversity of the former.

Recent advances in this method have been made on the basis of machine learning techniques. Convolutionary neural networks, which are used in computer vision, have been successfully proven in the integration of many learning patterns to increase more sensitive predictions with spatial resolution. Lopez-Alvis et al. consider one such neural network approach: Variation Autocoder (VAE).

Variant coders are able to not only “record” past educational information. They can generate new patterns that match the types of patterns included in the images, but are not identical to them. The authors test this capability by comparing VAEs using individually inserted images with those studied in a set of images between synthetic and real observational data.

One of the key findings of the study is that VAEs trained using image sets work better than those based on a single import. In fact, the integrated VAE performs for almost both synthetic and field data approximately as well as the best educational image. Thus, instead of looking for a “proper fit” model by performing multiple inversions with different inputs, combining training inputs into one VAE and performing only one inversion is significantly more efficient.

This study was published in the publication Journal of Geophysical Research: Hard Earth.


When it comes to AI, can we get rid of data sets?


More information:
J. Lopez-Alvis et al., Geophysical Inversion Using Variation Autodecoder for Modeling Previous Spatial Uncertainty, Journal of Geophysical Research: Hard Earth (2022). DOI: 10.1029 / 2021JB022581

Presented by the American Geophysical Union

This story has been republished with respect to Eos, which is hosted by the American Geophysical Union. Read the original story here.

Quote: Test of mechanical engineering approach to geophysical inversion (April 1, 2022) Retrieved April 1, 2022 from https://phys.org/news/2022-04-machine-approach-geophysical-inversion.html

This document is copyrighted. Except in the case of any fair transaction for the purpose of private study or research, no part may be reproduced without our written permission. Content is provided for informational purposes only.

Leave a Comment