Testing a machine learning approach to geophysical inversion

Ground penetrating radar, shown above, is one source of observations for which geophysical inversion is used. Credit: Comprehensive Nuclear Test Ban Treaty, CC BY 2.0

A common problem in geosciences is the need to infer an unseen physical structure based on limited observations. For example, a ground-penetrating radar observation attempts to infer the subterranean structure without any in situ measurements. This class of problems is called inversion, in which an assumed physical model is repeatedly adjusted until it is consistent with observations.

Inversion results can be strongly affected by the choice of models, which acts as a Bayesian prior. And because models are generally less complex than the physical world, the process can also result in an oversimplified solution. To combat these difficulties, it is common to supplement a theoretical model with known real-world instances, such as evidence gathered from outcrops or boreholes. This combination can result in a number of model permutations to provide more realistic diversity for the prior.

Recent advances in this approach have been made based on machine learning techniques. Convolutional neural networks similar to those used in computer vision have proven effective in integrating many training samples to produce more nuanced priors with increased spatial resolution. Lopez-Alvis et al. examine one such neural network approach: the variational autoencoder (VAE).

Variational autoencoders are able to do more than just “regurgitate” past training data. They can generate new samples consistent, but not identical, with the types of patterns observed in the input images. The authors test this ability by comparing VAEs trained using individual input images with those trained on sets of images across synthetic and real observational data.

One of the main results of the study is that VAEs trained using image collections seem to perform better than those based on a single input. In fact, the combined VAE performs almost as well as the best training image for synthetic and terrain data. So rather than finding the “right match” pattern by performing many inversions with different inputs, it is significantly more efficient to combine the training inputs into a single VAE and perform a single inversion.

This study is published in the Journal of Geophysical Research: Solid Earth.

When it comes to AI, can we ditch datasets?

More information:

J. Lopez-Alvis et al, Geophysical inversion using a variational autoencoder to model an assembled spatial a priori uncertainty, Journal of Geophysical Research: Solid Earth (2022). DOI: 10.1029/2021JB022581

Provided by
American Geophysical Union

This story is republished courtesy of Eos, hosted by the American Geophysical Union. Read the original story here.

Testing a Machine Learning Approach to Geophysical Inversion (2022, April 1)
retrieved April 1, 2022
from https://phys.org/news/2022-04-machine-approach-geophysical-inversion.html

This document is subject to copyright. Other than fair use for purposes of private study or research, no
any part may be reproduced without written permission. The content is provided for information only.

Disclaimer! Verve Times is an automatic aggregator of all the media in the world. In each content, the hyperlink to the main source is specified. All trademarks belong to their rightful owners, all materials to their authors. If you are the content owner and do not want us to publish your materials, please contact us by email – [email protected]. Content will be deleted within 24 hours.

Sherry J. Basler