Reproduce “A Test-Time Learning Approach to Reparameterize the Geophysical Inverse Problem with a Convolutional Neural Network” on OpenResearch

About this paper

Anran Xu, Lindsey J. Heagy

Researchers at the University of British Columbia developed Deep Image Prior Inversion (DIP-Inv), an unsupervised test-time learning method that reparameterizes the Direct Current (DC) resistivity inverse problem using a Convolutional Neural Network. This approach accurately recovers subsurface structures, such as dipping conductors and proximate anomalies, with fewer artifacts and better resolution compared to traditional Tikhonov regularization methods.

Key insights

• Convolutional Neural Network (CNN) architectures implicitly provide regularization for ill-posed geophysical inverse problems, effectively replacing explicit smoothness terms.
• The bi-linear upsampling operator within the CNN is a critical component for achieving the observed smoothing and regularization effects.
• The number of hidden blocks in the CNN and the use of dropout layers offer mechanisms to control the strength of this implicit regularization.

Summary by alphaXiv.