Joint Learning for Seismic Inversion

Collaborators: Ahmad Mustafa, Motaz Alfarraj, and Ghassan AlRegib


Goal/Motivation: Leverage beneficial information from multiple datasets in a joint learning setup to mitigate high labeled data requirements for seismic impedance inversion

Challenges: While transfer learning is a popular option to leverage knowledge from existing well-seismic datasets on new data, it falls short in delivering optimal performance for various reasons: firstly, there does not exist a massive repository of labeled data to learn good representations on like it does for natural images; secondly, the distributions of well logs can differ substantially from one survey to another leading potentially to performance deterioration with the application of transfer learning.

High Level Description of the Work: We propose a joint learning setup whereby multiple networks with similar architectures are trained simultaneously on different dataset. A soft penalty imposing regularization on the weights is imposed that forces the networks to search for their respective solutions in the proximity of each other to leverage mutually beneficial information. This results in the networks using cross-domain information while still specializing on their respective tasks. We benchmark the proposed framework against a variety of traditional transfer learning schemes and demonstrate our method to outperform. 

Related Publications:

  1. A. Mustafa, and G. AlRegib, "Joint Learning for Seismic Inversion: An Acoustic Impedance Estimation Case Study," in Expanded Abstracts of the SEG Annual Meeting, Houston, TX, Oct. 11-16 2020. [PDF][Code][Video]

  2. A. Mustafa, M. Alfarraj, and G. AlRegib, "Joint Learning for Spatial Context-Based Seismic Inversion of Multiple Datasets for Improved Generalizability and Robustness," in Geophysics, vol. 86, no. 4, Mar. 26 2021. [PDF][Code]