Semi-supervised Sequence Modeling for Elastic Impedance Inversion

 Collaborators: Motaz Alfarraj and Ghassan AlRegib

Goal/Motivation: To incorporate physics-based forward model in traditional learning-based seismic inversion frameworks to unburden the setup from high labeled data requirements

Challenges: Seismic inversion is a challenging geophysics application involving the estimation of well properties away from the well locations using seismic data. Owing to the lack of wells in many cases, traditional learning algorithms are prone to overfitting, leading to inaccurate estimations of property profiles for subsurface data. 

High Level Description of the Work: Where seismic inversion is a challenging application considering the high labeled data requirements, it benefits from the availability of physics-based models. We put forward the hypothesis that combining data-driven learning with physics-based forward modeling can potentially lead to better results than would have been possible working with only labeled data alone. We validate this hypothesis with impedance inversion performed on the popular industry benchmark marmousi 2 dataset in both elastic and non-elastic scenarios. 

References 

1. M. Alfarraj, and G. AlRegib, "Semi-Supervised Sequence Modeling for Elastic Impedance Inversion," in Interpretation, Aug. 2019. [PDF][Code]