Using Generative Adversarial Networks as a Fast Forward Operator for Hydrogeological Inverse Problems.

Area of Science:

• Hydrogeology
• Computational Science
• Geophysics

Background:

• Subsurface characterization relies on inverse techniques for hydrogeological modeling.
• Solving inverse problems requires forward operators, which become computationally intensive with increasing parameters.
• Existing methods face computational challenges in complex subsurface modeling.

Purpose of the Study:

• To investigate conditional generative adversarial networks (cGAN) as an emulator for forward operators in hydrogeological inverse problems.
• To assess the accuracy and computational feasibility of using cGAN within the adaptive importance sampling method (PoPEx).
• To evaluate cGAN's potential to replace traditional forward operators.

Main Methods:

• Utilized conditional generative adversarial networks (cGAN) to emulate the forward operator.
• Integrated the trained cGAN model into the posterior population expansion (PoPEx) method.
• Trained the cGAN model on channelized geological structures.

Main Results:

• The cGAN model successfully reproduced state variables for given parameter fields.
• Integration of cGAN in PoPEx yielded satisfactory results for subsurface characterization.
• Computational time was reduced by up to 80% using cGAN as a surrogate forward model.

Conclusions:

• Conditional generative adversarial networks (cGAN) can effectively emulate forward operators in hydrogeological inverse problems.
• The cGAN-PoPEx integration offers a computationally efficient alternative to traditional methods.
• The primary limitation is the significant training time required for cGAN model development.