Connectivity-informed drainage network generation using deep convolution generative adversarial networks.

Area of Science:

• Earth Sciences
• Computational Modeling
• Geomorphology

Background:

• Stochastic network modeling faces computational challenges in generating sufficient samples for statistical analysis.
• Existing methods for simulating drainage networks are often computationally intensive and time-consuming.

Purpose of the Study:

• To apply Deep Convolutional Generative Adversarial Networks (DCGANs) for efficient reproduction of drainage networks.
• To develop and evaluate a novel connectivity-informed method for improving DCGAN performance in network generation.

Main Methods:

• Utilized DCGANs to generate drainage networks from existing samples, bypassing repetitive stochastic modeling.
• Developed a connectivity-informed method converting network images to directional flow information and binary layers representing connectivity constraints.
• Compared DCGANs trained on original images, directional information only, and connectivity-informed directional information.

Main Results:

• The connectivity-informed method significantly improved DCGAN training effectiveness and accuracy in reproducing drainage networks.
• This approach offered a more compact representation of network complexity and connectivity compared to other methods.
• Generated drainage networks using the enhanced method showed superior fidelity to the original samples.

Conclusions:

• DCGANs, particularly with the connectivity-informed method, offer an efficient solution for generating complex network models.
• The approach is highly applicable to high-contrast images in earth and material sciences, aiding the study of networks and fractures.
• This research demonstrates a powerful application of AI in accelerating scientific discovery and analysis in geosciences.