Surrogate models for Rock-Fluid Interaction: A Grid-Size-Invariant Approach

Researchers develop grid-size-invariant neural network surrogate models for predicting rock-fluid interactions in porous media, offering a computationally cheaper alternative to traditional high-fidelity simulations. The approach demonstrates that UNet++ architecture outperforms standard UNet for this application, enabling significant memory reduction during training while maintaining prediction accuracy.

This research addresses a fundamental computational challenge in subsurface modeling by replacing expensive physics-based simulations with machine learning surrogates. Traditional high-fidelity models require extensive computational resources to achieve reliable predictions, making them impractical for scenarios requiring multiple runs such as uncertainty quantification and optimization studies. The study's innovation lies in the grid-size-invariant framework, which trains neural networks on smaller domains but applies them to larger computational grids, dramatically reducing training memory requirements without sacrificing predictive performance.

The competitive landscape of surrogate modeling has evolved significantly as machine learning techniques mature. Reduced-order models represent one approach, using sequential compression and prediction networks, but the paper demonstrates that single-network grid-size-invariant models achieve superior results. The choice of architecture matters substantially—UNet++ architectures incorporate dense skip connections that better capture multi-scale features relevant to fluid-rock interactions compared to standard UNet designs. This architectural advantage proves particularly valuable for complex, non-linear phenomena like fluid-induced rock dissolution where the solid field changes dynamically.

For the subsurface engineering and petroleum industry, this development enables faster feasibility studies and risk assessments in carbon sequestration, geothermal energy, and hydrocarbon extraction. The grid-size invariance property specifically addresses a practical constraint: training data often comes from smaller, more manageable simulations, yet practitioners need predictions across varied field scales. By eliminating this domain mismatch problem, the methodology accelerates adoption of AI-enhanced reservoir modeling across industry applications where computational efficiency directly impacts project economics and decision timelines.

• →Grid-size-invariant neural networks can infer on computational domains larger than training domains, reducing memory costs during model development.
• →UNet++ architecture outperforms standard UNet for rock-fluid interaction surrogate modeling with superior prediction accuracy.
• →The approach handles non-static solid fields caused by fluid-induced dissolution, a challenge traditional reduced-order models struggle with.
• →Single neural network surrogates demonstrate better performance than two-stage reduced-order model pipelines for this application.
• →The methodology enables faster uncertainty quantification and optimization studies by replacing computationally expensive high-fidelity simulations.