AI and PINNs in Fractured Carbonate Reservoir Simulation

This article provides a comprehensive overview of how petroleum reservoir engineers utilize Physics-Informed Neural Networks (PINNs) and Artificial Intelligence (AI) for history matching to simulate multiphase fluid flow in fractured carbonate rock formations. By integrating fundamental laws of fluid dynamics directly into deep learning architectures, engineers overcome the severe computational bottlenecks of traditional numerical simulators. Furthermore, automated AI-driven history matching calibrates these complex geological models against real-world production data with unprecedented speed and accuracy.

The Challenge of Fractured Carbonate Formations

Carbonate reservoirs hold a massive portion of the world's oil and gas reserves, yet they are notoriously difficult to model. Unlike uniform sandstone, carbonate rocks are highly heterogeneous. They feature complex pore networks containing microscopic matrix pores, large dissolution cavities (vugs), and extensive natural fracture networks.

When simulating multiphase fluid flow—the simultaneous movement of oil, water, and gas—through these formations, engineers must account for dual-porosity and dual-permeability systems. Fluids flow rapidly through fractures but slowly bleed out of the tight rock matrix. Traditional reservoir simulation relies on dividing the reservoir into millions of grid blocks and solving Partial Differential Equations (PDEs) using finite-difference or finite-volume methods. In fractured carbonates, the extreme contrast in fluid velocity between the matrix and the fractures requires exceptionally fine grids and tiny time steps, making traditional simulations computationally expensive and slow.

Physics-Informed Neural Networks (PINNs) for Multiphase Flow

To bypass the limitations of grid-based numerical solvers, engineers use Physics-Informed Neural Networks. Standard machine learning models struggle in reservoir engineering because subsurface data is sparse, and purely data-driven models often produce predictions that violate the laws of physics (such as creating mass out of nowhere).

PINNs solve this by embedding the governing physical equations directly into the neural network's loss function. During training, the AI does not just learn from historical data; it is penalized if its predictions violate fundamental fluid dynamics. Engineers program the PINN's loss function to include:

  • Conservation of Mass: Ensuring the total mass of oil, water, and gas remains constant within the system.
  • Darcy’s Law for Multiphase Flow: Governing how each fluid phase moves based on pressure gradients, rock permeability, fluid viscosity, and relative permeability.
  • Capillary Pressure and Gravity Effects: Accounting for the physical forces that trap fluids in the tight carbonate matrix or drive them through vertical fractures.

Because PINNs are mesh-free, they do not require the complex gridding of fractures. Instead, engineers often couple two distinct neural networks: one representing the rock matrix and another representing the fracture network. A transfer function links the two networks, accurately simulating the fluid exchange between the matrix and fractures in a fraction of the time required by conventional simulators.

AI-Driven History Matching

Building a forward simulation model with PINNs is only half the battle. Because engineers cannot see the exact layout of underground fractures, the initial geological model is full of uncertainties. To make the model reliable, engineers perform "history matching." This is an inverse problem where engineers adjust the model’s unknown parameters (like fracture permeability and porosity) until the simulator's output matches historical field data, such as bottom-hole pressure and water-cut rates.

In fractured carbonates, manual history matching is nearly impossible due to the millions of variables involved. Engineers use AI to automate and accelerate this process through several methods:

  • Surrogate Modeling: Once a PINN is trained, it acts as a high-speed surrogate model. It can run a full-field multiphase flow simulation in seconds rather than days. This allows engineers to run tens of thousands of realizations to find the best match.
  • Ensemble Smoother with Multiple Data Assimilation (ES-MDA) and Deep Learning: AI algorithms, such as Variational Autoencoders (VAEs) or Generative Adversarial Networks (GANs), are used to parameterize complex fracture networks into a low-dimensional latent space. Optimization algorithms then adjust these latent variables to match production history while strictly preserving the realistic geological structures of the carbonate rock.
  • Reinforcement Learning: AI agents are trained to systematically tweak reservoir parameters. The agent receives a "reward" when the discrepancy between simulated multiphase flow and actual historical production decreases, rapidly converging on an accurate reservoir characterization.

The Integrated Workflow

In practice, petroleum engineers combine PINNs and history-matching AI into a unified, continuous loop. First, generative AI creates multiple plausible realizations of the carbonate fracture network based on seismic and well-log data. Next, the PINN rapidly simulates multiphase fluid flow across these realizations, ensuring all predictions obey physical laws. Finally, the history-matching AI compares the PINN outputs to real-world sensor data from the oilfield, automatically updating the geological parameters until the simulated fluid flow mirrors reality. This synergy allows engineers to accurately forecast future production and optimize recovery strategies in the world's most complex reservoirs.