Abstract: Pore-scale simulations accurately describe transport properties of fluids in the subsurface. These simulations enhance our understanding of applications such as assessing hydrogen storage efficiency and forecasting CO$_2$ sequestration processes in underground reservoirs. Nevertheless, they are computationally expensive due to their mesoscopic nature. In addition, their stationary solutions are not guaranteed to be unique, so multiple runs with different initial conditions must be performed to ensure sufficient sample coverage. These factors complicate the task of obtaining representative and reliable forecasts. To overcome the high computational cost hurdle, we propose a hybrid method that couples generative diffusion models and physics-based modeling. Upon training a generative model, we synthesize samples that serve as the initial conditions for physics-based simulations. We measure the relaxation time (to stationary solutions) of the simulations, which serves as a validation metric and early-stopping criterion. Our numerical experiments revealed that the hybrid method exhibits a speed-up of up to 8.2 times compared to commonly used initialization methods. This finding offers compelling initial support that the proposed diffusion model-based hybrid scheme has potentials to significantly decrease the time required for convergence of numerical simulations without compromising the physical robustness.
Introduction
Simulating multiphase fluid transport in subsurface fractures is critical for climate change mitigation strategies, including carbon sequestration and large-scale hydrogen storage. Direct Numerical Simulations (DNS), such as the Lattice Boltzmann Method (LBM), provide high-fidelity insights into these processes but are notoriously computationally intensive. These simulations often require millions of iterations to reach a stationary state.
A significant bottleneck in this process is the initialization of the simulation. Traditional methods typically rely on random noise or simple geometric heuristics (e.g., Euclidean distance-based placement). Because these initial states are often physically unrealistic, the solver must expend massive computational effort simply to "relax" the system into a physically consistent state before it can progress toward a final solution.
Methods: Geometric Conditioning and Hybrid Workflow
The proposed method introduces a hybrid scheme that couples Denoising Diffusion Probabilistic Models (DDPMs) with traditional physics solvers. The core innovation is the use of geometric conditioning. Instead of generating generic fluid patterns, the diffusion model is conditioned on the specific fracture geometry by feeding it a binary image of the rock fracture as an auxiliary channel.
The workflow consists of three main steps:
- Generation: The trained diffusion model synthesizes a "near-optimal" fluid configuration for a given, unseen fracture geometry.
- Initialization: This generated configuration is used as the "warm start" initial condition for a physics-based LBM simulator.
- Simulation: The simulator runs until it reaches a stable, low-energy stationary solution.
The model was trained on a dataset of 1,000 synthetic 2D fractures. Unlike standard computer vision tasks that rely on pixel-wise metrics like Mean Squared Error (MSE), the primary validation metric here is the relaxation time—the number of iterations required for the physics solver to reach convergence.
Results and Discussion
The hybrid method demonstrates substantial efficiency gains. By providing a starting point that approximates the underlying physics (minimizing free energy), the diffusion-based approach significantly reduces the workload of the downstream solver.
Key performance metrics include:
- Speedup: The hybrid method achieved an average speedup of 8.24x compared to random initialization and 1.76x compared to Euclidean distance-based methods.
- Iteration Reduction: The average number of iterations required for convergence dropped from 2.588 million (random) to just 0.137 million (diffusion)—a 95% reduction.
- Predictability: The standard deviation of convergence times also decreased by 95%, making simulation timelines far more predictable.
Conclusion
This research demonstrates that generative AI can learn to approximate complex physical states, serving as a powerful tool for accelerating traditional numerical simulations. While currently limited to 2D fractures, the success of this hybrid scheme provides a clear path toward scaling to 3D geometries and more complex fluid properties. By reducing the computational barrier, this approach facilitates more comprehensive and reliable forecasting for critical subsurface energy and environmental applications.
Summary updated