Accelerating Multiphase Flow Simulations with Denoising Diffusion Model Driven Initializations

Jaehong Chung, Agnese Marcato, Eric Guiltinan, Tapan Mukerji, Hari Viswanathan, Yen Ting Lin, Javier E. SantosarXiv 2024 · DOI · arXiv

Abstract

This study introduces a hybrid fluid simulation approach that integrates generative diffusion models with physics-based simulations, aiming at reducing the computational costs of flow simulations while still honoring all the physical properties of interest. These simulations enhance our understanding of applications such as assessing hydrogen and CO$_2$ storage efficiency in underground reservoirs. Nevertheless, they are computationally expensive and the presence of nonunique solutions can require multiple simulations within a single geometry. To overcome the computational cost hurdle, we propose a hybrid method that couples generative diffusion models and physics-based modeling. We introduce a system to condition the diffusion model with a geometry of interest, allowing to produce variable fluid saturations in the same geometry. While training the model, we simultaneously generate initial conditions and perform physics-based simulations using these conditions. This integrated approach enables us to receive real-time feedback on a single compute node equipped with both CPUs and GPUs. By efficiently managing these processes within one compute node, we can continuously evaluate performance and stop training when the desired criteria are met. To test our model, we generate realizations in a real Berea sandstone fracture which shows that our technique is up to 4.4 times faster than commonly used flow simulation initializations.

Introduction: The Initialization Bottleneck

Simulating how fluids like CO₂ or Hydrogen move through the tiny, jagged cracks (fractures) in underground rocks is critical for carbon sequestration and green energy storage. However, these "pore-scale" simulations are computationally brutal. Standard physics-based solvers, like the Lattice Boltzmann Method (LBM), start from a rough guess of where the fluids are and iteratively move toward a stable, minimum-energy state. If the initial guess is poor (e.g., random placement), the solver might spend hundreds of thousands of iterations just "fixing" the initial mess before it even begins the real work. This research addresses this "initialization bottleneck"—the massive computational cost required to reach a physically valid starting point.

Methods: Hybrid Diffusion-LBM Workflow

The proposed hybrid workflow combines Denoising Diffusion Probabilistic Models (DDPM) with traditional LBM solvers. Unlike standard image-generating AI, this model respects the "walls" of the rock through an auxiliary binary channel ($G$) that encodes the fracture geometry. This ensures the diffusion model only "paints" fluids within the available pore space.

U-Net architecture for the diffusion model, showing the integration of geometric conditioning G to predict the next denoising step.

In a clever hardware utilization move, the team trained the diffusion model on a GPU while simultaneously running LBM simulations on idle CPUs within the same compute node. They used the "number of iterations to convergence" as a real-time feedback metric to decide when the AI was "smart" enough to stop training.

Datasets

  • Synthetic Fractures: A dataset of 1,000 synthetic fractures generated using pySimFrac. By varying parameters like the Hurst exponent and surface roughness, they created 13,000 unique simulation cases (1,000 geometries × 13 saturation levels).
  • Real-World Benchmark: To test generalization, they used micro-CT scans of a real Berea sandstone fracture. This is a standard benchmark in geophysics, representing much more complex, "messy" geometry than the synthetic training data.

Key Insights: Physics over Pixels

A critical finding of this research is that standard statistical metrics like Mean Squared Error (MSE) only tell half the story. While MSE might plateau, the physical utility of the generated images (measured by how quickly the LBM solver converged) continued to improve. This suggests the model was learning the underlying "logic" of fluid interfaces beyond simple pixel-wise accuracy.

MSE Loss and LBM iterations to convergence vs training iterations. Note how physical performance continues to improve even after the loss plateaus.

By providing a "warm start" that is already near a local energy minimum, the diffusion model bypasses the chaotic early stages of a simulation where the solver often struggles with unphysical fluid overlaps.

Key Results

  • Significant Speedup: On the real Berea sandstone dataset, the diffusion-based initialization was 4.4x faster than random initialization and 2.7x faster than the Euclidean distance-based method (a common industry heuristic).
  • Iteration Reduction: During training, the average number of iterations required for the solver to reach convergence dropped from 2.6 million to just 140,000—a 95% reduction.
  • Efficiency: The entire training process took only about one hour on a single NVIDIA A100 GPU, making it a practical pre-processing step for large-scale simulation campaigns.

Conclusion and Future Work

This study demonstrates that generative AI can significantly accelerate complex physical simulations by providing high-quality initial conditions. While currently limited to 2D slices, future work aims to scale the diffusion model to 3D volumes. Additionally, predicting initial velocity fields and incorporating varying fluid properties (like surface tension) as model inputs could further enhance simulation efficiency.

Summary updated

Abstract

This study introduces a hybrid fluid simulation approach that integrates generative diffusion models with physics-based simulations, aiming at reducing the computational costs of flow simulations while still honoring all the physical properties of interest. These simulations enhance our understanding of applications such as assessing hydrogen and CO$_2$ storage efficiency in underground reservoirs. Nevertheless, they are computationally expensive and the presence of nonunique solutions can require multiple simulations within a single geometry. To overcome the computational cost hurdle, we propose a hybrid method that couples generative diffusion models and physics-based modeling. We introduce a system to condition the diffusion model with a geometry of interest, allowing to produce variable fluid saturations in the same geometry. While training the model, we simultaneously generate initial conditions and perform physics-based simulations using these conditions. This integrated approach enables us to receive real-time feedback on a single compute node equipped with both CPUs and GPUs. By efficiently managing these processes within one compute node, we can continuously evaluate performance and stop training when the desired criteria are met. To test our model, we generate realizations in a real Berea sandstone fracture which shows that our technique is up to 4.4 times faster than commonly used flow simulation initializations.
BibTeX
@misc{chung2024accelerating,
  author = {Jaehong Chung and Agnese Marcato and Eric Guiltinan and Tapan Mukerji and Hari Viswanathan and Yen Ting Lin and Javier E. Santos},
  title = {{Accelerating Multiphase Flow Simulations with Denoising Diffusion Model Driven Initializations}},
  year = {2024},
  eprint = {2406.19333},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2406.19333},
}

Loading related content...

Figure