Updating DMD Operators for Changes in Domain Properties

This paper introduces a lightweight, data-free framework for updating Dynamic Mode Decomposition (DMD) surrogate models to accommodate changes in reservoir permeability and well locations, enabling rapid, high-fidelity optimization and uncertainty analysis for geological carbon storage without the computational cost of retraining.

The Gist

Imagine you are a geologist trying to figure out the best way to inject carbon dioxide (CO₂) deep underground to store it safely. To do this, you need to run complex computer simulations that act like a "digital twin" of the Earth's underground rock layers. These simulations are incredibly accurate, but they are also slow and expensive—like trying to predict the weather by building a new, full-scale atmosphere in your living room every time you want to know if it will rain tomorrow.

To speed things up, scientists use surrogate models. Think of these as "crystal balls" or "shortcuts." They are trained on a few expensive simulations and then learn to predict the future quickly. However, there's a catch: these crystal balls are fragile.

If the underground rock changes even slightly (for example, if the rock becomes more porous or "spongy," allowing fluids to flow faster), the old crystal ball breaks. Usually, the only solution is to throw it away, run new expensive simulations, and build a brand-new crystal ball. This defeats the purpose of having a shortcut in the first place.

This paper introduces a magic wand that fixes the crystal ball instead of throwing it away.

Here is how the authors' new method works, using simple analogies:

1. The "Fast-Forward" Button (Uniform Changes)

Imagine you have a movie of water flowing through a pipe.

• The Problem: If you suddenly make the pipe wider (increasing permeability), the water flows faster. If you just play the old movie at the same speed, the prediction is wrong.
• The Old Way: Stop the movie, film a new one with the wider pipe, and edit it. (Takes forever).
• The New Way: Keep the original movie, but hit the "Fast-Forward" button.
  • If the rock is 4 times more permeable, the authors show you how to mathematically speed up the movie by 4x.
  • They also adjust the "volume" (pressure) so it doesn't sound too loud or too quiet.
  • Result: You get a perfect prediction of the new, faster flow without ever filming a new scene. It's like realizing that if you run a race on a track with better shoes, you don't need to re-run the whole race; you just calculate how much faster you would have been.

2. The "Stretchy Map" (Uneven Changes)

Now, imagine the underground rock isn't just uniformly better; it's a patchwork quilt. Some areas are like smooth highways (high permeability), and others are like muddy swamps (low permeability).

• The Problem: A standard model treats every square inch of the map equally. It wastes its "brain power" (mathematical resources) studying the muddy swamps where nothing interesting happens, while missing the details on the highways where the CO₂ is actually rushing through.
• The New Way: The authors invented a "Stretchy Map."
  • Imagine a rubber map of the underground. When you stretch it, the "highway" areas (high permeability) get pulled apart and become huge, taking up more space on the map. The "swamp" areas get squished down.
  • Now, when the model looks at the map, it naturally spends more time and attention on the highways because they physically occupy more space on the stretched paper.
  • Result: The model becomes smarter about where the fluid is actually going, even if the rock properties change, without needing new data.

Why This Matters

In the real world, we often don't know the exact properties of the underground rock until we start drilling. We need to ask "What if?" questions constantly:

• "What if the rock is 50% more permeable?"
• "What if we drill the well in a slightly different spot?"

Before this paper: Answering these questions meant waiting days or weeks for new, expensive computer simulations.
After this paper: You can answer these questions in seconds. The authors' method updates the "shortcut" model instantly using math, keeping it accurate enough for real-world decisions but fast enough for real-time optimization.

The Bottom Line

The authors have found a way to update a model's "muscle memory" when the environment changes, rather than forcing it to relearn everything from scratch.

• Uniform changes? Just speed up the clock and adjust the volume.
• Messy, uneven changes? Stretch the map so the important parts get more attention.

This allows engineers to make faster, safer, and more efficient decisions about storing carbon underground, turning a slow, expensive process into a rapid, agile one.

Technical Summary

1. Problem Statement

In geological carbon sequestration and subsurface resource management, high-fidelity multiphase flow simulators are computationally expensive, making them unsuitable for real-time optimization, control, and uncertainty quantification. Data-driven surrogate models, specifically Dynamic Mode Decomposition (DMD) and DMD with control (DMDc), offer a reduced-order alternative.

However, a critical limitation exists: standard DMD surrogates are trained on a specific set of reservoir properties (e.g., permeability distribution, well locations). When these domain properties change (e.g., due to geological uncertainty or different well placement scenarios), the conventional approach requires regenerating high-fidelity simulation snapshots and retraining the surrogate. This process negates the computational speed advantages of using surrogates in the first place. The paper addresses the challenge of updating existing DMD operators to accommodate changes in permeability without retraining or generating new simulation data.

2. Methodology

The authors propose a lightweight, algebraic framework to update DMD operators based on two distinct strategies for permeability changes: Uniform Scaling and Anisotropic (Spatially Varying) Scaling.

A. Uniform Permeability Scaling

This approach handles cases where permeability changes by a constant factor ($\kappa$) across the entire domain.

• Theoretical Basis: Based on Darcy's law and the pressure equation, a uniform increase in permeability accelerates fluid transport speed. The authors derive that the new pressure field evolves $\kappa$ times faster, while the saturation field remains invariant in shape but accelerated in time.
• Time and Amplitude Rescaling:
  • Time: The time variable is stretched ($\tau = \kappa t$).
  • Pressure: A multiplicative amplitude correction ($\kappa^{-1}$) is applied to restore physical balance for compressible fluids.
• Operator Updates:
  • Eigenvalues: Continuous-time eigenvalues scale linearly ($\lambda_{new} = \kappa \lambda_{old}$). In discrete time, the state transition matrix eigenvalues are raised to the power of $\kappa$ ($\mu_{new} = \mu_{old}^\kappa$).
  • DMDc Operators: The authors derive algebraic rules to update the system matrix ($A$) and input matrix ($B$) directly. For non-integer $\kappa$, they utilize real Schur decomposition and analytic continuation of geometric sums to compute $A^\kappa$ and the corresponding $B^\kappa$ without explicit matrix inversion.
  • Snapshot Reconstruction: If the new time step does not align with the original output grid, linear or cubic interpolation is used to reconstruct snapshots at the required timestamps, followed by pressure amplitude correction.

B. Anisotropic (Spatially Varying) Permeability Scaling

This approach addresses heterogeneous permeability fields where flow is direction-dependent.

• Permeability-Conformal Coordinate Warp: The authors introduce a mathematical mapping (diffeomorphism) that warps the physical coordinate system. In this warped space, regions of high permeability occupy a larger volume (more grid cells), effectively redistributing the Degrees of Freedom (DoF) to prioritize hydraulically important zones.
• Energy Identity: The warp is designed such that performing DMD on the warped grid is mathematically equivalent to performing a permeability-weighted DMD on the original grid. This ensures the reduced basis captures high-permeability channels more accurately.
• Conservative Resampling: To preserve integral quantities (like total saturation), the authors propose a conservative resampling method that integrates the saturation field over the deformed physical cells before mapping to the warped grid.
• Operator Transformation:
  • Spatial Scaling: The spatial basis ($\Phi$) is scaled using a diagonal scaling matrix ($S_F$) derived from the Jacobian of the warp. The reduced operators are updated via similarity transformations involving Gram matrices ($G_L, G_R$) to avoid forming large dense matrices.
  • Temporal Scaling: Temporal scaling factors ($\Gamma$) are applied to the reduced dynamics matrix ($\tilde{A}$) and input matrix ($\tilde{B}$).
  • Combined Update: The framework provides formulas to apply both spatial and temporal scaling simultaneously, ensuring the physics remain consistent while adapting to new domain properties.

3. Key Contributions

1. Algebraic Update Framework: The paper introduces a novel method to update DMD/DMDc operators directly, eliminating the need for new high-fidelity simulations or retraining when permeability changes.
2. Dual-Strategy Approach: It distinguishes between uniform scaling (handled via time rescaling and eigenvalue modification) and anisotropic scaling (handled via coordinate warping and spatial basis redistribution).
3. Conservative Interpolation: The development of a conservative resampling technique for saturation fields during coordinate warping ensures mass conservation in the reduced-order model.
4. Efficient Implementation: The use of Gram matrices and Schur decompositions allows these updates to be computed efficiently, maintaining the computational speed advantage of the surrogate.

4. Results

The methods were validated on a multiphase flow CO₂ sequestration scenario with permeability multipliers ranging from 40% to 400% of the reference value.

• Uniform Changes:
  • The updated surrogate reduced Mean Absolute Error (MAE) for pressure by 1 to 3 orders of magnitude compared to a non-scaled (static) surrogate.
  • The Pressure Coefficient of Error (PCE) also showed significant improvement (1–2 orders of magnitude).
  • The method was particularly effective in the "far-wellbore" region, where errors dropped by three orders of magnitude, indicating accurate capture of pressure signal propagation.
• Non-Uniform (Heterogeneous) Changes:
  • The anisotropic warp approach maintained low global MAE (approx. $10^{-4}$) for moderate permeability changes, whereas the non-scaled model failed (errors $\approx 10^{-2}$).
  • Even for extreme 400% permeability scaling, the updated model preserved the large-scale structure and temporal evolution of the pressure field (lower PCE), though errors increased in highly nonlinear near-wellbore zones.
• Computational Efficiency: The update process executed hundreds of times faster than full retraining, enabling real-time "what-if" studies.

5. Significance

This work bridges a critical gap in data-driven subsurface modeling. By enabling surrogates to adapt to changing geological properties without retraining, the authors:

• Enable Real-Time Optimization: Facilitate rapid decision-making for well placement, injection scheduling, and pressure management in carbon storage projects.
• Enhance Robustness: Allow surrogates to remain reliable under geological uncertainty (e.g., varying permeability distributions) without the prohibitive cost of re-simulation.
• Preserve Physical Fidelity: The algebraic updates are grounded in the physics of Darcy flow, ensuring that the reduced-order models do not just fit data but respect the underlying physical time-scales and spatial dynamics.

The proposed framework represents a significant step toward practical, adaptive surrogate modeling for complex subsurface engineering applications.