A Multi-Fidelity Parametric Framework for Reduced-Order Modeling using Optimal Transport-based Interpolation: Applications to Diffused-Interface Two-Phase Flows

This paper presents a non-intrusive, multi-fidelity reduced-order modeling framework that utilizes Optimal Transport-based displacement interpolation to efficiently correct low-fidelity models and construct accurate parametric surrogates for complex, nonlinear two-phase flow simulations.

The Gist

Imagine you are trying to predict the weather. You have two tools:

1. The Supercomputer: It's incredibly accurate but takes weeks to run a single forecast. It's too slow to use for real-time decisions.
2. The Weather App: It's fast and gives you a quick guess, but it's often wrong because it uses simple rules and low-resolution data.

This paper introduces a clever "translator" that combines the speed of the Weather App with the accuracy of the Supercomputer. It does this using a mathematical concept called Optimal Transport, which the authors use to create a "smart shortcut" for complex physics problems.

Here is a breakdown of their method using simple analogies:

1. The Problem: Moving Shapes are Hard to Predict

In physics, many problems involve things moving and changing shape, like a drop of ink spreading in water or two fluids mixing.

• The Old Way (Linear Interpolation): Imagine trying to guess what a moving cloud looks like in the middle of the day by just taking a photo of it at 9 AM and 3 PM and blending them together. If the cloud moved 10 miles, a simple blend would just create a blurry, ghostly mess of the cloud in both places. It fails because it doesn't understand movement.
• The New Way (Optimal Transport): Instead of blending, this method asks: "If I have a pile of sand at point A and I want it at point B, what is the most efficient path for every grain of sand to take?" It tracks the movement of the "stuff" (the mass) rather than just the numbers. It's like watching a dance and predicting the next step based on the flow of the dancers, rather than just averaging their positions.

2. The Multi-Fidelity Trick: Fixing the Cheap Model

The authors realized that while the "Supercomputer" (High-Fidelity) is too slow, the "Weather App" (Low-Fidelity) is fast.

• The Strategy: They run the fast, cheap model. Then, they take a few snapshots from the slow, expensive model to see where the cheap model is wrong.
• The Correction: They don't just guess the error; they use the "sand moving" math (Optimal Transport) to figure out how the error moves and changes over time. They then "patch" the cheap model with this moving correction.
• The Result: You get a prediction that runs as fast as the cheap model but looks almost exactly like the expensive one.

3. The Parametric Twist: Learning to Adapt

Real-world problems change. Maybe the wind speed changes, or the temperature of the fluids is different.

• The Challenge: Usually, if you change a setting (like the wind speed), you have to run the expensive simulation all over again.
• The Solution: The authors created a two-step dance:
  1. Step 1 (Parameter Space): They use the "sand moving" math to guess what the expensive simulation would look like for a new wind speed, based on simulations they already did for nearby wind speeds.
  2. Step 2 (Time): They then use the "Multi-Fidelity" trick from above to fix the cheap model for this new scenario.
• The Result: They can predict the outcome for any setting (wind speed, temperature, etc.) almost instantly, without ever running the expensive simulation for that specific setting.

4. The Real-World Test: Mixing Fluids

To prove this works, they tested it on Two-Phase Flows. Imagine oil and water mixing, or a bubble rising in a liquid. These are notoriously difficult to simulate because the boundary between the two fluids is sharp, moves fast, and breaks apart into tiny droplets.

• The Test: They simulated a bubble being stretched by a swirling wind (Rider-Kothe vortex) and a heavy fluid falling through a lighter one (Rayleigh-Taylor instability).
• The Outcome:
  • Traditional methods (like the "blending" approach) turned the sharp bubble into a blurry, unrecognizable blob.
  • Their new method kept the bubble sharp, tracked its movement perfectly, and even predicted how it would break apart into tiny droplets.
  • They achieved this with 99% less computing power than the full simulation.

Summary Analogy

Think of the High-Fidelity Model as a master painter creating a masterpiece, but it takes them a month.
Think of the Low-Fidelity Model as a child's crayon drawing. It's fast, but it lacks detail.

The authors' framework is like a smart art teacher.

1. The teacher looks at the child's drawing (Low-Fidelity).
2. The teacher glances at a few of the master's paintings (High-Fidelity snapshots) to see where the child went wrong.
3. Instead of just telling the child "draw it better," the teacher uses a special map (Optimal Transport) to show the child exactly how to move their crayon strokes to match the master's flow and movement.
4. The result? The child produces a masterpiece in minutes that looks like it took a month to paint.

Why does this matter?
This technology allows engineers and scientists to run complex simulations (like designing better jet engines, predicting oil spills, or modeling blood flow) in seconds or minutes instead of days, making advanced physics accessible for real-time decision-making.

Technical Summary

Here is a detailed technical summary of the paper "A Multi-Fidelity Parametric Framework for Reduced-Order Modeling Using Optimal Transport-Based Interpolation: Applications to Diffused-Interface Two-Phase Flows."

1. Problem Statement

The paper addresses the computational challenges associated with simulating advection-dominated, nonlinear partial differential equations (PDEs), specifically focusing on two-phase flows with moving interfaces (diffuse-interface methods).

• High-Fidelity (HF) Bottleneck: Accurate simulations (e.g., using Finite Element or Finite Volume methods) require high-dimensional discretizations ($N_h \gg 1$), making them computationally prohibitive for real-time applications, uncertainty quantification, or design optimization.
• Limitations of Traditional ROMs: Standard Reduced-Order Modeling (ROM) techniques, such as Proper Orthogonal Decomposition (POD), rely on linear subspaces. These fail for advection-dominated problems because the solution manifold has a slow decay of the Kolmogorov $n$-width. Moving features (like interfaces) cause "Gibbs phenomena" and require an impractically large number of modes to capture the physics accurately.
• Data Scarcity and Cost: Generating sufficient HF snapshots for training ROMs is expensive. Furthermore, many practical problems involve multi-fidelity scenarios (where cheap Low-Fidelity/LF models exist but are inaccurate) and parametric dependencies (where system behavior changes with physical parameters like material properties or boundary conditions).

2. Methodology

The authors propose a data-driven, non-intrusive framework that leverages Optimal Transport (OT) theory, specifically Displacement Interpolation (DI), to overcome the limitations of linear ROMs. The methodology is built on three main pillars:

A. Optimal Transport and Displacement Interpolation

Instead of linearly interpolating solution snapshots (which smears moving features), the method treats solution fields as probability distributions.

• Wasserstein Geometry: It utilizes the Wasserstein distance ($W_2$), which measures the "cost" of transporting mass from one distribution to another. This metric is sensitive to spatial shifts, making it ideal for moving interfaces.
• Geodesic Paths: The method constructs geodesic paths between snapshots in the Wasserstein space. This allows for the generation of synthetic intermediate states that preserve the physical movement of features rather than just blending values.
• Entropic Regularization: To make the OT problem computationally tractable, the authors use Sinkhorn algorithms with entropic regularization, enabling efficient parallelization (e.g., on GPUs).

B. Multi-Fidelity OT-ROM (MF-OT-ROM)

This strategy corrects a computationally cheap Low-Fidelity (LF) model using sparse High-Fidelity (HF) data.

• Residual Interpolation: Instead of interpolating the full solution, the framework interpolates the residual field ($r = u_{HF} - u_{LF}$).
• Process:
  1. Compute residuals at sparse HF checkpoints.
  2. Decompose residuals into positive and negative parts (to satisfy OT non-negativity constraints).
  3. Apply displacement interpolation in time to the residual components.
  4. Reconstruct the final HF approximation: $u_{approx} = u_{LF} + r_{interp}$.
• Advantage: This captures the movement of the discrepancy between LF and HF models, which is often concentrated at the moving interface.

C. Parametric Multi-Fidelity OT-ROM (PMF-OT-ROM)

This extends the framework to handle parametric dependencies ($\mu$) in addition to time ($t$).

• Two-Level Hierarchical Strategy:
  1. Parametric Level: Generate synthetic HF checkpoints for an unseen parameter $\mu^*$ by performing displacement interpolation between HF snapshots of neighboring training parameters.
  2. Temporal Level: Compute the residual between these synthetic HF checkpoints and the LF solution at $\mu^*$, then interpolate this residual in time using the MF-OT-ROM strategy.
• Alternative Strategy: The paper also discusses interpolating the HF solution directly in parameter space and then correcting the LF solution, noting that the residual-based approach often offers better stability.

3. Key Contributions

1. Novel Frameworks: Introduction of MF-OT-ROM and PMF-OT-ROM, which combine multi-fidelity modeling with optimal transport-based interpolation.
2. Handling Nonlinearity: Demonstrating that displacement interpolation in Wasserstein space effectively handles the slow decay of the Kolmogorov $n$-width in advection-dominated problems, preserving sharp interfaces and boundedness (values staying within physical limits, e.g., $0 \le \phi \le 1$).
3. Non-Intrusive Approach: The method does not require modifying the governing equations or the solver code; it operates purely on the output data (snapshots).
4. Physical Consistency: Unlike linear interpolation which creates unphysical intermediate values (smearing), the OT-based approach generates physically plausible synthetic snapshots that follow geodesic paths.

4. Numerical Results and Validation

The framework was tested on two challenging benchmarks using the conservative Allen-Cahn equation coupled with a five-equation compressible two-phase flow model:

A. Rider-Kothe Vortex (Advection Test)

• Setup: A circular bubble deformed by a shear flow and then restored.
• Results:
  • POD vs. DI: POD projection resulted in significant smearing of the interface and unphysical intermediate values. Displacement Interpolation (DI) accurately preserved the sharp interface and boundedness.
  • Multi-Fidelity: Correcting a coarse mesh (LF) using DI on residuals significantly reduced errors compared to the uncorrected LF solution. With sufficient checkpoints ($N_c=80$), the corrected solution was nearly indistinguishable from the HF solution.
  • Parametric: Successfully predicted solutions for unseen initial bubble radii.

B. Rayleigh-Taylor Instability (Complex Breakup)

• Setup: Two fluids of different densities interacting under gravity, leading to complex interface breakup.
• Results:
  • Dynamic Complexity: The method successfully captured the acceleration of the interface and the formation of complex ligaments and droplets.
  • Interfacial Area: A critical metric for two-phase flows. The uncorrected LF model under-predicted the total interfacial area due to lack of resolution. The OT-corrected model converged toward the HF reference, accurately predicting the surface area growth.
  • Parametric Sensitivity: The framework handled varying Atwood numbers (density ratios). While accuracy decreased slightly for highly complex breakup scenarios (high Atwood numbers) with limited data, the large-scale dynamics and interface positions were correctly predicted.

5. Significance and Impact

• Bridging the Gap: The work provides a robust bridge between expensive, high-fidelity diffuse-interface simulations and the need for fast, real-time surrogate models.
• Overcoming Linear Limitations: It offers a concrete solution to the "moving feature" problem that has long plagued linear ROMs, without resorting to complex deep learning architectures that require massive datasets and lack interpretability.
• Efficiency: By leveraging cheap LF models and correcting them with sparse HF data via OT, the method drastically reduces the computational cost of exploring parameter spaces and performing uncertainty quantification for complex multi-physics problems.
• Future Directions: The authors suggest adaptive sampling for checkpoints, barycentric interpolation for higher-dimensional parameter spaces, and the use of machine learning to map parameters directly to OT plans.

In conclusion, this paper presents a mathematically rigorous and practically effective framework for reduced-order modeling of complex, moving-interface flows, demonstrating that Optimal Transport is a superior geometric tool for handling the nonlinear dynamics inherent in two-phase flow simulations.