Local Well-Posedness of a Modified NSCH-Oldroyd System: PINN-Based Numerical Illustrations

This paper establishes the local well-posedness of a diffusion-enhanced modified Navier-Stokes-Cahn-Hilliard-Oldroyd system motivated by thrombus modeling and validates the model through Physics-Informed Neural Network (PINN) simulations that utilize Metropolis-Hastings sampling based on energy decay.

The Gist

The Big Picture: Modeling a Blood Clot

Imagine your bloodstream is a busy highway. Sometimes, a traffic jam forms—a blood clot (thrombus). This isn't just a static pile of cars; it's a messy, shifting mess where the "cars" (blood cells) interact with the "road" (fluid dynamics) and the "road conditions" (elasticity) change constantly.

Scientists want to build a computer simulation to predict how these clots form, move, and dissolve. This paper does two main things:

1. Mathematically: It proves that a new, improved set of rules for this simulation actually works and won't break down.
2. Computationally: It uses a special type of AI (called a PINN) to run these simulations and show that the new rules produce stable, realistic results.

---

Part 1: The "Leaky" Old Model vs. The "Reinforced" New Model

The Problem with the Old Model:
Think of the previous mathematical model for blood clots like a house built with a weak foundation. It worked okay for simple rooms, but if you tried to simulate a "shockwave" (like a clot forming rapidly), the walls would crumble.

• The Flaw: In the old model, the "elasticity" (how stretchy the clot is) was set to zero in the pure blood areas. This is physically weird. It's like saying a rubber band has no stretchiness until you pull it, but in reality, the material properties should exist everywhere, just at different strengths.
• The Missing Ingredient: The old model also lacked a "smoothing" mechanism (diffusion) for the deformation variable. Imagine trying to draw a sharp line with a marker that keeps bleeding ink everywhere; the lines get blurry and unstable. The math needed a way to keep those lines crisp.

The Solution (The Modified System):
The author, Woojeong Kim, fixed the model by adding two things:

1. Generalized Elasticity: Instead of turning elasticity "off" in the blood, they made it a smooth transition. Now, the model understands that blood has some elasticity, and clots have more. It's like upgrading from a lightbulb that is either "On" or "Off" to a dimmer switch that allows for a smooth gradient.
2. Diffusion (The Stabilizer): They added a "diffusion term" to the deformation equation. Think of this as adding a little bit of friction or a "shock absorber" to the system. It prevents the simulation from vibrating wildly or exploding when things change too fast.

The Result:
The paper proves mathematically (using a concept called Local Well-Posedness) that with these new rules, the system is stable. It guarantees that if you start with a specific initial state, the simulation will produce a unique, sensible result for a certain amount of time, rather than turning into nonsense.

---

Part 2: The AI Detective (PINNs)

Now that the rules are fixed, how do we solve them? These equations are so complex that traditional calculators (supercomputers using standard methods) struggle with the sharp, jagged edges where blood meets the clot.

Enter PINNs (Physics-Informed Neural Networks).

• The Analogy: Imagine trying to learn a song. A traditional method is to read the sheet music note-by-note (grid-based). A PINN is like a musician who listens to the song and tries to hum it back, but they are forced to follow the laws of music theory (the physics equations) while they learn.
• The Challenge: The "song" here has a very sharp, high-pitched note (the shock interface where the clot forms). Standard AI often misses these sharp notes, smoothing them out too much.

The "Metropolis-Hastings" Trick:
To catch these sharp notes, the author used a clever sampling technique called Metropolis-Hastings.

• The Metaphor: Imagine you are a detective looking for a hidden treasure. If you walk randomly, you might miss it. But if you have a "heat map" that tells you where the energy (or the treasure) is highest, you can focus your steps there.
• The Application: The AI looks at the "Total Energy" of the system. It knows the most chaotic, interesting action happens where the energy is changing rapidly (the shock interface). So, instead of sampling the whole domain evenly, the AI "throws more darts" at the high-energy zones. This allows it to resolve the sharp details of the clot formation much better.

---

Part 3: The Experiments (The "What If" Scenarios)

The author ran several simulations to test the new model:

1. The Static Clot (Case A): A clot that sits still. The model confirmed it stays still, proving the math is stable.
2. The Diffusive Clot (Case B): A clot that spreads out like ink in water. By tweaking the "viscosity" (thickness) and "permeability" (how easily fluid flows through), they showed the model could simulate a clot dissolving or spreading naturally.
3. The Two-Clot Dance (Case C): Two clots start far apart and slowly merge into one. The new model, with its better energy handling, showed them merging cleanly, whereas the old model might have gotten confused at the boundary.
4. The Thin Interface (Case D): This is the hardest case. Imagine a clot with a razor-thin edge. This is where the "diffusion" and the "AI sampling" really shined. Without the new tricks, the simulation would crash or look blurry. With them, the AI captured the thin edge accurately.

The Takeaway

This paper is a bridge between rigorous math and smart AI.

• The Math: Proves that if you add a little bit of "friction" (diffusion) and fix the "elasticity" settings, the physics of blood clots can be described by a stable, solvable system.
• The AI: Shows that by teaching a neural network to focus its attention on the most energetic parts of the problem, we can simulate complex biological events (like clotting) with high precision.

In short: The author built a better, more stable engine for simulating blood clots and taught an AI how to drive it through the roughest terrain without crashing.

Technical Summary

1. Problem Statement

The paper addresses the mathematical and numerical challenges associated with modeling thrombus (blood clot) dynamics using a coupled Navier–Stokes–Cahn–Hilliard–Oldroyd (NSCH–Oldroyd) system.

• Context: Previous models (specifically referencing Kim et al. [6]) successfully established local well-posedness but lacked a diffusion term in the deformation variable ($F$).
• Deficiencies of Previous Models:
  • Analytical Instability: The absence of diffusion in the deformation equation made the system less robust for deriving a priori estimates and controlling higher-order terms.
  • Physical Inconsistency: In the original model, viscoelasticity was zero in the pure blood region ($\phi=1$), leading to zero elastic energy in blood, which is physically unrealistic.
  • Numerical Difficulty: Simulating sharp phase-field interfaces (shock regions) where the solution exhibits rapid spatial gradients is notoriously difficult for standard numerical methods, including Physics-Informed Neural Networks (PINNs).

2. Methodology

A. Mathematical Formulation (The Modified System)

The author proposes a diffusion-enhanced modification of the governing equations to address the stability issues while preserving the dissipative energy structure.

• Governing Equations: The system couples the incompressible Navier-Stokes equations (velocity $u$, pressure $p$), the Cahn-Hilliard equation (phase field $\phi$), and the Oldroyd-B type equation for the deformation gradient ($F$).
• Key Modifications:
  1. Generalized Viscoelasticity: Replaced the binary parameter $\lambda_e$ with a continuous function $\nu(\phi)$, ensuring non-zero viscoelasticity in both blood and thrombus regions.
  2. Diffusion Term: Added a diffusion term $k(\nu(\phi)\Delta F + 2\nabla\nu(\phi)\cdot\nabla F)$ to the deformation equation. This stabilizes the system analytically and numerically.
  3. Boundary Conditions: Neumann boundary conditions are imposed on $\phi$, $F$, and their derivatives to ensure energy dissipation.

B. Theoretical Analysis (Local Well-Posedness)

The paper provides a rigorous proof of local well-posedness (existence and uniqueness of strong solutions) for the modified system in dimensions $d=2,3$.

• Energy Estimates: The author derives a total energy functional $E(x,t)$ comprising kinetic, mixed (phase-field), and elastic energies. It is proven that the system satisfies a dissipative energy law ($\frac{d}{dt}E \leq 0$).
• A Priori Estimates: High-order Sobolev estimates are established for $u$, $\phi$, and $F$ and their time derivatives. This involves:
  • Using the Faedo-Galerkin method to construct approximate solutions.
  • Deriving bounds for higher-order norms (e.g., $H^3$ for velocity, $H^4$ for phase field) using Sobolev embeddings and interpolation inequalities.
  • Handling the non-linear coupling terms via careful estimation of the deformation gradient and chemical potential.
• Existence & Uniqueness:
  • Existence: Proven via the compactness of the Galerkin approximations using the Aubin-Lions lemma.
  • Uniqueness: Proven by analyzing the difference between two potential solutions and applying Gronwall's inequality to show the difference vanishes.

C. Numerical Methodology (PINNs)

To illustrate the modified system, the author employs Physics-Informed Neural Networks (PINNs) with advanced training strategies.

• Architecture: A deep neural network (10 layers, 128 neurons) mapping $(x, y, t)$ to $(u_1, u_2, \phi, F_{ij}, p)$.
• Training Strategy:
  • Window-Sweeping Method: The time domain is split into small overlapping intervals. Transfer learning is used to pass solutions from the previous interval to the next, ensuring stability in time-marching.
  • Auto-Adaptive PINN (AA-PINN): To resolve the sharp gradients at the thrombus-blood interface, the paper utilizes Metropolis-Hastings sampling. Instead of uniform sampling, collocation points are sampled based on the total energy density. This concentrates computational resources in regions of high energy change (shock interfaces).
• Loss Function: A composite loss function includes PDE residuals, boundary conditions, initial conditions, and a "backward compatibility" term for the overlapping time windows.

3. Key Contributions

1. Theoretical Advancement:

   • Construction of a diffusion-enhanced NSCH-Oldroyd model that preserves the physical energy dissipation structure.
   • Proof of local well-posedness for strong solutions, establishing a rigorous mathematical foundation that was missing in previous thrombus models.
   • Demonstration that the added diffusion term allows for higher-order regularity estimates, making the system analytically tractable.

2. Numerical Innovation:

   • Development of a PINN framework capable of simulating complex fluid-structure interactions with viscoelasticity.
   • Implementation of energy-adaptive sampling (via Metropolis-Hastings) to overcome the "gradient explosion" problem at phase interfaces, significantly reducing $L_\infty$ errors in challenging cases (e.g., thin interfaces or multiple thrombi).
   • Validation of the window-sweeping technique for long-time simulations of stiff PDE systems.

4. Results

The paper presents numerical illustrations for several representative thrombus cases:

• Case A (Static Thrombus): Serves as a baseline, showing stable, static behavior with minimal energy dissipation.
• Case B (Diffusive Thrombus): Demonstrates how varying the coefficient $\lambda$ affects the diffusion rate of the thrombus, showing smoother phase transitions.
• Case C (Two Thrombi): Simulates the merging of two clots. The results show that higher $\lambda\gamma$ (double-well potential strength) leads to distinct phase separation and effective clot gathering.
• Case D (Thin Interface): The most challenging case with a small interface thickness ($h=0.035$).
  • Without AA-PINN: High error in capturing the initial configuration.
  • With AA-PINN: The $L_\infty$ error decreased by 55.29%, and the sampling distribution successfully concentrated on the shock interface, capturing the sharp gradients accurately.

Benchmarking:

• The simulations demonstrated mass conservation and divergence-free errors within acceptable bounds.
• Energy dissipation was tracked over time, confirming the system evolves toward a static equilibrium as predicted by the theoretical energy law.

5. Significance

• Mathematical Rigor: The paper bridges the gap between physical modeling and rigorous analysis for thrombus dynamics, providing a stable framework for future data-assimilation applications.
• Computational Efficiency: By integrating Auto-Adaptive PINNs with energy-based sampling, the study offers a solution to one of the most persistent challenges in phase-field modeling: resolving sharp interfaces without excessive computational cost or numerical instability.
• Clinical Relevance: The improved model and simulation techniques provide a more reliable tool for understanding thrombus formation and evolution, which is critical for developing diagnostic tools and treatment strategies for cardiovascular diseases.
• Future Directions: The work lays the groundwork for parameter calibration and data-driven diagnostics, suggesting that such models can eventually be used to infer physical parameters from clinical imaging data.