Revisiting Conservativeness in Fluid Dynamics: Failure of Non-Conservative PINNs and a Path-Integral Remedy

This paper demonstrates that standard non-conservative Physics-Informed Neural Networks (PINNs) fail to capture correct shock speeds in unsteady fluid dynamics due to violations of Rankine–Hugoniot conditions, and proposes a path-integral framework based on DLM theory to successfully restore physical fidelity in primitive-variable formulations.

The Gist

The Big Picture: The "Traffic Jam" Problem

Imagine you are trying to simulate how cars move on a highway, specifically when a sudden traffic jam (a shock wave) forms.

In physics, there are two main ways to write the rules for how these cars move:

1. The "Conservative" Way: You count the total number of cars in every section of the road. If a car leaves one section, it must enter the next. Nothing is lost or created out of thin air. This is like balancing a checkbook perfectly.
2. The "Non-Conservative" Way: You describe how fast each individual car is moving and how the speed changes from one car to the next. It feels more intuitive (like talking about speed limits), but it doesn't strictly track the total "stuff" (cars) moving through the system.

The Problem: When a traffic jam happens (a shock), the "Non-Conservative" way often gets the math wrong. It might predict the jam moving at 10 mph when it should be moving at 50 mph. In fluid dynamics (like air hitting a supersonic jet), getting the speed of the shock wrong means your simulation is physically impossible.

The New Contender: PINNs (The "Smart Predictor")

The paper investigates a new type of solver called Physics-Informed Neural Networks (PINNs). Think of a PINN not as a calculator, but as a super-smart student trying to learn the laws of physics by guessing and checking.

• How it works: The student is given a set of rules (the equations) and some boundary conditions (start and end points). The student makes a guess, checks how far off they are from the rules, and adjusts their brain (weights) to get closer.
• The Goal: Can this "student" learn to predict shock waves correctly, even if we give it the "Non-Conservative" (intuitive) rules instead of the strict "Conservative" ones?

The Experiment: What Happened?

The authors tested this student on three different scenarios:

1. The Smooth Water (Shallow Water Equations)

• Scenario: Ripples in a pond with a bumpy bottom. No sudden crashes, just smooth waves.
• Result: The "Smart Student" (PINN) did a great job! It didn't matter if they used the strict rules or the intuitive rules. The student learned to keep the total amount of water (mass) correct just by being told to minimize errors globally.
• Analogy: If you are pouring water into a bowl, as long as you don't spill, it doesn't matter if you measure the flow by the cup or by the second. The result is the same.

2. The Traffic Jam (Burgers' Equation)

• Scenario: A sudden stop where cars pile up instantly.
• Result:
  • Old School Math (Numerical Methods): If you use the "Non-Conservative" rules, the traffic jam stops moving entirely. It freezes in place. Wrong!
  • The Smart Student (PINN): Surprisingly, the student learned the correct speed even with the "Non-Conservative" rules!
  • Why? Because the student was smooth and continuous. It didn't have the "jagged edges" that confused the old math methods.

3. The Supersonic Crash (Sod Shock Tube / Euler Equations)

• Scenario: A high-speed explosion where air is compressed violently. This is the real test.
• Result: This is where the paper found a major failure.
  • When the student tried to learn the "Non-Conservative" rules for this high-speed crash, it failed. It predicted the shock wave moving at the wrong speed.
  • The Culprit: To make the math work, the student had to add a little bit of "friction" (artificial viscosity) to smooth out the crash. But this friction introduced a hidden "ghost force" that pushed the shock wave in the wrong direction. The student was stable, but physically wrong.

The Solution: The "Path-Integral" Detour

The authors realized that the "Non-Conservative" rules were missing a crucial piece of information: How you get from Point A to Point B matters.

Imagine you are hiking from a valley to a mountain peak.

• Conservative View: You only care about the height difference between the start and finish.
• Non-Conservative View: You care about the steepness of the path right now.
• The Problem: If there is a cliff (a shock), "steepness right now" is undefined. You can't calculate it.

The Fix: The Path-Integral Remedy
The authors taught the student a new trick called Path-Conservative Learning.

• Instead of just looking at the start and end points, the student is forced to imagine a specific path connecting the two states (like drawing a line on a map between the valley and the peak).
• The student must calculate the "cost" of the journey along that specific line.
• The Result: By forcing the student to respect this path, the "ghost force" disappears. The student finally learns the correct shock speed, even while using the intuitive "Non-Conservative" rules.

The Takeaway

1. Conservation is King (usually): For high-speed, crashing flows, you usually must use the strict "Conservative" math to get the right answer.
2. PINNs are tricky: Even smart AI models can get the physics wrong if the underlying math doesn't account for how things change across a crash.
3. The Bridge: You can use the more intuitive "Non-Conservative" math with AI, but you have to add a special "Path-Integral" rule. This rule acts like a safety net, ensuring that even though you are using simple rules, the AI still respects the complex laws of physics when things break or crash.

In a nutshell: The paper shows that while "intuitive" math is easier to write, it's dangerous for high-speed crashes. However, by teaching AI to "walk the path" between states, we can make the intuitive math safe and accurate again.

Technical Summary

1. Problem Statement

The paper addresses a fundamental dilemma in Computational Fluid Dynamics (CFD) and Physics-Informed Neural Networks (PINNs): the choice between conservative (divergence form) and non-conservative (primitive variable) formulations of governing equations.

• The Core Issue: While non-conservative forms (using primitive variables like velocity $u$, pressure $p$, and density $\rho$) are often more intuitive and computationally efficient for smooth flows, they historically fail to capture correct shock speeds in discontinuous flows. This is because they do not inherently satisfy the Rankine–Hugoniot (RH) jump conditions, leading to erroneous shock propagation.
• The PINN Context: The authors investigate whether modern PINNs, specifically those using Adaptive Weight and Viscosity (PINNs-AWV), can overcome this limitation. They hypothesize that while PINNs might succeed in steady-state or scalar problems, they may fail in unsteady, shock-dominated systems (like the Euler equations) due to the non-vanishing source terms introduced by viscous regularization, which violate the RH conditions.
• The Gap: There is a lack of rigorous frameworks that allow non-conservative PINNs to accurately simulate transient, high-speed flows with shocks without reverting to conservative variables.

2. Methodology

The study employs a multi-faceted approach, comparing classical numerical methods against PINN architectures across various test cases.

A. Theoretical Framework

• Hierarchy of Conservativeness: The authors define three levels:
  1. Mathematical: The PDE is in divergence form.
  2. Numerical: The discretization preserves discrete integral balances (flux consistency).
  3. Physical: The solution preserves fundamental principles (positivity, entropy, equilibrium states).
• Path-Conservative Theory: To remedy non-conservative failures, the paper utilizes the Dal Maso–LeFloch–Murat (DLM) theory. This approach defines non-conservative products across discontinuities by integrating the system Jacobian along a specific path in state space ($\Phi(s)$) connecting left and right states.

B. PINN Architecture (PINNs-AWV)

The authors use an Adaptive Weight and Viscosity architecture where:

• Adaptive Viscosity: A sub-network learns the optimal artificial viscosity ($\nu$) to stabilize shocks without excessive smearing.
• Adaptive Weights: Loss weights are adjusted based on gradient norms to prevent PDE residuals from dominating the training near discontinuities.
• Path-Integral PINN (PI-PINN): A novel modification where a Path-Integral Loss is added. This loss term enforces the generalized RH condition by calculating the path integral of the Jacobian and comparing it to the flux jump:
  $$\mathcal{L}_{path} = \left\| \left( \int_0^1 A(\psi(\tau)) d\tau \right) \Delta V - \Delta F \right\|^2$$

C. Test Cases

1. Burgers' Equation: A scalar conservation law used to test mathematical conservativeness and weak solution behavior.
2. Shallow Water Equations (SWE): Used to test "well-balanced" properties and conservation in the presence of variable bottom topography.
3. Sod Shock Tube (1D Euler): A transient, shock-dominated problem to evaluate shock speed accuracy in unsteady systems.
4. Supersonic Flow over a Wedge (2D Steady Euler): A steady-state problem to test if shock location accuracy can be achieved without shock speed errors.

3. Key Contributions

1. Identification of PINN Failure in Unsteady Non-Conservative Systems:
   The paper demonstrates that while PINNs-AWV can solve scalar (Burgers) and steady-state Euler equations using non-conservative forms, they fail to predict correct shock speeds in unsteady Euler systems (Sod shock tube). The failure is attributed to the fact that viscous regularization in primitive variables introduces non-conservative source terms that do not vanish as viscosity $\to 0$, leading to different weak limits than the true physical solution.

2. Development of Path-Integral PINNs (PI-PINN):
   The authors propose a rigorous remedy by integrating DLM theory into the PINN loss function. By explicitly enforcing the path-consistent jump condition, the PI-PINN successfully recovers correct shock speeds in non-conservative frameworks, bridging the gap between primitive-variable intuition and physical conservation laws.

3. Comparative Analysis of Conservation Levels:
   The study clarifies that "conservation" is not binary. It shows that:

   • Conservative discretizations are essential for correct shock speeds in unsteady flows.
   • Non-conservative schemes can work for steady flows or scalar equations if the solution is smooth.
   • Global mass constraints in PINNs can mitigate errors in steady flows but are insufficient for unsteady shock dynamics without path-consistency.

4. Results

• Scalar Problems (Burgers): Both conservative and non-conservative PINNs (with adaptive viscosity) successfully captured the correct shock speed. The non-conservative form worked because the identity $u u_x = (u^2/2)_x$ holds for smooth neural approximations.
• Shallow Water Equations:
  • Numerical Methods: Non-conservative FDM failed to preserve mass and equilibrium (lake-at-rest). Conservative FVM preserved mass but required well-balancing for equilibrium.
  • PINNs: Interestingly, PINNs (even non-conservative ones) performed well in this smooth, steady-like scenario by enforcing a global mass constraint in the loss function, outperforming standard non-conservative FDM.
• Sod Shock Tube (Unsteady Euler):
  • Failure: Standard non-conservative PINNs (even with adaptive viscosity) predicted a shock speed that deviated significantly from the exact solution over time.
  • Success: The PI-PINN (Path-Integral) formulation successfully recovered the exact shock location and speed, matching the performance of conservative PINNs and numerical solvers.
• Supersonic Wedge (Steady Euler):
  • Both conservative and non-conservative PINNs correctly predicted the shock location.
  • However, the non-conservative PINN exhibited a higher mass defect ($3.39 \times 10^{-3}$) compared to the conservative PINN ($3.62 \times 10^{-4}$), highlighting that while steady shock location can be found without strict conservation, global physical consistency is compromised.

5. Significance and Conclusion

The paper establishes that conservativeness is a hierarchical requirement (Mathematical $\subset$ Numerical $\subset$ Physical) that cannot be bypassed in high-speed, transient simulations.

• For Classical CFD: It reinforces the necessity of conservative schemes for shock-capturing but validates the utility of path-conservative methods for complex multiphysics systems where conservative forms are unavailable.
• For Machine Learning in CFD: It provides a critical warning: simply applying PINNs to non-conservative primitive variables is insufficient for unsteady shock problems.
• The Path Forward: The proposed Path-Integral PINN (PI-PINN) offers a rigorous mathematical bridge. It allows researchers to utilize the flexibility and intuition of primitive variables in neural networks while guaranteeing physical fidelity (correct shock speeds) through path-consistent loss functions. This enables the development of "primitive-variable" solvers for transient, high-speed compressible flows that were previously thought to require conservative formulations.

In summary, the paper proves that while non-conservative PINNs fail in unsteady shock regimes due to regularization artifacts, this failure can be mathematically remedied by embedding the Dal Maso–LeFloch–Murat path-integral framework directly into the neural network's training objective.