Physics-Informed Neural Network for Elastic Wave-Mode Separation

This paper proposes a computationally efficient Physics-Informed Neural Network (PINN) that solves a scalar Poisson equation to accurately separate P and S elastic wave modes in both homogeneous and non-homogeneous media, outperforming traditional methods by reducing transverse wave leakage.

The Gist

Imagine you are standing in a crowded room where two different groups of people are shouting at the same time. One group is shouting in a low, rumbling bass (like a P-wave, or pressure wave), and the other is shouting in a high-pitched, wiggly squeal (like an S-wave, or shear wave).

In the real world, especially underground where oil, gas, and earthquakes happen, these "voices" get mixed up. When a wave hits a rock layer, a boundary, or a salt dome, it doesn't just keep its original voice; it often splits, bounces, and turns into the other type of wave. This is called mode conversion.

For scientists trying to "see" inside the Earth (like a doctor using an ultrasound), this mix-up is a nightmare. They need to separate the bass from the squeal to understand what the rocks are made of. If they can't separate them, their map of the underground is blurry and wrong.

The Old Way: The Heavy Lifting

Traditionally, scientists have used complex math (called the Helmholtz decomposition) to untangle these waves. Think of this like trying to separate the red and blue threads from a tangled ball of yarn.

The old method requires solving a massive, multi-dimensional puzzle. It's like asking a team of 100 mathematicians to solve a giant 3D jigsaw puzzle where every piece has three different colors. It works, but it's slow, computationally expensive, and requires a lot of brainpower (or computer power) to get right, especially in 3D space.

The New Idea: The Smart Assistant (PINN)

This paper introduces a new, smarter way to do this using Physics-Informed Neural Networks (PINNs).

Think of a PINN not as a human mathematician, but as a super-smart student who has been taught the "Laws of Physics" before they even start studying.

• Standard AI is like a student who memorizes thousands of flashcards of tangled yarn and tries to guess how to untangle them by pattern recognition. It needs a huge library of examples to learn.
• PINN is like a student who knows the rules of how yarn works (it can't stretch infinitely, it has tension, etc.). Because they know the rules, they can figure out how to untangle a new ball of yarn they've never seen before, even with very few examples.

The Big Breakthrough: Simplifying the Math

The authors of this paper made a clever shortcut.

In the old "vector" method, the computer had to track the wave's movement in three directions (up/down, left/right, forward/back) simultaneously. It was like juggling three balls at once.

The authors realized they could solve a simpler scalar equation (a single number) instead.

• The Analogy: Instead of juggling three balls, they found a way to just track the pressure of the air in the room. Once you know the pressure, you can mathematically deduce where the up/down and side-to-side movements are.
• The Result: This reduces the workload significantly. In 3D problems, it cuts the computational cost by a factor of three. It's like going from juggling three balls to just watching one, and still knowing exactly where the other two are going.

What They Tested

They tested this "Smart Student" in two scenarios:

1. The Simple Room (Homogeneous Model): A uniform block of rock. The AI separated the waves perfectly, matching the results of the old, heavy-duty math methods.
2. The Complex Cave (Non-Homogeneous Model): A realistic underground model with a giant salt dome (like the oil fields in Brazil). This is where waves get messy and convert into each other.
   • The Outcome: The AI successfully untangled the waves, even in this messy environment. It did a slightly better job than the old methods at stopping "leakage" (where a bit of the S-wave accidentally stays in the P-wave picture).

The Trade-off

There is one catch. The "Smart Student" (PINN) is currently a bit slower at solving the problem once than the old, highly optimized computer programs.

• Old Method: Fast, but rigid. If the shape of the underground changes, you have to rewrite the code.
• New Method: Slightly slower to start, but incredibly flexible. If the underground has weird shapes or noisy data, the AI adapts without needing a complete rewrite.

The Bottom Line

This paper shows that we can use a "Physics-Smart" AI to untangle the complex mess of underground waves by solving a simpler math problem. It's like replacing a heavy, complicated crane with a nimble, rule-following robot that can handle messy, real-world construction sites just as well as the clean ones.

This means better, clearer images of the Earth's interior, which helps us find resources more safely and understand earthquakes better, all while using less computer power than the traditional heavy-hitters.

Technical Summary

1. Problem Statement

In non-homogeneous elastic media, the interaction of elastic waves with material discontinuities causes mode conversion, where longitudinal P-waves and transverse S-waves scatter into both modes. This phenomenon complicates the accurate interpretation of physical properties in fields like seismology, non-destructive testing, and material characterization.

To address this, wavefield separation is required to decompose the total displacement field ($u$) into pure P and S modes.

• Traditional Approach: The standard method uses Helmholtz decomposition, which typically requires solving a vector Poisson equation ($\nabla^2 w = u$).
• Limitation: Solving a vector equation increases computational complexity and neural network size as spatial dimensions increase (e.g., 2D or 3D), making it less scalable for large-scale problems.

2. Methodology

The authors propose a Physics-Informed Neural Network (PINN) framework that utilizes a scalar formulation of the Poisson equation to separate elastic modes, significantly reducing computational overhead.

A. Mathematical Formulation

Instead of solving the vector Poisson equation, the authors solve an auxiliary scalar Poisson equation:
$$\nabla^2 w = \nabla \cdot u$$
where $u$ is the known elastic wavefield (source term) and $w$ is the scalar potential. The P and S modes are then reconstructed using:

• P-wave: $u_p = \nabla w$
• S-wave: $u_s = u - u_p$

This scalar approach reduces the number of output variables compared to the vector approach (which requires solving for $w_x, w_y, w_z$), offering a computational cost reduction factor of $1/2$ in 2D and $1/3$ in 3D.

B. Neural Network Architecture

• Type: Fully connected feed-forward network with Fourier feature embedding.
• Input: Spatial coordinates $x = (x, z)$.
• Output: The scalar potential $w(x)$.
• Fourier Features: To mitigate spectral bias (the tendency of neural networks to struggle with high-frequency functions), the input coordinates are mapped to a higher-dimensional space using sine and cosine functions:
  $$\gamma(x) = (\sin(2\pi B^T x), \cos(2\pi B^T x))^T$$
  where $B$ contains frequencies sampled from a Gaussian distribution. This allows the network to learn high-frequency wavefield details without increasing trainable parameters.
• Activation Function: Sine function (chosen for its suitability in modeling oscillatory phenomena).

C. Training Strategy

The network is trained by minimizing a composite loss function:
$$\mathcal{L} = \mathcal{L}_{BC} + \mathcal{L}_{PDE}$$

• $\mathcal{L}_{BC}$ (Boundary Condition): Enforces zero boundary conditions on the potential $w$.
• $\mathcal{L}_{PDE}$ (Physics Constraint): Enforces the scalar Poisson equation ($\nabla^2 w - \nabla \cdot u = 0$) at collocation points.
• Optimizer: AdamW with a learning rate decay schedule.

3. Key Contributions

1. Scalar PINN Formulation: The primary contribution is adapting the Helmholtz decomposition to a scalar Poisson equation within a PINN framework. This avoids the complexity of vector-valued outputs, enabling a single-output network architecture.
2. Dimensional Scalability: By reducing the problem to a scalar field, the method offers a scalable reduction in computational cost and network complexity compared to existing vector-based PINN approaches.
3. Single-Network Efficiency: Unlike previous works that used two independent networks for horizontal and vertical components, this method uses a single network to solve the problem, further reducing training and inference costs.
4. Robustness to Noise and Geometry: The method demonstrates the ability to handle irregular geometries and noisy data without requiring substantial reformulation, a known advantage of PINNs over grid-based numerical solvers.

4. Results

The method was validated on two scenarios and compared against a conventional numerical solver based on the Discrete Sine Transform (DST).

A. Homogeneous Model

• Setup: 1 km $\times$ 1 km domain with constant velocity ($v_p=4.0$ km/s, $v_s=2.35$ km/s).
• Performance: The PINN results closely matched the conventional solver with a global Mean Squared Error (MSE) of $1.79 \times 10^{-6}$.
• Observation: Minimal artifacts were observed at boundaries, attributed to the difficulty of representing sharp variations.
• Speed: The PINN inference time (9.36 ms) was slower than the optimized FFT-based solver (0.61 ms) but is considered acceptable given the flexibility gains.

B. Non-Homogeneous (Realistic) Model

• Setup: A complex Brazilian pre-salt model featuring water, marine shale, and a salt body with significant velocity contrasts.
• Performance: The PINN successfully separated modes in the complex medium with a global MSE of $4.94 \times 10^{-4}$.
• Key Finding: The proposed approach substantially suppressed transverse wave leakage (residual S-wave contamination in the P-mode) compared to conventional numerical methods.
• Limitation: As with the homogeneous case, some errors persisted at the physical domain boundaries where wavefronts interact with the edges.

5. Significance and Conclusion

This paper demonstrates that Physics-Informed Neural Networks can effectively solve elastic wave-mode separation problems by leveraging a scalar formulation of the Helmholtz decomposition.

• Efficiency: The scalar approach significantly reduces the dimensionality of the neural network problem, making it more scalable for 3D applications than previous vector-based PINN methods.
• Accuracy: The method achieves accuracy comparable to established numerical solvers while offering superior suppression of mode leakage (crosstalk).
• Future Impact: The approach is transferable to other scenarios involving elastic wave propagation in solids, such as seismic imaging, non-destructive testing, and material characterization, particularly where complex geometries or limited data availability make traditional grid-based methods challenging.

In summary, the authors provide a computationally efficient, single-network solution for elastic wave separation that balances physical accuracy with the flexibility of deep learning.