Physics-Informed Neural Network Approach for Surface Wave Propagation in Functionally Graded Magnetoelastic Layered Media

This paper proposes and validates a physics-informed neural network (PINN) framework, benchmarked against an analytical solution, to accurately model the dispersion of SH-waves in a pre-stressed, gravity-influenced, functionally graded magnetoelastic layered composite structure.

The Gist

Imagine the Earth's crust and the materials we build with aren't just solid blocks, but more like layered cakes or stacked blankets. Sometimes, these layers are made of special materials that change their properties as you go deeper (like a cake that gets sweeter the lower you go). These are called Functionally Graded Materials.

Now, imagine you want to know how a ripple (a wave) travels through this layered cake. In the real world, this happens with earthquakes (seismic waves) or when testing new materials for airplanes. But calculating exactly how these ripples move is incredibly hard because the layers are different, they are under pressure (pre-stressed), and they interact with magnetic fields and gravity.

Traditionally, scientists have to use heavy, complex math (like solving a giant puzzle with thousands of pieces) to figure this out. This paper introduces a new, smarter way to solve the puzzle using Artificial Intelligence (AI).

Here is the breakdown of what the researchers did, using simple analogies:

1. The Problem: A Noisy, Stressed-Out Cake

The researchers are studying a specific type of wave called an SH-wave (think of it as a "side-to-side" shiver, like shaking a rug).

• The Setup: They have a top layer (the "cake") sitting on a bottom layer (the "floor").
• The Complications:
  • Gravity: The layers are heavy, pulling down.
  • Pressure: The layers are already squished (pre-stressed).
  • Magnetism: The materials react to magnetic fields.
  • Grading: The materials get stiffer or softer as you go deeper.

Trying to calculate how a wave moves through this messy, changing environment using old-school math is like trying to predict the path of a pinball in a machine where the bumpers are constantly moving and changing shape.

2. The Old Way vs. The New Way

• The Old Way (Analytical Solution): Scientists derive a massive, complicated equation. It's like writing a 50-page manual on how to bake the cake perfectly. It works, but it's slow, rigid, and if you change one ingredient (like the magnetic field), you have to rewrite the whole manual.
• The New Way (PINN - Physics-Informed Neural Network): Instead of writing a manual, the researchers built a smart robot student (the Neural Network).
  • The Twist: Usually, AI needs thousands of photos of ripples to learn. But here, they didn't give the robot photos. Instead, they gave the robot the laws of physics (the rules of how waves must behave).
  • The Training: The robot tries to guess the answer. Every time it guesses wrong, the computer checks: "Did you follow the laws of physics? Did you respect gravity? Did you respect the magnetic pull?" If the robot breaks a rule, it gets a "penalty" (loss).
  • The Goal: The robot keeps adjusting its internal settings until it finds a solution that satisfies all the physical laws perfectly.

3. The "Trainable" Speed

One of the coolest tricks in this paper is how they find the speed of the wave.

• Usually, you have to guess the speed, run the math, and see if it works.
• Here, the AI treats the speed of the wave as a "knob" it can turn. The AI turns the knob, checks the physics, and keeps turning until the physics equations balance out perfectly. It's like tuning a radio until the static disappears and the music is clear.

4. Did It Work? (The Results)

The researchers tested their "Robot Student" against the "50-page Manual" (the old math).

• The Verdict: The robot was almost perfect. Its predictions matched the manual's answers with extremely high accuracy (less than 3% error, and often much less).
• The Error Check: They looked at the "mistakes" the robot made. The mistakes were tiny, random, and didn't have any weird patterns. This means the robot truly understood the physics, not just memorized a pattern.

5. What Did They Learn About the Waves?

By using this AI, they discovered some interesting things about how these waves behave in their "layered cake":

• Stiffness Matters: If the top layer gets "stiffer" (due to changes in the material), the wave moves slower. If the bottom layer gets stiffer, the wave moves faster.
• Pressure: Squeezing the top layer makes the wave go faster. Squeezing the bottom layer makes it go slower.
• Magnetism: Changing the angle of the magnetic field or how magnetic the material is changes the wave's speed, acting like a dimmer switch for the wave's energy.
• Thickness: A thicker top layer helps the wave stay trapped and move faster.

Why Does This Matter?

This isn't just about math puzzles. This technology is a super-tool for engineers and geologists.

• Earthquakes: It helps us understand how seismic waves travel through the Earth's complex, layered interior, which helps predict earthquake damage.
• Materials: It helps design better composite materials for airplanes or bridges that can withstand stress and vibration.
• Efficiency: Instead of spending weeks running complex simulations, this AI method can solve these problems quickly and accurately, even for materials we haven't fully understood yet.

In a nutshell: The researchers taught an AI to "think like a physicist" to solve a very difficult wave problem. The AI learned the rules, tuned the wave speed, and gave answers that matched the best human math, proving that AI can be a powerful partner in understanding the physical world.

Technical Summary

1. Problem Statement

The study investigates the propagation of Shear Horizontal (SH) waves in a complex, layered composite structure. The physical system consists of:

• Upper Layer: A pre-stressed, functionally graded (FG), magnetoelastic, orthotropic layer of finite thickness ($H$).
• Substrate: A pre-stressed, functionally graded, orthotropic half-space.
• External Influences: The system is subjected to gravity and initial horizontal stress.
• Material Properties: Both layers exhibit continuous variation in mechanical properties (stiffness, density) with depth (functional grading) and possess orthotropic symmetry. The upper layer is also coupled with electromagnetic fields (magnetoelasticity).

The Challenge:
Traditional analytical and numerical methods (e.g., Finite Element Method) for solving the governing partial differential equations (PDEs) in such highly heterogeneous, coupled, and non-linear systems often require fine discretization, complex mesh generation, and significant computational effort. Furthermore, deriving closed-form analytical solutions for the dispersion relation in such complex configurations is mathematically intricate. The paper aims to address these challenges by developing a Physics-Informed Neural Network (PINN) framework to solve the dispersion problem without explicit meshing.

2. Methodology

A. Analytical Formulation (Benchmark)

Before applying the PINN, the authors derived an exact analytical solution to serve as a benchmark:

1. Governing Equations: The equations of motion for the magnetoelastic layer and the orthotropic half-space were formulated, incorporating initial stress, gravity, and Lorentz forces (from Maxwell's equations).
2. Material Grading: Material properties were assumed to vary exponentially with depth ($e^{\beta z}$).
3. Solution Derivation:
   • For the layer, the solution was expressed using trigonometric functions after transforming the variable-coefficient PDE.
   • For the half-space, the governing equation was reduced to a Whittaker differential equation due to the combined effects of gravity and functional grading. The solution involved Whittaker functions ($W_{-s,0}$).
4. Dispersion Relation: By applying boundary conditions (stress-free surface, displacement/stress continuity at the interface, and radiation condition at infinity), a transcendental dispersion equation was derived linking phase velocity ($c$) and wave number ($k$).

B. Physics-Informed Neural Network (PINN) Framework

The core contribution is the application of PINNs to solve the same problem as an eigenvalue identification task.

• Network Architecture:
  • Two separate fully connected feed-forward neural networks were used: one for the layer (outputting real and imaginary parts of displacement) and one for the half-space (outputting real displacement).
  • Activation Function: Hyperbolic tangent ($tanh$) was chosen for its smoothness and infinite differentiability, crucial for automatic differentiation (AD).
  • Input/Output: Input is the depth coordinate ($z$); outputs are displacement amplitudes.
• Trainable Parameters:
  • Standard network weights and biases ($\theta$).
  • Phase Velocity ($c$): Treated as an auxiliary trainable parameter. This transforms the problem from a standard PDE solver into an eigenvalue solver where the network learns the specific $c$ that satisfies the physics.
• Loss Function: The total loss ($L_{total}$) is a weighted sum of residuals:
  1. PDE Loss: Residuals of the governing differential equations in both media.
  2. Boundary Loss: Stress-free condition at the top surface.
  3. Interface Loss: Continuity of displacement and stress at the layer-half-space interface.
  4. Far-Field Loss: Decay condition for the half-space.
  5. Amplitude Normalization: To prevent trivial zero solutions.
• Optimization: The Adam optimizer was used to minimize the loss function. The phase velocity $c$ was updated alongside network weights for each prescribed wave number $k$.

3. Key Contributions

1. Novel Application of PINNs: This is one of the first studies to apply PINNs to magnetoelastic SH-wave propagation in functionally graded media under gravity and initial stress.
2. Eigenvalue Learning Strategy: The authors successfully formulated the dispersion relation as an optimization problem where the phase velocity is a learnable parameter, eliminating the need to solve transcendental equations numerically.
3. Mesh-Free Solution: The approach avoids the mesh generation and discretization errors inherent in FEM/FDM, offering a continuous solution representation.
4. Comprehensive Parametric Study: The study systematically analyzes the effects of:
   • Heterogeneity parameters ($\beta$).
   • Initial stress ($\Im$).
   • Gravity (Biot's parameter $G$).
   • Magnetic field angle ($\phi$) and permeability.
   • Layer thickness ($H$).
5. Rigorous Validation: Extensive error analysis using $L_1, L_2, L_\infty$, RMSE, and relative error norms confirms the accuracy of the PINN against the analytical benchmark.

4. Results and Findings

A. Validation and Accuracy

• Agreement: The dispersion curves ($c$ vs. $k$) generated by the PINN showed excellent agreement with the analytical solutions.
• Error Metrics:
  • Relative errors ranged from $5.12 \times 10^{-4}$ to $2.81 \times 10^{-2}$.
  • Absolute errors were predominantly in the order of $10^{-3}$ to $10^{-2}$.
  • The $L_\infty$ norm was bounded below $10^{-1}$, indicating no localized instabilities.
• Convergence: The loss function (PDE, interface, and total) decreased progressively, reaching final values of $\approx 10^{-5}$, confirming stable optimization.

B. Parametric Analysis Insights

• Heterogeneity ($\beta$):
  • Increasing heterogeneity in the upper layer decreased phase velocity (reduced effective stiffness).
  • Increasing heterogeneity in the half-space increased phase velocity (enhanced effective rigidity).
• Initial Stress:
  • Increased stress in the layer increased phase velocity (stiffening effect).
  • Increased stress in the half-space decreased phase velocity (softening effect).
• Gravity: Increasing the Biot gravity parameter ($G$) reduced the phase velocity.
• Layer Thickness: Thicker layers led to higher phase velocities due to better energy confinement.
• Magnetic Effects:
  • Increasing the magnetic field angle ($\phi$) increased phase velocity.
  • Increasing induced magnetic permeability decreased phase velocity.

C. Architectural Sensitivity

• Activation Functions: The model was robust across different activation functions (Sigmoid, Tanh, Swish, Arctan, Softplus), with negligible differences in relative error ($\approx 1.68 \times 10^{-2}$).
• Network Depth/Width: Increasing the number of neurons slightly improved accuracy for shallow networks. However, increasing the number of hidden layers beyond a moderate count (e.g., >5) did not significantly improve accuracy, suggesting the problem is well-suited to moderate-depth architectures.

5. Significance and Conclusion

This paper demonstrates that Physics-Informed Neural Networks are a powerful, accurate, and flexible alternative to traditional numerical methods for analyzing wave propagation in complex, heterogeneous, and coupled physical systems.

• Scientific Impact: It provides a new computational paradigm for solving non-linear eigenvalue problems in geophysics and material science, specifically for functionally graded materials (FGMs) which are critical in aerospace and structural engineering.
• Practical Utility: The method can be extended to inverse problems (e.g., identifying material properties from wave data) and real-time monitoring of structural health in layered media.
• Reliability: The study confirms that PINNs can handle the mathematical complexity of magnetoelasticity, gravity, and initial stress simultaneously, offering a mesh-free solution with high predictive fidelity.

In summary, the proposed PINN framework successfully captures the dispersion characteristics of SH-waves in a pre-stressed, magnetoelastic, functionally graded layered medium, validating its potential as a standard tool for advanced wave propagation analysis.