🔬 physics

A Physics-Informed Neural Network Frameworkfor Elastodynamic Wave Propagation inBimaterial Systems

This study proposes a physics-informed neural network framework that accurately models transient elastodynamic wave propagation in bimaterial systems by embedding governing equations and boundary conditions into the learning process, serving as a robust and computationally efficient surrogate model validated against finite-element simulations for applications in high-rate solid mechanics.

Original authors: Sonal Ankush Chibire, Jenn-Terng Gau, Bo Zhang

Published 2026-07-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Sonal Ankush Chibire, Jenn-Terng Gau, Bo Zhang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a giant, high-speed camera that can freeze time to watch how a shockwave travels through a sandwich made of steel and aluminum. Now, imagine trying to predict exactly how that wave bounces, twists, and squishes the metal without actually building the sandwich or running the expensive, slow-motion simulations every single time. That is the challenge this paper tackles.

The researchers built a "smart guesser" called a Physics-Informed Neural Network (PINN). Think of this not as a magic black box that just memorizes answers, but as a student who is forced to study the rulebook of physics before taking a test. Instead of just looking at data, this student is taught the laws of how waves move through solid objects (specifically the Navier–Lamé equations) and is told, "You must follow these rules, or you fail."

To see if their smart student was any good, the team set up a virtual experiment mimicking a Split Hopkinson Pressure Bar (SHPB). This is a classic setup where a steel bar smashes into a specimen to send a shockwave through it. In this case, the specimen was a tiny cylinder, 2.5 mm long, made of two different metals: steel and aluminum. They ran a super-detailed, high-fidelity computer simulation using ANSYS Workbench Explicit Dynamics to create the "ground truth"—the perfect, albeit slow and expensive, answer key.

The main finding is that their PINN framework, which learned from both the physics rules and a little bit of data from that expensive simulation, could predict how the wave moved through the steel and aluminum with impressive accuracy. It successfully tracked the axial (lengthwise) and radial (sideways) movements of the metal. For instance, when the wave hit the steel input face, the PINN predicted the displacement history almost perfectly, matching the simulation's timing for when the wave arrived and how high the peak was. It even got the "Poisson effect" right—where the metal gets fatter when you squish it lengthwise.

However, the paper is careful to note where the "smart student" started to stumble. While the predictions were great for the initial wave and the aluminum side, things got a bit wobbly later on. When the wave bounced around inside the steel specimen multiple times (after about 20 microseconds), the PINN's prediction of the stress and strain became a bit smoother than the jagged, complex reality shown in the simulation. The authors suggest that these derivative-based quantities (like stress and strain) are harder to predict perfectly than the raw movement (displacement), especially when the wave gets complicated with reflections and material mismatches.

Crucially, the paper rules out the idea that this method is just a simple interpolation tool that only works for the exact moments it was trained on. They tested the model on a time window it had never seen before (200 to 400 microseconds) without any extra training. The model successfully continued the story, predicting the wave's behavior in that unseen future with strong agreement, proving it learned the underlying physics rather than just memorizing a list of numbers.

The confidence here is high for the displacement predictions and the general wave behavior, but the authors explicitly state that the stress and strain predictions, while good for the main event, show larger discrepancies in the later, messy stages of the simulation. They didn't claim to have solved all wave problems forever; instead, they demonstrated a framework that acts as a fast, continuous "surrogate model." Once trained, this model can spit out answers for different times or slightly different materials without needing to run the heavy, slow computer simulations again.

In short, the paper shows that by teaching a neural network the laws of physics, you can create a fast, reliable tool for watching how shockwaves travel through mixed-metal systems. It's not a perfect crystal ball for every single tiny detail of a complex crash, but it's a powerful, physics-respecting shortcut that gets the big picture right and saves a ton of computing time.

Technical Summary: A Physics-Informed Neural Network Framework for Elastodynamic Wave Propagation in Bimaterial Systems

Problem Statement
Transient elastodynamic wave propagation in heterogeneous solids, particularly within bimaterial systems, presents significant challenges for computational mechanics. While conventional Finite Element Methods (FEM) provide high-fidelity solutions for stress-wave transmission and reflection, they are computationally expensive when repeated simulations are required for parametric studies, optimization, or inverse analysis. This is especially critical in high-rate loading scenarios, such as those modeled by the Split Hopkinson Pressure Bar (SHPB), where material discontinuities and imperfect interfaces strongly influence wave behavior. There is a need for efficient surrogate modeling techniques that maintain physical consistency while reducing computational costs.

Methodology
The authors propose a Physics-Informed Neural Network (PINN) framework designed to solve the axisymmetric equations of linear elasticity governing transient wave propagation in a steel-aluminum bimaterial specimen.

Governing Equations: The model is based on the Navier-Lamé equations for a linear, isotropic elastic solid. The formulation utilizes radial ( $u_r$ ) and axial ( $u_x$ ) displacement components in a cylindrical coordinate system. Strain and stress tensors are derived from these displacements using Hooke's law and constitutive relations for isotropic materials.
PINN Architecture: A fully connected feed-forward neural network maps spatial-temporal coordinates $(r, x, t)$ to displacement fields $(u_r, u_x)$ . Automatic differentiation is employed to compute the necessary spatial and temporal derivatives to evaluate the governing partial differential equations (PDEs) and their residuals without numerical differentiation.
Loss Function: The network is trained by minimizing a composite loss function comprising:
- Physics-based terms: Residuals of the governing PDEs, initial conditions (undeformed state at rest), boundary conditions (axisymmetry, traction-free surfaces, prescribed impact velocity), and interface conditions (continuity of axial displacement and normal stress; vanishing shear traction).
- Data-driven terms: Soft constraints derived from high-fidelity FEM simulations performed using ANSYS Workbench Explicit Dynamics. These data points include displacement histories at specific nodes and face-averaged values.
Training Strategy: The Adam optimizer is used to minimize the weighted total loss. A progressive weighting strategy is adopted, initially prioritizing displacement observations and boundary conditions before gradually enforcing the governing physics to improve convergence.

Key Contributions

Framework Development: The study establishes a PINN framework specifically tailored for axisymmetric elastodynamic wave propagation in bimaterial systems, explicitly handling material interfaces and frictionless contact conditions.
Hybrid Validation: The approach integrates high-fidelity ANSYS Explicit Dynamics simulations not only as a validation benchmark but also as supplementary data constraints during training to enhance predictive accuracy.
Surrogate Modeling: The trained network functions as a continuous surrogate model capable of predicting displacement, stress, and strain fields at arbitrary spatial locations and time instants without requiring additional FEM simulations.

Results
The proposed framework was validated against ANSYS Explicit Dynamics results for a steel-aluminum SHPB configuration:

Displacement Accuracy: The PINN demonstrated excellent agreement with FEM results for both axial and radial displacement histories at representative monitoring nodes on the steel input and aluminum output faces. It accurately captured wave arrival times, peak displacements, and elastic unloading.
Generalization: The model successfully generalized to unseen time intervals (200–400 µs) without retraining, maintaining smooth and physically consistent predictions, although minor deviations were noted in late-time radial responses due to complex wave reflections.
Face-Averaged Responses: Post-processed face-averaged displacements (critical for SHPB analysis) showed improved agreement with FEM solutions compared to nodal predictions, as cross-sectional averaging reduced local numerical fluctuations.
Stress and Strain: The network accurately reproduced the dominant compressive stress peaks and initial strain responses. However, discrepancies were observed in the post-peak response, particularly for the steel specimen, where the PINN predicted a smoother response compared to the oscillatory behavior in the FEM solution. This suggests that while displacement predictions are robust, derivative-based quantities (stress/strain) are more sensitive to late-time wave interactions.

Significance and Claims
The paper claims that integrating PINNs with explicit finite-element analysis provides an accurate and computationally efficient framework for elastodynamic wave propagation in heterogeneous solids. The primary significance lies in the ability to generate a continuous surrogate model that significantly reduces the computational cost associated with repeated FEM simulations in parametric studies and engineering analyses. The authors position this work as an effective approach for high-rate solid mechanics and impact engineering applications, noting that while the method shows promise, displacement fields are generally more robustly predicted than stress and strain fields in complex transient scenarios. Future work is identified as extending the framework to other bimaterial combinations and refining physics enforcement for late-time wave reflections.

More like this