PINN framework for elastodynamic wave propagation in bimaterials

Sonal Ankush Chibire, Jenn-Terng Gau and Bo Zhang submitted a paper to arXiv on 7 Jul 2026 presenting a physics-informed neural network framework for transient elastodynamic wave propagation in bimaterial systems (arXiv:2607.06479). The network embeds the axisymmetric equations of linear elasticity and enforces initial, boundary and interface conditions through a physics-informed loss while using ANSYS Workbench Explicit Dynamics finite-element simulations for validation and as supplementary training constraints.

What did the authors build and how does it work?

The authors built a PINN-based surrogate that directly incorporates the axisymmetric equations of linear elasticity and the corresponding initial, boundary and interface conditions into the training loss, enabling the network to model transient elastodynamic waves in a bimaterial specimen. The study focuses on a steel-aluminum specimen representative of a Split Hopkinson Pressure Bar configuration and trains the network with physics constraints plus high-fidelity finite-element data.

The framework embeds the governing elastodynamic equations inside the loss function so the network enforces physics as it learns. The PINN is trained to reproduce axial and radial displacement histories and the evolution of stress and strain while explicitly modelling wave transmission and reflection at the bimaterial interface. The authors also report mesh-sensitivity studies and tests on additional material combinations to demonstrate numerical robustness and generality.

How was the PINN validated and what did it reproduce?

High-fidelity finite-element simulations performed with ANSYS Workbench Explicit Dynamics were used both for validation and as supplementary data constraints during training; the trained network reproduces displacement and stress/strain histories with close agreement to the finite-element solutions. The network accurately predicts wave transmission and reflection across the bimaterial interface, reproduces axial and radial displacement histories, face-averaged responses, and captures the dominant stress and strain evolution reported by the ANSYS models.

The paper states the trained network can "predict wave responses at previously unseen time instants" and can do so for modified material properties without requiring additional finite-element simulations, providing a continuous surrogate model for elastodynamic analysis. The arXiv submission metadata shows the paper was submitted on 7 Jul 2026 (arXiv:2607.06479) and the initial upload was 3,079 KB in size ([v1] Tue, 7 Jul 2026 16:34:55 UTC).

Why it matters

Embedding the governing equations into the loss produces a surrogate that can interpolate in time and across material-property variations without rerunning costly explicit finite-element solves, which the authors highlight as a computationally efficient approach for high-rate solid mechanics and impact engineering applications. If the PINN maintains fidelity to explicit dynamics across materials and meshes, practitioners can use it to explore parameter spaces or run many-query tasks faster than repeated ANSYS explicit runs.

This approach also bridges data-driven models and conventional simulation: the PINN uses physics to constrain learning and uses ANSYS outputs as anchors, reducing reliance on purely data-hungry black-box surrogates while retaining the continuous, differentiable structure of neural networks.

What to watch

Look for follow-up arXiv versions or accompanying code and data releases: the paper's arXiv entry includes toggles for "Code, Data and Media Associated with this Article". Future signals of adoption will include extensions to experimental validation, broader material sets, and public benchmarks comparing PINN surrogates against explicit finite-element runtimes and accuracy.