Group B, Poster #308, Research Computing (RC)
Forward and Inverse Problems with the Elastic Wave Equation: A Comparison of Traditional Numerical and Machine-Learning Methods
Poster Image:

Poster Presentation
2025 SCEC Annual Meeting, Poster #308, SCEC Contribution #14653 VIEW PDF
nformed neural networks (PINNs) to solve both forward and inverse problems related to the elastic wave equation. By analyzing the computational benefits and limitations of these two approaches, we hope to identify the comparative strengths of PINNs versus traditional methodologies.
For the forward problem, we systematically benchmark three numerical strategies for the elastic wave equation—FDM, soft-enforced PINNs, and hard-enforced PINNs—under two optimization schemes (Adam and L-BFGS). FDM delivers the lowest errors and shortest runtimes when grid resolution is sufficient, but hard-enforced PINNs narrow the gap dramatically. With hard enforcement, L-BFGS training reaches near-FDM accuracy. Soft enforcement proves sensitive to sharp localized features, whereas hard enforcement stabilizes convergence and preserves accuracy. Across all trials, L-BFGS consistently outperforms Adam on both speed and fidelity, highlighting optimizer choice as a key design factor. Overall, the results position hard-enforced, L-BFGS-optimized PINNs as an attractive mesh-free alternative when exact constraint satisfaction is possible and grid-based methods become cumbersome or expensive. Next, we consider a classical problem of inferring elastic moduli given limited surface observations. Given the material density, we set up the PINN to solve for the material displacement as well as the shear modulus. Two networks are defined to approximate both the displacement field and the shear modulus. The composite loss function couples the PDE residual, initial/boundary conditions, surface observations, and an optional regularizer on the surface shear modulus. A warm-up phase with Adam optimizes both nets on mini-batches, followed by full-batch L-BFGS to reach machine precision.
SHOW MORE
For the forward problem, we systematically benchmark three numerical strategies for the elastic wave equation—FDM, soft-enforced PINNs, and hard-enforced PINNs—under two optimization schemes (Adam and L-BFGS). FDM delivers the lowest errors and shortest runtimes when grid resolution is sufficient, but hard-enforced PINNs narrow the gap dramatically. With hard enforcement, L-BFGS training reaches near-FDM accuracy. Soft enforcement proves sensitive to sharp localized features, whereas hard enforcement stabilizes convergence and preserves accuracy. Across all trials, L-BFGS consistently outperforms Adam on both speed and fidelity, highlighting optimizer choice as a key design factor. Overall, the results position hard-enforced, L-BFGS-optimized PINNs as an attractive mesh-free alternative when exact constraint satisfaction is possible and grid-based methods become cumbersome or expensive. Next, we consider a classical problem of inferring elastic moduli given limited surface observations. Given the material density, we set up the PINN to solve for the material displacement as well as the shear modulus. Two networks are defined to approximate both the displacement field and the shear modulus. The composite loss function couples the PDE residual, initial/boundary conditions, surface observations, and an optional regularizer on the surface shear modulus. A warm-up phase with Adam optimizes both nets on mini-batches, followed by full-batch L-BFGS to reach machine precision.
SHOW MORE