Speaker
Description
Comparative Analysis of Neural Surrogate Architectures for Acoustic Wave Propagation in Heterogeneous Media
M․V․Minasyan, Z․H․ Mkrtchyan, A․M․Minasyan
Institute of Applied Problems of Physics of the National Academy of Sciences of the Republic of Armenia, 25 Hrachya Nersisyan, Yerevan, Armenia, 0014
The study presents a comparative analysis of three scientific machine learning architectures, Physics-Informed Neural Networks (PINNs), Fourier Neural Operators (FNO), and Deep Operator Networks (DeepONet), applied as surrogate models for acoustic wave propagation in heterogeneous media. Traditional numerical solvers such as FDTD and FEM are computationally prohibitive in multi-query scenarios, and neural surrogates have demonstrated orders-of-magnitude inference speedup. However, the three architectures encode wave physics through fundamentally different mechanisms, resulting in distinct performance characteristics that have not been systematically compared.
The analysis evaluates reported results across problem categories of increasing complexity (homogeneous and layered media, heterogeneous velocity fields, and scattering-dominated domains) and compares the architectures on prediction accuracy, computational cost, data efficiency, noise robustness, and frequency-dependent performance. The spectral bias problem, the systematic underrepresentation of high-frequency wave content, is examined as the principal shared limitation, with each architecture offering partial mitigation through its structural properties: PDE-constrained optimization in PINNs, spectral convolution in FNO, and branch-trunk decomposition in DeepONet.
The analysis demonstrates that the three architectures occupy complementary niches: PINNs for inverse problems and data-scarce settings, FNO for high-throughput forward simulation on regular domains, and DeepONet for geometrically flexible and noise-tolerant applications. Current gaps, specifically the absence of standardized acoustic benchmarks and the neglect of uncertainty quantification, are identified as priorities for advancing the field.