NAP-Wave: A Neural Adaptive Plane-Wave Architecture for High-Frequency Helmholtz Problems
Organizations: Basque Center for Applied Mathematics (BCAM), Bilbao, Spain · University of the Basque Country (UPV/EHU), Leioa, Spain · Ikerbasque: Basque Foundation for Science, Bilbao, Spain
Abstract
Training neural networks to simulate highly oscillatory wavefields presents optimization challenges, often hindered by spectral bias and poorly conditioned physics-based loss landscapes. To address these challenges and enhance convergence, we propose NAP-Wave, a joint neural architecture and training strategy that expands the target scattered wavefield in an adaptive neural plane-wave basis. The basis functions are extracted from all hidden layers and modulated by trainable plane-wave carriers, embedding the high-frequency oscillatory physics directly into the network's trial space. NAP-Wave is trained on a physics-informed loss built from the Lippmann-Schwinger integral equation, so no differentiation of the network with respect to its input is required. We develop a decoupled optimization strategy in which Adam updates the parameters of the neural plane-wave basis, while the linear combination coefficients are computed exactly by a least-squares (LS) solver. We couple this with an adaptive-depth training strategy in which the network starts shallow and grows progressively layer-by-layer. At each growth step, we avoid the loss spikes typical of dynamic network growth using the LS-computed coefficients, the warm-started parameters, and the preservation of the first- and second-order optimizer moments of the previous stage. Ablation studies confirm that the plane-wave basis, the exact coefficient solve, and the progressive schedule independently accelerate convergence and improve accuracy. Validated on exact-solution benchmarks, NAP-Wave achieves relative L2 errors of order 1%. On highly heterogeneous geophysical benchmarks, it achieves relative errors of 5%, 9%, and 10%, significantly outperforming published PINN methods in both accuracy and convergence time. To ensure reproducibility, the code is available at https://github.com/Hamza-Kamil/NAP-WAVE.
Figures & tables
| Parameter | Fluid-cylinder scattering | 3D radial velocity gradient | Marmousi | Overthrust | Otway |
| Domain (km) | |||||
| Background velocity (km/s) | |||||
| Source position (km) | |||||
| Frequency (Hz) | |||||
| Background wavenumber (km -1 ) |
| Setting | Fluid-cylinder scattering | 3D radial velocity gradient | Marmousi | Overthrust | Otway |
| Depth / width | |||||
| Basis functions | |||||
| Quadrature grid | |||||
| Total iterations | |||||
| Test grid |
| Reference | Method | Marmousi Relative | Overthrust Relative | Otway Relative |
| Abedi et al. [ 2 ] | Gabor PINN | 21.3% | 21.1% | not reported |
| Abedi et al. [ 1 ] | Hybrid GI+PDE | 9.1% | 13.2% | 14.0% |
| NAP-Wave (ours) | Neural Adaptive Plane-Wave basis functions |
| Adaptive | Fixed | |||||
| Depth | GI Loss | Rel. (%) | Time (min) | GI Loss | Rel. (%) | Time (min) |
| 2 | 76.02 2.82 | 0.31 0.00 | 79.61 3.28 | 0.34 0.01 | ||
| 3 | 27.01 1.29 | 0.62 0.00 | 43.26 2.45 | 0.73 0.01 | ||
| 4 | 16.29 0.83 | 1.05 0.00 | 21.20 0.93 | 1.28 0.01 | ||
| 5 | 11.44 0.64 | 1.59 0.00 | 14.42 0.88 | 2.01 0.01 | ||
| 6 | 9.61 0.67 | 2.24 0.01 | 11.81 0.21 | 2.90 0.01 | ||
| Adaptive | Fixed | ||||
| Basis / -update | Width | Step to | Time (min) | Step to | Time (min) |
| Plane-wave, -LS | 80 | 17,314 | 3.91 | 32,211 | 8.27 |
| Plane-wave, -LS | 128 | 7,853 | 2.30 | 8,228 | 3.32 |
| Plane-wave, -trained | 80 | 165,262 | 8.64 | 346,188 | 18.19 |
| Plane-wave, -trained | 128 | 105,917 | 7.67 | 119,746 | 8.81 |
| Standard, -LS | 80 | 20,033 | 4.47 | 39,637 | 9.89 |
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
| Problem setting | Network and trial space | Discretization and training | |||
| Spatial dimension, | , | Target depth, current depth stage | , | Quadrature points, uniform grid | |
| , , | Physical, padded domain, taper region | Architecture vector; width of layer | , | Grid spacings | |
| , | Angular, temporal frequency, | Network parameters; their number | , | Taper widths, horizontal and vertical | |
| , | Velocity field, background velocity | , , | Weights, biases, activation of layer | , , | Loss, reduced, empirical reduced loss |
| , | Local, background wavenumber, | , | Total basis, active basis at depth | , | Tikhonov parameter, decay schedule |
| , , , | Squared slowness, background, contrast, tapered | -th complex neural basis function | , | Iterations at stage , total budget | |
| Operation | Cost | Memory |
| Forward pass | ||
| Green operator (FFT) | ||
| Assemble | ||
| Normal matrix | ||
| Right-hand side | ||
| Cholesky factorization |
Neural PDE Solvers
Neural PDE solvers use trainable models to approximate solutions or solution operators for partial differential equations, often with physical constraints.
See all 297 papers