Spectral bias mitigation in physics-informed neural surrogates for nonlinear structural dynamics: Fourier feature encoding versus Kolmogorov–Arnold representations

Standard multilayer perceptrons (MLPs) employed in physics-informed neural networks (PINNs) suffer from spectral bias—the systematic inability to learn high-frequency components in multi-frequency structural response signals. Recently, Kolmogorov–Arnold Networks (KANs) with learnable edge activations have been proposed as alternatives that could inherently mitigate this limitation. This study presents a rigorous, parameter-matched (~16,400 parameters per model) comparison of four neural architectures—Fourier-enhanced KAN (FPIKAN), Fourier-MLP, plain MLP, and standalone KAN—across four canonical nonlinear dynamics benchmarks: the Bouc–Wen hysteretic oscillator, the Duffing oscillator with cubic stiffness, the forced Van der Pol oscillator, and a three-degree-of-freedom shear building. A systematic 2 × 2 factorial ablation study isolates the individual contributions of (i) Fourier feature encoding and (ii) backbone architecture choice. Results demonstrate unambiguously that Fourier feature encoding is the dominant factor, reducing relative L2 errors by 98.6–100% when applied to MLP across all benchmarks. Fourier-MLP achieves the lowest error in all four benchmarks (L2 = 0.014, with L2 < 10-3 for Duffing and Van der Pol). Without Fourier encoding, KAN outperforms MLP by up to 3.5× (Van der Pol: L2 = 0.164 vs. 0.574) due to its learnable Chebyshev activations acting as implicit spectral decomposers. Paradoxically, adding Fourier encoding to KAN degrades performance in two benchmarks (Duffing: -201.7%; Van der Pol: -9.0%), and FPIKAN consistently underperforms Fourier-MLP. Analysis of gradient norms through Chebyshev polynomial layers reveals oscillatory gradient magnitudes incompatible with the smooth optimization landscape required by second-order physics residual terms. These findings provide critical, experimentally validated guidance for selecting neural architectures in computational structural dynamics.

Authors

Institutions

Publication Details

Journal
The Journal of Supercomputing
Published
2026-11-19
DOI
https://doi.org/10.1007/s11227-026-08964-x
Primary Topic
Model Reduction and Neural Networks
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Spectral bias mitigation in physics-informed neural surrogates for nonlinear structural dynamics: Fourier feature encoding versus Kolmogorov–Arnold representations

Salih Berkan Aydemı̇r
4 links
The Journal of Supercomputing
Model Reduction and Neural Networks
article

Spectral bias mitigation in physics-informed neural surrogates for nonlinear structural dynamics: Fourier feature encoding versus Kolmogorov–Arnold representations

Salih Berkan Aydemı̇r
article en

Abstract

Standard multilayer perceptrons (MLPs) employed in physics-informed neural networks (PINNs) suffer from spectral bias—the systematic inability to learn high-frequency components in multi-frequency structural response signals. Recently, Kolmogorov–Arnold Networks (KANs) with learnable edge activations have been proposed as alternatives that could inherently mitigate this limitation. This study presents a rigorous, parameter-matched (~16,400 parameters per model) comparison of four neural architectures—Fourier-enhanced KAN (FPIKAN), Fourier-MLP, plain MLP, and standalone KAN—across four canonical nonlinear dynamics benchmarks: the Bouc–Wen hysteretic oscillator, the Duffing oscillator with cubic stiffness, the forced Van der Pol oscillator, and a three-degree-of-freedom shear building. A systematic 2 × 2 factorial ablation study isolates the individual contributions of (i) Fourier feature encoding and (ii) backbone architecture choice. Results demonstrate unambiguously that Fourier feature encoding is the dominant factor, reducing relative L2 errors by 98.6–100% when applied to MLP across all benchmarks. Fourier-MLP achieves the lowest error in all four benchmarks (L2 = 0.014, with L2 < 10-3 for Duffing and Van der Pol). Without Fourier encoding, KAN outperforms MLP by up to 3.5× (Van der Pol: L2 = 0.164 vs. 0.574) due to its learnable Chebyshev activations acting as implicit spectral decomposers. Paradoxically, adding Fourier encoding to KAN degrades performance in two benchmarks (Duffing: -201.7%; Van der Pol: -9.0%), and FPIKAN consistently underperforms Fourier-MLP. Analysis of gradient norms through Chebyshev polynomial layers reveals oscillatory gradient magnitudes incompatible with the smooth optimization landscape required by second-order physics residual terms. These findings provide critical, experimentally validated guidance for selecting neural architectures in computational structural dynamics.

The Journal of SupercomputingVol. 82(17)
Amasya Üniversitesi (TR)
Model Reduction and Neural Networks
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon