EUVIMEDEUVIMED entdecken
Crossref

Deep Wavelet Diffeomorphism: A New Approach to Manifold Reshaping for PINN Training

Shijiao Gao, Carlo Cattani, Shuli Mei

International Journal of Wavelets, Multiresolution and Information Processing · 03.11.2026

DOI: 10.1142/s0219691326500402 ↗

Kurz gefasst

Fachartikel in der Zeitschrift International Journal of Wavelets, Multiresolution and Information Processing, online erschienen am 03.11.2026. Verfasst von Shijiao Gao und 2 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Physics-Informed Neural Networks (PINNs) face challenges with high-gradient flows because they require dense collocation points and are susceptible to local optima, while traditional wavelet methods suffer from parameter sensitivity and are difficult to couple with physical constraints. To overcome these limitations, we propose Deep Wavelet Diffeomorphism PINN (DWD-PINN), which combines both approaches through an integrated manifold-reshaping and low-dimensional training mechanism. Leveraging wavelet multi-resolution analysis, DWD-PINN identifies physically critical regions and constructs a smooth, invertible diffeomorphic mapping. This mapping transforms the distorted physical training manifold into a smooth computational manifold and adaptively concentrates collocation points in feature-rich areas. Training a PINN in this transformed space significantly simplifies nonlinear optimization by smoothing the loss landscape and enabling more uniform gradient propagation. On two-dimensional Navier–Stokes benchmarks, DWD-PINN reduces the number of collocation points by 50% and training time by 37.2%, while robustly reducing the loss to the order of 10 −5 without sacrificing accuracy. This framework thus offers a data-efficient, physics-informed solver for complex PDE simulations.

Zitieren

APA

Gao, S., Cattani, C., & Mei, S. (2026). Deep Wavelet Diffeomorphism: A New Approach to Manifold Reshaping for PINN Training. International Journal of Wavelets, Multiresolution and Information Processing. https://doi.org/10.1142/s0219691326500402

Vancouver

Gao S, Cattani C, Mei S. Deep Wavelet Diffeomorphism: A New Approach to Manifold Reshaping for PINN Training. International Journal of Wavelets, Multiresolution and Information Processing. 2026. doi:10.1142/s0219691326500402

BibTeX
@article{Gao2026,
  author = {Gao, Shijiao and Cattani, Carlo and Mei, Shuli},
  title = {Deep Wavelet Diffeomorphism: A New Approach to Manifold Reshaping for PINN Training},
  journal = {International Journal of Wavelets, Multiresolution and Information Processing},
  year = {2026},
  doi = {10.1142/s0219691326500402}
}
Kostenlos starten →