Physics-Informed Neural Networks with Penalty Strategy for Domain Parameterization in Isogeometric Analysis

The Journal of Geometric Analysis · Springer Nature · 2026-09-28 · DOI: 10.1007/s11424-026-5379-1

Аннотация

In Isogeometric Analysis (IGA), constructing high-quality analysis-suitable computational domain parameterizations is a crucial issue. In recent years, learning-based parameterization methods have attracted attention with the development of machine learning and deep learning. This paper proposes a novel learning-based domain parameterization method for IGA. The proposed method employs Physics-Informed Neural Networks (PINNs) to solve Partial Differential Equations (PDEs) associated with parameterizations. To achieve high-quality parameterizations, the loss function incorporates both angular distortion and area distortion during the construction process. Furthermore, the authors introduce a penalty function-based Jacobian regularization strategy that enhances network training stability and ensures the generation of analysis-suitable parameterizations. After training the network, the authors use tensor-product B-splines to fit the results and construct the parameterizations. Numerical examples demonstrate that the proposed method achieves superior orthogonality and uniformity compared to existing deep learning-based parameterization approaches for complex planar computational domains. Additionally, the proposed method can be extended to volume parameterizations for 3D computational domains.

Читать у издателя →

Похожие статьи