aNeuSF: A Factorized Neural Surrogate for Efficient Shape-Function Construction in Numerical Coarsening
Numerical coarsening simulates structures by embedding them into a coarse, regular background mesh while maintaining high simulation accuracy and avoiding tedious preprocessing and challenging volume mesh construction. Its practical applicability, however, is often limited by the high cost of constructing material-aware shape functions for individual coarse elements. We present aNeuSF, a factorized neural surrogate for efficient shape-function construction in numerical coarsening. Instead of learning the complete coarsening map directly, aNeuSF decomposes coarse-element shape functions into an analytical Bézier-order interpolator and a learned boundary-to-fine mapping predicted from coarse-element density fields. This analytical-learned factorization reduces the dimensionality of the learning target, improves trainability, and provides a natural mechanism for preserving key shape-function properties, including partition of unity and rigid-motion consistency. It further supports selectable Bézier orders, allowing different accuracy-efficiency trade-offs within a unified formulation. Experiments on linear-elastic static problems show that aNeuSF achieves competitive simulation accuracy while delivering substantial batched acceleration over conventional numerical coarsening baselines, with up to two orders of magnitude speedup in high-resolution and high-order settings.
Authors
- Xuwei Wang
- Haroon Ijaz (ORCID: https://orcid.org/0009-0004-1839-3778)
- Ming Li
- Wei Chen
Institutions
- Zhejiang University (CN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-20
- DOI
- https://doi.org/10.21203/rs.3.rs-10068181/v1
- Primary Topic
- Model Reduction and Neural Networks
- Type
- preprint