Reproduce “Precise Gradient Discontinuities in Neural Fields for Subspace Physics” on OpenResearch

About this paper

Mengfei Liu, Yue Chang, Zhecheng Wang, Peter Yichen Chen, Eitan Grinspun

Researchers from the University of Toronto and MIT CSAIL developed a neural field construction method that precisely represents gradient discontinuities in reduced-order physical simulations. This approach enables the accurate and generalizable simulation of complex phenomena like heterogeneous materials and evolving creases, outperforming traditional neural fields in capturing sharp transitions while maintaining near real-time performance for interactive design and optimization.

Key insights

• Gradient discontinuities can be precisely modeled by lifting the input to a higher dimension using a specific non-smooth (in gradient) but continuous lifting function, such as a smoothly clamped distance function.
• The approach aligns with theoretical aspects of elliptic PDEs, where eigenfunctions of weighted Laplace operators naturally possess gradient discontinuities at material interfaces.
• Avoiding embedding interface geometry into network weights allows the same neural representation to generalize across varying shapes and material distributions without requiring retraining.

Summary by alphaXiv.