arXiv · 2611.22099
Differentiable High-Order Mesh $r$-Adaptivity
Abstract
High-order mesh r-adaptivity typically entails minimizing a nonlinear functional using its gradients and, for Newton-based methods, its Hessian. These derivatives can be expensive to derive, must be updated with any changes to the functional, and errors in the derivatives directly impact the convergence of r-adaptivity. Additionally, derivatives of the r-adaptivity operator are required for consistent sensitivity analysis for inverse problems. This work develops a framework for differentiable r-adaptivity based on the Target-Matrix Optimization Paradigm. Given a pointwise functional defining the local mesh-adaptation objective, the proposed framework uses $\partial$FEM to automatically generate pointwise first- and second-order derivatives. These local derivatives are then assembled into a global gradient vector and a Hessian operator through the finite element operator decomposition. The derivative kernels support portable CPU and GPU execution, with partial assembly and tensor sum factorization enabling matrix-free application of the Hessian operator. The framework also provides consistent sensitivities of the r-adaptivity operator through automatically generated mixed derivatives and implicit differentiation of the mesh stationarity equations. Numerical experiments show GPU solve times comparable to those of hand-derived TMOP implementations on a benchmark problem, improvement in r-adaptivity convergence due to accurate computation of the derivatives, and demonstrate the use of mesh sensitivities for material-interface identification in an inverse heat-conduction problem.
Explore related subjects
Keep this discovery
Explore connections, maps & timelines
Ketan Mittal, Julian Andrej, John Camier, Veselin Dobrev, Tzanio Kolev, Leonardo Molinari, Vladimir Tomov. 2026-11-17. Differentiable High-Order Mesh $r$-Adaptivity. https://arxiv.science/abs/2611.22099
Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite arXiv Science by Indiciae: BibTeX · RIS.