Axion: A differentiable, GPU-native finite element framework,demonstrated on phase-field fracture

We introduce Axion (Autodifferentiable, XLA-compiled, Implicit, On-device Numerics), a differentiable, GPU-native finite element framework written in JAX. The element kernels, Newton and Krylov iterations, bound-constrained updates, staggered coupling of fields, and the loop over load increments are expressed as a single program that is just-in-time compiled and executed on the GPU without host intervention. Consistent tangent operators are obtained by forward-mode automatic differentiation of the element residual, and derivatives of the discrete solution with respect to model parameters are obtained by differentiating the same program. Global operators are applied element by element without assembling a sparse matrix, octree mesh refinement is performed within fixed-size arrays, and a cell-based domain decomposition runs the same program on several GPUs. Phase-field fracture is implemented as the first module because it combines fine mesh resolution, bound-constrained damage, and repeated nonlinear solves over a loading history, which together make it computationally expensive. On two- and three-dimensional single-edge-notched tension benchmarks, a uniformly refined Axion computation completes faster than the corresponding adaptive FEniCSx computation. At identical discretization, the two-dimensional benchmark runs 37.9 times faster on one GPU than a FEniCSx implementation on one CPU. Strong and weak scaling efficiencies on four GPUs are 92% and 96%, respectively. Large-scale demonstrations include hydraulic fracture in a poroelastic medium (0.38M degrees of freedom), fracture of a multi-material premolar under contact loading (0.69M degrees of freedom), and a concrete gravity dam (96M degrees of freedom) on a single GPU. The implementation is available as open-source software.

Authors

Publication Details

Journal
arXiv (Cornell University)
Published
2026-11-18
DOI
https://doi.org/10.48550/arxiv.2611.23682
Primary Topic
Numerical methods in engineering
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Axion: A differentiable, GPU-native finite element framework,demonstrated on phase-field fracture

Somdatta Goswami, Wei Wang, Meenu Krishnan U.
arXiv (Cornell University)
Numerical methods in engineering
preprint

Axion: A differentiable, GPU-native finite element framework,demonstrated on phase-field fracture

Somdatta Goswami, Wei Wang, Meenu Krishnan U.
preprint en

Abstract

We introduce Axion (Autodifferentiable, XLA-compiled, Implicit, On-device Numerics), a differentiable, GPU-native finite element framework written in JAX. The element kernels, Newton and Krylov iterations, bound-constrained updates, staggered coupling of fields, and the loop over load increments are expressed as a single program that is just-in-time compiled and executed on the GPU without host intervention. Consistent tangent operators are obtained by forward-mode automatic differentiation of the element residual, and derivatives of the discrete solution with respect to model parameters are obtained by differentiating the same program. Global operators are applied element by element without assembling a sparse matrix, octree mesh refinement is performed within fixed-size arrays, and a cell-based domain decomposition runs the same program on several GPUs. Phase-field fracture is implemented as the first module because it combines fine mesh resolution, bound-constrained damage, and repeated nonlinear solves over a loading history, which together make it computationally expensive. On two- and three-dimensional single-edge-notched tension benchmarks, a uniformly refined Axion computation completes faster than the corresponding adaptive FEniCSx computation. At identical discretization, the two-dimensional benchmark runs 37.9 times faster on one GPU than a FEniCSx implementation on one CPU. Strong and weak scaling efficiencies on four GPUs are 92% and 96%, respectively. Large-scale demonstrations include hydraulic fracture in a poroelastic medium (0.38M degrees of freedom), fracture of a multi-material premolar under contact loading (0.69M degrees of freedom), and a concrete gravity dam (96M degrees of freedom) on a single GPU. The implementation is available as open-source software.

arXiv (Cornell University)
Numerical methods in engineering
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon