betaFeedback

WorksPhysics and Astronomy

Machine learning of strength surfaces enables high regularization length phase-field fracture

Abstract

Classical phase-field models of brittle fracture capture toughness-controlled crack growth but not the material's strength surface, which governs crack nucleation. The theory of Kumar et al. (2020) incorporates the strength surface through a nucleation driving force, and Chockalingam (2025) derived this driving force for arbitrary strength surfaces described by nn material parameters, with which the phase-field strength surface (PFSS) converges to the material strength surface as the regularization length ε→0\varepsilon \to 0 and matches it exactly at nn chosen states for any ε\varepsilon. In practice, however, the strength surface must be supplied as an analytic form chosen and calibrated case by case, and for the widely used analytic surfaces (n=2n=2) accurate strength reproduction requires ε\varepsilon commensurate with the Irwin characteristic length lchl_{ch}, imposing prohibitive mesh requirements for large structures. We introduce a neural-network (NN) strength function that is learned directly from strength data, is physics-constrained by construction, and provides far more calibration parameters than analytic surfaces; the calibration is generalized to any number of strength states and to enforce openness in chosen directions. A single architecture captures synthetic data from the Mohr-Coulomb, Drucker-Prager, Mogi-Coulomb and 3D Hoek-Brown surfaces and experimental data for graphite, natural rubber and styrene-butadiene rubber. The resulting PFSS is accurate up to ε=10 lch\varepsilon = 10\,l_{ch}, whereas that of the exact analytic surfaces performs comparatively poorly even at ε=0.5 lch\varepsilon = 0.5\,l_{ch}. Finite-element simulations validate the theory across fracture regimes. By decoupling strength fidelity from the regularization length, the approach substantially reduces the mesh resolution required for phase-field fracture simulations of large structures.

Paper details

vlnc:W_46ekg34u
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
arXiv nonexclusive license
Usage

views · 6

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…