Thermal-shock cracks no longer need seeded defects: a mesh-free deep energy method lets them nucleate, branch and arrest on their own
Synopsis
Phase-field simulation of thermal-shock fracture has required defects placed in advance or meshes fitted to the crack, so where cracks nucleate, how far apart they sit and which ones keep growing could not come out of the computation. A mesh-free deep energy method represents displacement and phase field in one multiresolution network and the transient temperature in a second, estimates both energies on integration points redrawn every iteration, reproduces the straight-to-oscillatory-to-branched path sequence as the temperature jump grows on a glass strip, and forms crack arrays of the experimentally observed type on a quenched ceramic slab.
Interpretation
Phase-field simulation of thermally driven brittle fracture can now compute where a crack array nucleates, how far apart its cracks sit, and which of them propagate and which arrest, without prescribing crack paths or nucleation sites and without a mesh fitted to the cracks. Earlier neural energy-based treatments of thermal fracture all started from defects placed in advance, so nucleation, spacing and selective propagation under a transient temperature field were not computed. On two notched plates the initiation peak loads agree with an independent finite element reference in the cooled, isothermal and heated cases, and the peak displacement agrees to within one load step.
The method puts displacement and phase field in one network and the transient temperature in a second, each with its own multiresolution feature-grid encoding, with the temperature obtained from the incremental functional of the backward Euler heat equation. Both the mechanical energy and the thermal functional are estimated on integration points redrawn at every iteration, so no point set is held fixed and no mesh is refined around the crack band or the thermal boundary layer. On the quenched ceramic slab the temperature network agrees with a backward Euler one-dimensional solution on the same time steps at the midpoint of the long face at every load step, and the departure stays small at both thermal loads.
Irreversibility is imposed on the crack set only, and carrying the optimizer state across load steps with learning rates decaying within each step keeps a uniformly cooled surface from turning into a damaged film before cracks nucleate. With the penalty acting on the whole domain, every small stochastic optimizer update is retained and accumulates into a damaged film. On the quenched ceramic slab only the crack-set threshold, the carried optimizer state and the within-step learning-rate decay used together kept the bands at a nearly constant width and the cooled face clean.
Perspective
The demonstrated scope is two-dimensional quasi-static brittle fracture: two notched plates, a directionally quenched glass strip, a quenched ceramic slab and a quenched ceramic annulus, with the temperature either computed or given analytically, and each quench example using a single realization of the random material field.
Results depend on the resolution of the neural representations and on the optimization procedure, and a systematic convergence study has not been made; a larger iteration budget changes the crack spacing on the slab, and on the strip the paths near transitions between propagation regimes change with every increase of the budget; the settings that control spurious damage accumulation were tested under limited conditions and do not remove every artifact; a damaged layer persists along the bore of the annulus at the largest thermal load; under the convective condition the method forms several times more near-surface cracks than the finite element computation, a difference that arises at nucleation; the AT2 strength is tied to the regularization length.
