PaperlayerView exact revision on arXiv

Source-linked AI summary

Assessing Nonlinear Elimination Preconditioning for Trust-Region Phase-Field Fracture

Tianchen Hu

arXiv:2609.05458v1cs.CEmath.APmath.NA

TL;DR

Phase-field fracture solvers struggle with crack-front localization: monolithic Newton can stall, while alternate minimization converges slowly. This paper introduces an on-demand nonlinear-elimination trust-region Newton method and finds that energy globalization supplies robustness, while elimination reduces iterations without demonstrating a total-work advantage over well-globalized monolithic Newton.

  • Problem

    Nonconvex coupled fracture energies can make monolithic Newton stall near advancing crack fronts, while alternate minimization may require hundreds of sweeps per load step.

  • Method

    The method combines an energy Steihaug–Toint trust region, primal-dual active-set irreversibility, and an on-demand field-split sweep eliminating an algebraically selected hard set spanning displacement and damage.

  • Results

    Energy globalization lets monolithic Newton complete every loading history without cutbacks; nonlinear elimination reduces outer iterations by 19–25% for brittle and 17% for ductile fracture, but remains only competitive with monolithic Newton in assembly-work proxy.

  • Takeaways & Limitations

    Nonlinear elimination is an iteration-reduction mechanism whose on-demand gate bounds overhead, with no demonstrated total-work advantage over well-globalized monolithic Newton.

  • Takeaways & Limitations

    The sweep is an approximate nonlinear block relaxation, and its pre-sweep active set can freeze advancing degrees of freedom or raise the KKT residual during rapid front propagation.

Abstract

from arXiv · show

Each quasi-static load step of phase-field fracture is a bound-constrained minimization of a nonconvex, coupled displacement-damage energy under an irreversibility bound on the damage. Monolithic Newton stalls once the nonlinearity localizes at the advancing crack front, and staggered (alternate-minimization) schemes converge slowly there. We present an on-demand nonlinear-elimination preconditioned trust-region Newton method: an energy Steihaug-Toint trust region, a primal-dual active set for irreversibility, and a bound-constrained field-split sweep that eliminates an algebraically-identified "hard set" spanning both fields before each step. The elimination is applied on demand -- triggered by the coupled Newton's own stalling and otherwise skipped -- so the method reduces to monolithic Newton at no surcharge where the step is already healthy. We find the robustness to come from the energy trust region: with that globalization fixed, monolithic Newton already completes every loading history without cutbacks, where residual-merit Newton death-spirals, alternate minimization stalls, and the full-field sweep loses robustness. Against that well-globalized baseline, the on-demand elimination cuts outer nonlinear iterations by 19-25% (brittle) and 17% (ductile), with always-on elimination reaching 26-28% and about $39\%$ at the ductile nucleation step. Measured machine-independently, as a full-mesh-equivalent assembly-work proxy rather than wall-clock, it is competitive with -- not faster than -- monolithic Newton (within about 10%), whereas an always-on sweep adds up to 30%. Nonlinear elimination is thus an iteration-reduction mechanism whose overhead the on-demand gate bounds, with no demonstrated total-work advantage over well-globalized monolithic Newton.

1 Introduction

Phase-field fracture combines nonconvex coupled displacement–damage minimization with damage irreversibility, making crack-front propagation difficult for standard solvers. The paper combines energy trust-region globalization, bound-constrained active sets, and on-demand coupled nonlinear elimination to target this difficulty.

  • Motivation: Nonconvex softening makes the coupled Hessian indefinite and causes monolithic Newton residual plateaus near advancing crack fronts.Alternate minimization is robust but can require hundreds of sweeps per step near propagation.
  • Globalization: Residual-merit line searches can stall in the softening regime because their descent direction need not decrease the energy.The method instead globalizes the energy with a Steihaug–Toint trust region that handles negative curvature explicitly.
  • Evaluation design: The paper isolates nonlinear elimination’s contribution by holding the energy trust-region globalization fixed and comparing against well-globalized monolithic Newton.This design addresses the confounding of globalization and preconditioning in earlier comparisons.
  • Nonlinear elimination: The nonlinear-elimination preconditioner identifies an algebraic hard set spanning displacement and damage, avoiding explicit crack tracking.The hard set targets the strongly nonlinear crack-front region before the global Newton step.
  • Bound constraints: Irreversibility is enforced through one primal–dual active set applied to the coupled operator and elimination sub-solves.A merit-safeguarded sweep prevents active-set oscillation during rapid crack propagation.
  • On-demand activation: On-demand activation runs the elimination only when monolithic coupling stalls, so healthy steps pay no elimination surcharge.The implementation also restricts finite-element assembly to elements incident to the hard set.

2 Problem formulation

Each quasi-static load step minimizes a total potential over coupled displacement and damage variables subject to boundary conditions and damage irreversibility. The resulting bound-constrained variational problem uses the energy gradient as residual and a primal–dual active set for its inequality constraint.

  • Problem statement: Each load step minimizes total potential energy over the coupled displacement–damage vector x = (u, d).The damage field satisfies d ∈ [0, 1].
  • Bounds: Damage irreversibility imposes the lower bound d ≥ d_n, where d_n is the converged phase field from the previous load step.The upper bound is the physical bound d ≤ 1.
  • Variational structure: For the considered phase-field models, the coupled residual satisfies R = ∇Ψ, making Ψ an exact energy merit.The coupled operator can nevertheless become indefinite near the softening limit point.
  • Constraint treatment: The damage bound makes each load step a bound-constrained variational inequality rather than a smooth root-finding problem.A primal–dual active set handles the bound constraint.
  • Energy: The total potential combines damage-degraded stored bulk energy with regularized fracture surface energy.The ductile model additionally includes stored plastic energy.

3 Method

The method combines an energy-based Steihaug–Toint trust region, primal–dual active sets, and a hard-set nonlinear-elimination sweep within an inexact Newton framework. The sweep preserves the original energy residual and is safeguarded, restricted, and activated only when coupled Newton stalls.

  • Outer trust-region Newton: The inexact-Newton method solves a coupled energy trust-region subproblem with Steihaug–Toint truncated CG and an SPD field-split preconditioner.The trust region handles indefinite coupled Hessians through radius truncation and negative-curvature termination.
  • Bound constraints: Irreversibility is enforced by a primal–dual active set, with active degrees of freedom eliminated consistently from the coupled operator and field-split solves.The damage block holds active increments fixed, while convergence is tested using the reduced residual and active-set stationarity.
  • Acceptance and globalization: The decoupled acceptance test accepts a step when either energy merit or reduced complementarity residual improves, shrinking the radius only when both fail.This avoids energy-ratio stagnation near constrained minimizers, where the energy can be flat while the reduced residual remains nonzero.
  • Nonlinear elimination: The nonlinear preconditioner is a multiplicative displacement-then-damage field-split sweep, restricted to the hard set and used only to generate an improved starting iterate.The outer trust-region step remains based on the original residual, preserving the variational energy globalization.
  • On-demand activation: The sweep is activated after stalled reduced-residual contraction or trust-region rejection, with hysteresis turning it off once contraction resumes.The gate avoids charging healthy elastic and post-crack iterations while targeting process-zone iterations.
  • Sweep safeguarding: The merit-safeguarded sweep is necessary because freezing a rapidly advancing active set can increase the KKT residual and destabilize unconditional application.The sweep therefore remains outside the trust-region ratio test only after its own merit safeguard accepts it.

4 Results

Across brittle and ductile benchmarks, energy-merit globalization preserves robustness while nonlinear elimination reduces outer iterations, especially near crack nucleation. On-demand activation limits overhead, leaving total assembly work essentially comparable to monolithic Newton, whereas always-on or full-field sweeps can be costly or fragile.

  • Robustness: The energy-merit baseline completes all loading histories with zero cutbacks and identical physics, while residual-merit and alternate-minimization baselines fail or become impractical.Residual-merit monolithic Newton death-spirals, and the residual-merit field-split sweep requires 2465 iterations by step 68; alternate minimization also fails to complete.
  • Brittle fracture: 28% on AT2 and 26% on AT1: NEPIN reduces monolithic outer iterations from 1100 to 795 and from 1011 to 753, respectively.The reduction is concentrated in the nucleation and propagation spike, while elastic-regime counts are nearly identical.
  • Ductile fracture: 39% at ductile nucleation: NEPIN reduces the single-step monolithic count from 199 to 121, but the overall reduction is 10%.Elastic and post-nucleation propagation steps require about 8 iterations for both methods, concentrating the benefit at nucleation.
  • On-demand activation: On-demand activation reduces block-solve work by 13–50% relative to blind NEPIN while recovering 89% of AT2 and 74% of AT1’s iteration reduction.Activation fractions rise from 42.5% for AT2 to 75.2% for ductile cases, concentrating the surcharge where monolithic Newton stalls.
  • Total work: Adaptive NEPIN stays within [−8%, +10%] of monolithic Newton in full-mesh-equivalent assembly work, while blind NEPIN costs up to +30%.The outer full-mesh residual and Jacobian assemblies dominate cost and are unchanged by hard-set restriction, bounding the total-work benefit.

5 Discussion

The energy merit supplies robustness, while nonlinear elimination reduces iterations when applied selectively to localized coupled nonlinearity. Its benefits depend on hard-set locality and do not yet translate into lower total assembly work.

  • Energy merit: Every solver family completes with zero cutbacks under the energy merit, unlike residual merit, which fails on the tested problems.The energy trust region is identified as the source of robustness across brittle and ductile fracture.
  • Robustness and on-demand cost: The on-demand gate avoids sweep overhead on easy steps by activating only when monolithic Newton stalls.Monolithic Newton already converges in about four iterations on most brittle steps, so always-on elimination largely wastes its extra sub-solves there.
  • Hard-set localization and selection: AT2 produces a broad damage band of about 38% of the mesh, whereas AT1 produces a compact band of about 5%.Raising the AT2 damage threshold from 10^-2 to 0.25 shrinks the median band from 38% to 3% with outer iterations changing from 795 to 802.
  • Hard-set localization and selection: Tightening the damage threshold improves locality rather than speed because full-mesh coupled assembly dominates total work.The displacement front can still span the mesh at nucleation and crack-through, limiting residual-based selectivity.
  • Dissipative (ductile) extension: The framework extends to dissipative J2 plasticity when the incremental problem retains a variational potential whose stationarity gives the coupled residual.Local constitutive return mapping supplies the inner minimization, preserving the solver’s variational block structure.

6 Conclusions and future work

The method is robust under energy globalization and reduces nonlinear iterations, but its present local elimination does not yet reduce total work. Future directions target tighter front resolution, global softening modes, and variational strength models.

  • 6 Conclusions and future work: The energy trust region makes monolithic Newton robust, while on-demand elimination reduces iterations by 19–25% on brittle tests and 17% overall on the ductile test.Always-on elimination reaches 26–28% on brittle cases and about 39% at ductile nucleation.
  • 6 Conclusions and future work: The present method has no demonstrated total-work advantage because elimination overhead offsets its iteration savings against well-globalized monolithic Newton.This conclusion follows the reported assembly-work comparison and the method’s robustness baseline.
  • 6 Conclusions and future work: The current local sweep addresses local imbalance but not the global softening limit point, and it advances the crack front only a sliver per iteration.A converged elimination on a tightly confined hard set is proposed to make iteration savings large enough for a genuine cost reduction.
  • 6 Conclusions and future work: Future work proposes global-mode corrections for crack-opening near-null directions or continuation to attack the limit point.These mechanisms complement, rather than replace, local hard-set elimination.
  • 6 Conclusions and future work: A single-potential phase-field model with a genuine strength surface would extend the solver to strength-driven nucleation.The proposed extension relies on retaining the variational structure required by the solver.

Declaration of generative-AI use

The author used Anthropic Claude models to assist with software implementation and comparison-harness setup, then reviewed and verified all AI-assisted output.

  • Declaration of generative-AI use: Claude models assisted with software implementation and setup of the comparison harness.The models listed are Opus 4.8, Sonnet 5, and Opus 5.
  • Declaration of generative-AI use: The author states that all AI-assisted output was reviewed and verified and takes responsibility for the manuscript’s content and results.

A Governing equations and constitutive models

The appendix formulates brittle and ductile phase-field fracture through a variational potential with displacement, damage, and constitutive internal variables. Local plastic minimization preserves the global residual and symmetric coupled tangent needed by the solver.

  • Governing equations and constitutive models: The solver depends on the variational structure of the energy rather than on the specific boundary-value or constitutive choices.
  • Governing equations: Stationarity of Ψ over displacement and damage is constrained by the irreversibility bound and yields coupled momentum and gradient-damage equations.The bound is enforced through a primal–dual active set as a variational inequality.
  • Elasticity, degradation, and fracture (brittle): The brittle models use small-strain isotropic elasticity with spectral tension/compression splitting and AT1 or AT2 crack geometric functions.AT2 uses α(d)=d^2, while AT1 uses α(d)=d.
  • Elasticity, degradation, and fracture (brittle): Brittle and ductile cases use degradation functions that distinguish quadratic brittle degradation from a strength-controlled rational form for the ductile testbed.The ductile strength parameter fixes peak stress independently of the phase-field length scale.
  • Plasticity (ductile): The ductile model uses small-strain J2 plasticity with associative flow, power-law hardening, and plastic internal variables at quadrature points.The phase field degrades the yield surface and plastic energy contributes to the crack-driving energy.
  • Plasticity (ductile): At each load step, local minimization over plastic variables defines a reduced incremental potential while displacement and damage remain fixed.The radial return mapping performs this local minimization and returns the constitutive internal variables.
  • Plasticity (ductile): The envelope theorem and consistent constitutive tangent preserve R=∇Ψ, symmetry, and the coupled 2×2 block structure for the ductile problem.Consequently, the same solver applies without modification.

B Implementation

The implementation uses a custom coupled MOOSE–PETSc executor with separate displacement and damage systems coupled through a PETSc MatNest. Field-split and elimination blocks use reduced-space variational-inequality Newton sub-solves with algebraic multigrid preconditioning.

  • The custom coupled executor runs in MOOSE, drives PETSc, and places phase-field kernels and materials in an application layer.
  • Displacement and damage are represented as separate MOOSE nonlinear systems, while PETSc MatNest stores their diagonal matrices and off-diagonal coupling blocks.
  • Each field-split or elimination block solves a reduced-space variational-inequality Newton problem using PETSc SNESVINEWTONRSLS.
  • Algebraic multigrid through hypre BoomerAMG preconditions the sub-solves, with Eisenstat–Walker controlling inexact linear tolerances.
  • The reduced-space solver updates only inactive free degrees of freedom, while active degrees of freedom remain fixed.
Loading 2609.05458v1…