PaperlayerView exact revision on arXiv

Source-linked AI summary

Hyperreduction accelerates finite element model updating for the discovery of neural materials

Hagen Holthusen, Matthias Knipper, Kevin Linka

arXiv:2610.31322v1cs.CEcs.AI

TL;DR

Structural experiments provide indirect measurements for material discovery, while simulation-based identification requires repeated nonlinear finite element solves. The paper combines POD, ECSW, and reduced-model adjoints, adapting the reduction as network parameters evolve; reported results show faster training without systematic accuracy loss.

  • Problem

    Simulation-based neural-material identification repeatedly solves nonlinear finite element problems, and reduced bases and sampled elements must remain reliable as network parameters change.

    Source: §1 · p. 3

  • Method

    The approach combines POD projection, ECSW internal-energy approximation, and reduced-model adjoint sensitivities, with bootstrap and canary epochs adapting the reduction as weights evolve.

    Source: §1 · p. 3, §2.4 · p. 9, §5 · p. 33

  • Results

    Case 1 training was 13 times faster with a 2.5% force error below the full-order reference; Case 3 reduced epochs cost more than two orders less than canary epochs.

    Source: §4 · p. 29

  • Takeaways & Limitations

    Hyperreduction makes finite element model updating of neural materials computationally affordable while preserving accommodation of arbitrary geometries and boundary conditions and sparse or noisy data.

    Source: §5 · p. 33

    Identified-material accuracy is limited by the information in the data because reaction force constrains energy only along deformation states activated by the experiment.

    Source: §5 · p. 33

Abstract

from arXiv · show

Finite element model updating is among the most general approaches to discover constitutive models from structural experiments: since the displacement field is computed rather than measured, it accommodates arbitrary geometries, boundary conditions, and sparse or noisy data. However, every loss evaluation requires a nonlinear finite element solve per load step, and training neural materials requires many such evaluations. Intrusive model order reduction lowers this cost while retaining boundary conditions and variational structure. How can a reduced model remain accurate while the network weights change at every training step? We combine proper orthogonal decomposition and energy-conserving sampling and weighting with adjoint sensitivities of the reduced model, and adapt both the trial space and the sampled elements along the optimization path. The constitutive law is a polyconvex neural network: a logarithmic normalization term, together with a zero-anchored input convex architecture and a growth term, renders the energy polyconvex, coercive, and non-negative by construction, independent of the trained weights. We demonstrate the approach on two experimental datasets and one synthetic three-dimensional example. For a two-phase specimen, training time decreases from nine hours to 41 minutes compared with a full-order identification, without loss of accuracy. In the three-dimensional example, six modes replace 48 672 degrees of freedom, 116 of 15 625 elements suffice to evaluate the internal forces, and training takes 13 minutes. In all cases, the identified material reproduces the measured reaction force and predicts the displacement field from the force alone. Hyperreduction makes the discovery of neural materials from structural data affordable while retaining the advantages of finite element model updating.

1. Introduction

Neural materials must be inferred from structural measurements such as displacement fields and reaction forces, rather than stress–strain pairs. The paper proposes reducing the finite element model and adapting that reduction as network parameters evolve.

1. Introduction

  • Structural experiments provide displacement fields and reaction forces, so material laws must be inferred from indirect observables rather than measured stress–strain pairs.Neural networks can parameterize energy and dissipation potentials while satisfying thermodynamic principles by construction.Source: §1 · pp. 1–2
  • Simulation-based identification solves the boundary value problem, accommodating sparse or noisy data and, in principle, varied geometries, load paths, and boundary conditions.Other approaches can evaluate equilibrium directly on measured displacements or replace optimization with a database lookup or a learned forward pass.Source: §1 · p. 2
  • Each loss evaluation requires nonlinear finite element solves for every load step, while each gradient also requires an adjoint solve, making repeated neural-network training costly.Richer combinations of specimens and load cases increase the number of solves per loss evaluation.Source: §1 · p. 2
  • Projection-based reduction combines a POD displacement subspace with hyperreduction, which evaluates nonlinear internal forces on a small element subset.Cubature-based methods include ECSW and empirical cubature; learned reduced models have been reported to diverge outside their training range.Source: §1 · p. 2
  • Reduced models have been used in inverse problems, but a neural constitutive model changes many weights and biases at every optimization step, challenging basis and element-set reliability.Prior reduced models for explicit laws were noted to remain valid only within their training parameter range.Source: §1 · pp. 2–3
  • The study combines POD projection, ECSW energy approximation, and reduced-model adjoint sensitivities, building the basis from load-step snapshots and sampled parameter vectors.The loss compares measured and predicted reaction forces and, optionally, displacements.Source: §1 · p. 3

2. Methods

The methods formulate structural material discovery as bilevel optimization, combining a polyconvex neural constitutive law with model reduction adapted during training. The workflow uses alternating full-order and reduced solves, with POD, ECSW sampling, and loss comparisons guiding the reduced model’s construction and use.

2.1. Inverse material discovery

  • Identification is bilevel model updating: equilibrium is solved for each parameter set, while reaction-force mismatch and optional displacement mismatch train the material weights using an adjoint solve.The adjoint procedure’s cost is independent of the number of parameters.Source: §2.1 · pp. 3, 5

2.2. Reduction techniques

  • POD projects equilibrium displacements onto a snapshot basis, while ECSW assembles reduced residuals and tangents over a sparse, non-negatively weighted element subset, preserving an approximate potential.The reduced-space adjoint gives the exact gradient of the reduced model, rather than an approximation of the full-order gradient.Source: §2.2 · pp. 6–7

2.3. Neural Elasticity

  • The isotropic polyconvex material uses a zero-anchored ICNN with nonnegative weights; growth and logarithmic normalization ensure coercivity, nonnegative energy, and zero stress in the undeformed state.The architecture’s convex, non-decreasing network and the constructed energy terms make polyconvexity hold independently of the trained weights.Source: §2.3 · pp. 8–9

2.4. Workflow during hyperreduced inverse discovery

  • Because a fixed basis and sampled-element set may fail across training states, the workflow adapts the reduction to the evolving material weights.The reduced model is constructed from states near particular material parameters, so it may lose accuracy as training changes those parameters.Source: §2.2 · p. 6, §2.4 · p. 9
  • After bootstrap full-order epochs, periodic canary epochs determine whether subsequent training uses the reduced or full-order model.Canary checks compare the two solutions at identical weights; if the reduction is inaccurate, full-order solves continue until a later canary accepts it.Source: §2.4 · pp. 9–10
  • Canary solutions add snapshots when the current projection cannot represent them, prompting recomputation of the POD basis and ECSW element set when needed.Otherwise, the basis and sampled elements are retained; a greedy algorithm selects ECSW elements until its relative training error meets tolerance or the support reaches its maximum size.Source: §2.4 · pp. 9–10
  • POD selects the smallest reduced-space dimension whose relative snapshot projection error falls below a prescribed tolerance.The projection uses the first Nr columns of the POD basis; snapshot management caps the snapshot count and replaces the oldest snapshot at the limit.Source: §2.4 · pp. 9–10
  • Reduced-model acceptance requires both absolute and relative loss-deviation criteria, avoiding overly strict relative tolerances near convergence and permissive absolute tolerances at large loss.The full-order and reduced losses are compared at identical weights; if neither criterion is met, training returns to the full-order model until a later canary check.Source: §2.4 · p. 10

3. Results

The results cover heterogeneous-plate, perforated-specimen and synthetic three-dimensional identifications, including comparisons of force-only and force-plus-displacement losses. They show that hyperreduced force-based identifications can preserve force accuracy and predict displacement fields, while reduction efficiency depends on deformation localization and predictions remain limited outside states probed by the data.

3.1. Case 1: Heterogeneous material system

  • 13× less training time—0.69 h versus 8.95 h for the full-order reference—while the hyperreduced and full-order force errors remained comparable.The median epoch times were 1.23 s and 33.4 s, respectively; the reduced model was accepted at every canary epoch after bootstrap.Source: §3.1 · p. 12, Figure 2 · p. 13
  • Adaptivity kept the reduction aligned with changing parameters: the trial space grew from Nr = 9 to Nr = 13, and the active ECSW set from 186 to 356 elements.The trial-space growth followed projection-error threshold crossings; the 356 active elements represented about 14 % of the 2,500-element mesh, with sampling concentrated at boundaries and around the inclusion.Source: §3.1 · pp. 13, 14, 16, Figure 6 · p. 17
  • The force-only identification achieved a 2.5 % relative force error and R2 = 0.9970, while its unseen displacement field had a 10.4 % relative error.The inclusion remained almost undeformed (Fyy ≈1.0), while the surrounding matrix stretched up to Fyy = 1.87, indicating qualitative separation of the phases from force alone; quantitative comparison of isolated phase laws was unavailable.Source: §3.1 · pp. 14–15
  • Across reduction-parameter runs, force errors ranged from 0.22 to 0.36 N versus 0.33 N for the full-order reference; POD and ECSW tolerances chiefly changed cost, not accuracy.Speed-up per epoch ranged from 8.6 to 60; looser snapshot updates could worsen accuracy, while the authors attributed remaining error primarily to non-uniqueness of force-based identification rather than reduction.Source: §3.1 · p. 17, Figure 7 · p. 18

3.2. Case 2: Homogeneous material, perforated specimen

  • Case 2 uses the perforated TT1 membrane to compare a force-only identification with one that also includes measured displacements at 11 load steps.The specimen has six elliptical holes and strongly heterogeneous deformation; both losses use machine force, while only the second includes the DIC displacement field.Source: §3.2 · pp. 18, 20, Figure 9 · p. 19
  • For the perforated specimen, the trial space grows from 26 to 27 modes; ECSW uses 613–801 of 1,518 elements (up to 53%), and 685 reduced epochs take 79% of 2.73 h.Without a full-order reference run, the authors state only that a reduced epoch is several times cheaper at this mesh size; deformation localization limits hyperreduction efficiency.Source: §3.2 · pp. 20, 23, Figure 11 · p. 21
  • On unseen TT2, both materials overestimate force by about 10%; the two-channel model predicts displacement with R2 = 0.998 in both directions, versus 0.892 and 0.997 for force-only training.Force RMSEs are 16.4 N and 16.5 N, respectively.Source: §3.2 · p. 23
  • The two perforated-specimen identifications have similar shear moduli (0.58 and 0.61 MPa) but bulk moduli of 4.35 and 1.16 MPa, respectively; neither supports near-incompressibility.The displacement field mainly affects the volumetric response, with the authors noting that out-of-plane kinematics are coarsely represented.Source: §3.2 · p. 23
  • On TT1, force-only training gives 1.80 N force RMSE and 3.2% displacement error; adding displacements reduces that error to 2.1% but raises force error to 4.2% (6.47 N).Force-only results include R2 = 0.9992 for force and R2 = 0.933 transverse and 0.997 loading for displacement; the two-channel results have displacement R2 = 0.991 and 0.998.Source: §3.2 · pp. 20, 23, Figure 13 · p. 24

3.3. Case 3: Heterogeneous material system with spherical inclusions

  • Six modes and 116 of 15,625 elements (<1%) yield 0.21 h training; estimated full-order training would take up to 4.5 h, or up to twenty times longer.The 116 active elements cluster around inclusions and clamped-face corners; full-order epochs dominate the adaptive training cost.Source: §3.3 · pp. 26, 28
  • 0.556 N force RMSE (R2 = 0.9995) and 3.2% displacement error (R2 = 0.990 transverse, 0.9998 loading) show that force-only identification reproduces the synthetic block’s response.The relative force error is 1.0%; displacement data did not enter the loss.Source: §3.3 · p. 26
  • The block’s force constrains its nearly uniaxial response, not shear or lateral contraction: matrix J reaches 1.66 versus reference 1.18 in equibiaxial tension, and inclusion energy remains less determined.The authors attribute these limits to data information content rather than reduction.Source: §3.3 · pp. 27–28

4. Discussion

Adaptive POD and ECSW reduce training cost while preserving the identified material’s structural response, provided the reduction is updated along the optimization path. The gains depend on deformation localization and are bounded by full-order checks and the information content of the experimental data.

4. Discussion

  • A 27-fold median-epoch and 13-fold total-training speed-up in Case 1 retained a 2.5% force error below the full-order reference.Across reduction-parameter variations, per-epoch speed-up ranged from 8.6 to 60 without systematic accuracy loss.Source: §4 · p. 29
  • Case 3’s reduction remained accurate as weights evolved when adapted along the optimization path, and reduced and full-order training followed nearly the same path in Case 1.In Case 3, bootstrap and canary epochs accounted for 85% of wall-clock time, so full-order checks constrain further gains.Source: §4 · p. 29, Figure 18 · p. 29
  • Case 2 required 27 modes and up to 53% of elements because deformation localized in the ligaments, limiting the efficiency of linear-subspace reduction.Its Fyy reached 3.94, and the specimen-specific speed-up remains open without a full-order reference.Source: §4 · pp. 29, 32
  • In Case 3, the force fit had RMSE 0.556 N and R2 = 0.9995, while all three displacement components outside the loss had R2 ≥ 0.990.The relative displacement error remained below 1.5%; the matrix fit was less accurate in deformation modes the experiment hardly activated.Source: Figure 19 · p. 30, Figure 20 · p. 31
  • Case 3’s trial space grew from 4 to 6 modes, and its ECSW rule assigned positive weights to 116 of 15,625 elements.The sampled elements clustered around inclusions and clamped-face corners; the matrix was represented by a few highly weighted elements.Source: Figure 18 · p. 29, Figure 21 · p. 31, Figure 22 · p. 32
  • Full-order checks bound the speed-up: canary epochs were under 2% of Case 1 epochs but consumed 41% of wall-clock time, while bootstrap and canary epochs took 85% in Case 3.Up-to-date snapshots govern accuracy, whereas POD and ECSW tolerances govern cost.Source: Figure 18 · p. 29, §4 · p. 32
  • Reaction force constrains stiffness level but not material uniquely; displacement data reveal phase-wise stiffness distribution and improve prediction of an unseen specimen.Experiments constrain energy only along activated deformation states, leaving Case 3 responses in simple shear and equibiaxial tension less determined.Source: Figure 20 · p. 31, §4 · p. 32
  • The study restricts its demonstrations to isotropic hyperelasticity; anisotropic discovery faces added identifiability demands, and reliable adaptive reduction for path-dependent inelastic states remains open.Changing ECSW samples in inelastic problems also requires reconstructing the histories of newly sampled elements.Source: §4 · p. 33

5. Conclusion

The study combines finite element model updating with adaptive reduction and reports order-of-magnitude training-cost savings without degrading identified materials. Its demonstrated scope is limited by localized deformation, full-order checks, and the information contained in the data.

5. Conclusion

  • Hyperreduction makes neural-material finite element model updating computationally affordable while retaining its advantages for arbitrary geometries, boundary conditions, and sparse or noisy data.The method projects equilibrium onto a POD subspace, approximates internal energy with weighted sampled elements, and uses reduced-model adjoint sensitivities.Source: §5 · p. 33
  • Across two experimental and one synthetic example, training cost fell by an order of magnitude without degrading identified materials; one reaction force also predicted displacement fields excluded from the loss.The reduced model remained accurate as weights evolved, and gains grew with full-order problem size.Source: §5 · p. 33
  • Efficiency decreases for localized deformation, full-order checks bound overall speed-up, and material accuracy remains limited by the data’s information content.Reaction force constrains energy only along deformation states activated by the experiment.Source: §5 · p. 33
  • Future work may investigate adaptive full-order checks, nonlinear trial manifolds for localized deformation, and extensions to anisotropic and inelastic materials.Source: §5 · p. 33

Declaration of generative AI and AI-assisted technologies in the writing process

The authors report using Claude to draft and revise text, improve language, and assist with source-code implementation and documentation; they reviewed and take responsibility for the published content.

Declaration of generative AI and AI-assisted technologies in the writing process

Appendix A. Polyconvexity and coercivity of the energy

The energy combines a neural network with stress normalization and a growth term, using normalized invariants and a zero-anchored input-convex architecture. The construction assigns the reference configuration zero energy and stress and permits tensile or compressive volumetric stresses.

Appendix A. Polyconvexity and coercivity of the energy

  • The energy combines neural network N, stress normalization ψnrm, and growth term ψgr, with normalized invariants set to zero at F = I.The network is an input convex neural network with non-negative weights and convex, non-decreasing activations, and is zero-anchored at the origin.Source: Appendix A · p. 34
  • The network’s first two input derivatives are non-negative, while the third-input derivative minus the fourth may have either sign, allowing tensile and compressive volumetric stresses.Source: Appendix A · p. 34
  • At the reference configuration, both normalizations give ψ(I) = 0 and P(I) = 0, and all three growth-energy terms vanish.Source: Appendix A · p. 34

Appendix A.1. Polyconvexity

The energy is made polyconvex through a zero-anchored, monotone input-convex network, convex normalizations, and a growth term. The growth term also supplies coercivity that the network alone cannot guarantee, with a weight-independent bound.

Appendix A.1. Polyconvexity

  • A zero-anchored network with non-negative weights and convex, non-decreasing activations is convex in its inputs, making its composition with the invariants polyconvex.The argument uses the convexity of I1 − 1 in F and I2 − 1 in cof F, and the affine dependence of ±(J − 1) on J.Source: Appendix A.1 · pp. 34–35
  • Both linear and logarithmic normalizations preserve convexity, and the convex growth term makes the overall energy polyconvex for either normalization.The logarithmic normalization relies on non-negative coefficients; the growth term is convex for ϵ > 0.Source: Appendix A.1 · p. 35
  • The growth term supplies coercivity because the network can be nearly linear, so its curvature cannot be relied on to ensure sufficient growth.The network and linear normalization together are not bounded from below; the explicit growth term provides the missing coercivity.Source: Appendix A.1 · p. 36
  • The coercivity bound is independent of trained network weights, but its constant grows with ϵ^-3 and becomes large when the growth parameter is small.The weight-independent bound involves only the growth parameter ϵ.Source: Appendix A.1 · p. 36

Appendix B. Mean-dilatation element (Q1/P0)

The mean-dilatation element modifies the kinematics to reduce volumetric locking while retaining a displacement-only formulation. It evaluates the energy using an element-wise volume change and requires no volumetric–isochoric split.

Appendix B. Mean-dilatation element (Q1/P0)

  • For nearly incompressible materials, enforcing J ≈ 1 at every quadrature point can overconstrain trilinear displacement elements and cause volumetric locking.The mean-dilatation approach reduces volume change to one value per element.Source: Appendix B · p. 36
  • The modified deformation gradient retains the isochoric part J^-1/3 F but replaces J with the element-wise constant volume change Θ_e.The potential evaluates the energy at this modified deformation gradient.Source: Appendix B · p. 37
  • The formulation uses trilinear Q1 displacements alone and requires no split of the energy into volumetric and isochoric parts.No additional unknowns occur.Source: Appendix B · p. 37

Appendix C. Case 2: Identification from force and displacements

Adding displacement data improves agreement of the displacement field on both specimens, but trades off against force agreement on the training specimen. On the unseen specimen, force overestimation remains similar to that of the force-based material.

Appendix C. Case 2: Identification from force and displacements

  • Displacement data improve field agreement on both specimens at the expense of force fit on TT1, while TT2 force overestimation remains about the same as for the force-based material.The comparison evaluates the identified material with full-order solves on training specimen TT1 and unseen specimen TT2.Source: Appendix C · p. 37
Loading 2610.31322v1…