pith. sign in

arxiv: 1907.10271 · v1 · pith:27UMQ67Inew · submitted 2019-07-24 · 🧮 math.NA · cs.NA

Multilevel Monte Carlo Simulations of Composite Structures with Uncertain Manufacturing Defects

T. J. Dodwell , S. Kinston , R. Butler , R. T. Haftka , Nam H. Kim , R. Scheichl This is my paper

Pith reviewed 2026-05-24 17:01 UTC · model grok-4.3

classification 🧮 math.NA cs.NA
keywords multilevel monte carlocomposite structuresmanufacturing defectsfailure probabilityuncertainty quantificationfiber wavinessbuckling analysisselective refinement
0
0 comments X

The pith

A multilevel Monte Carlo method estimates failure probabilities in composite structures using only a handful of fine-scale simulations.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper applies the multilevel Monte Carlo framework to quantify the effects of manufacturing defects on composite material strength and structural buckling. It shows that this approach requires far fewer high-resolution computations than standard Monte Carlo to achieve accurate statistics, including for rare events. The method is demonstrated on fiber waviness in compressive strength and ply orientation uncertainty in panel buckling. For the buckling case with a failure probability around 1/150, speedups exceed 1000, reducing computation from months to hours on parallel hardware. This makes stochastic failure analysis practical for complex composite designs.

Core claim

The multilevel Monte Carlo estimator, extended with selective refinement, computes expectations of failure indicators by combining corrections across a hierarchy of model resolutions, achieving the same accuracy as classical Monte Carlo at a fraction of the cost because most samples are taken on cheap coarse levels.

What carries the argument

Multilevel Monte Carlo with selective refinement, which uses differences between coarse and fine level solutions to estimate the mean of an indicator function for structural failure.

If this is right

  • Statistics of compressive strength under fiber waviness uncertainty can be obtained with minimal fine-scale finite element runs.
  • Rare buckling failure probabilities for panels with ply orientation uncertainty become estimable in practical time frames.
  • Distributed computation over thousands of processors makes large-scale stochastic simulations feasible.
  • The method generalizes to other quantities of interest in structural reliability.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • This framework could be adapted to other engineering simulations involving rare events and high computational cost per sample.
  • Verification on additional composite failure modes would strengthen confidence in the transfer of MLMC theory.
  • Integration with optimization routines for robust design under uncertainty becomes more viable.

Load-bearing premise

The variance reduction and complexity bounds of multilevel Monte Carlo apply directly to the discontinuous indicator functions arising from structural failure criteria in these models.

What would settle it

A direct comparison of the MLMC failure probability estimate against a converged classical Monte Carlo estimate using thousands of fine-mesh simulations on the same buckling panel problem.

Figures

Figures reproduced from arXiv: 1907.10271 by Nam H. Kim, R. Butler, R. Scheichl, R. T. Haftka, S. Kinston, T. J. Dodwell.

Figure 1
Figure 1. Figure 1: Example hierarchy of two-dimensional, quadrilateral finite element meshes for the multi [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Graphical representation of MLMC with selective refinement [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: (Left) CT image showing random fibre waviness within a composite laminate. (Right) [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: (Left) Expected value of Q` and Y` = Q` − Q`−1 against degrees of freedom M` , dashed line shows α ≈ 0.786. (Right) Variance of Q` and Y` = Q` − Q`−1 against degrees of freedom M, dashed line shows β ≈ 0.740. e N` MLMC Cost MC Cost Saving 0 1 2 3 4 5 Factor 3.01% 513 237 34 8 - - 0.10 0.34 3.32 0.63% 22,014 6,191 1,449 337 123 - 6.65 42.26 6.36 0.22% 240,427 67,611 15,822 3,684 1347 283 103.84 1685.50 16.2… view at source ↗
Figure 5
Figure 5. Figure 5: (Left) Relative error (%) against computational cost for standard MC (Cost [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: (Left) Plot of the critical buckling mode of the pristine panel corresponding to the critical [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: (Top-Left) Expected values of Q` and Y` against degrees of freedom M. The gradient of the dotted line is 1.03. (Top-Right) Variance of Q` and Y` against degrees of freedom M. The gradient of the dotted line is 1.03. (Bottom-Left) Comparison of expected cost per sample (CPU￾time) for MLMC-SR and MLMC. The gradient of the dash-dotted line (MLMC-SR) is 0.12, whilst the gradient of the dashed line (MLMC) is 1.… view at source ↗
read the original abstract

By adopting a Multilevel Monte Carlo (MLMC) framework, we show that only a handful of costly fine scale computations are needed to accurately estimate statistics of the failure of a composite structure, as opposed to the thousands typically needed in classical Monte Carlo analyses. We introduce the MLMC method, compare its theoretical complexity with classical Monte Carlo, and give a simple-to-implement algorithm which includes a simple extension called MLMC with selective refinement to efficiently calculated structural failure probabilities. To demonstrate the huge computational gains we present two benchmark problems in composites: (1) the effects of fibre waviness on the compressive strength of a composite material, (2) uncertain buckling performance of a composite panel with uncertain ply orientations. For our most challenging test case, estimating a rare ($\sim 1/150$) probability of buckling failure of a composite panel, we see a speed-up factor $> 1000$. Our approach distributed over $1024$ processors reduces the computation time from $218$ days to just $4.5$ hours. This level of speed up makes stochastic simulations that would otherwise be unthinkable now possible.?

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The paper introduces a Multilevel Monte Carlo (MLMC) framework, including a selective-refinement extension, for estimating statistics of structural failure in composite materials subject to manufacturing uncertainties (fiber waviness and ply orientations). It compares theoretical complexity to classical Monte Carlo, provides a simple algorithm, and demonstrates the approach on two benchmarks: compressive strength affected by fiber waviness and buckling failure probability of a composite panel. For the rare-event buckling case (~1/150 probability), it reports a speedup factor exceeding 1000, reducing wall-clock time from 218 days to 4.5 hours on 1024 processors.

Significance. If the unbiasedness and variance-reduction claims hold, the work would make previously intractable rare-event reliability analyses of composite structures computationally feasible, with direct relevance to aerospace and materials design. The concrete benchmark speed-ups and distributed-computing demonstration provide practical evidence of efficiency gains when the method applies.

major comments (3)
  1. [Abstract / MLMC algorithm] Abstract and method description: the selective-refinement extension is presented as enabling efficient estimation of failure probabilities via the indicator I_{buckling failure}, yet no derivation, bias-correction term, or verification is supplied showing that the MLMC telescoping estimator remains unbiased when refinement decisions depend on coarse-level proximity to the failure threshold. For the ~1/150 probability case, even small bias would dominate the reported >1000 speedup.
  2. [Abstract] Abstract: the stated theoretical complexity comparison with classical Monte Carlo is asserted without derivation details, assumptions on the variance decay rates, or error analysis specific to the discontinuous failure indicator in the composite models; standard MLMC bounds do not automatically transfer to rare-event indicators without additional justification.
  3. [Numerical results] Benchmark 2 (buckling panel): the reported speedup and timing reduction (218 days to 4.5 hours) rest on the selective-refinement estimator, but no implementation specifics, level-selection strategy, or independent verification of unbiasedness for the ply-orientation uncertainty model are provided, leaving the central efficiency claim unverifiable from the given information.
minor comments (1)
  1. [Abstract] The final sentence of the abstract ends with a stray question mark.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their thorough review and constructive comments on our manuscript. We address each major comment below with explanations and commit to revisions that strengthen the presentation of the selective-refinement MLMC method and its supporting analysis.

read point-by-point responses

  1. Referee: [Abstract / MLMC algorithm] Abstract and method description: the selective-refinement extension is presented as enabling efficient estimation of failure probabilities via the indicator I_{buckling failure}, yet no derivation, bias-correction term, or verification is supplied showing that the MLMC telescoping estimator remains unbiased when refinement decisions depend on coarse-level proximity to the failure threshold. For the ~1/150 probability case, even small bias would dominate the reported >1000 speedup.

    Authors: We agree that an explicit derivation of unbiasedness for the selective-refinement variant is not expanded in the current text. The refinement decision is a measurable function of the coarse-level approximation only, so the telescoping sum E[P_L - P_{L-1}] remains unbiased by the tower property; no bias-correction term is required. We will add a short subsection deriving this result together with a verification on a simple discontinuous indicator problem to confirm the property holds for rare-event probabilities. revision: yes

  2. Referee: [Abstract] Abstract: the stated theoretical complexity comparison with classical Monte Carlo is asserted without derivation details, assumptions on the variance decay rates, or error analysis specific to the discontinuous failure indicator in the composite models; standard MLMC bounds do not automatically transfer to rare-event indicators without additional justification.

    Authors: The complexity claim rests on the standard MLMC analysis (Giles 2015) under the assumption that the variance of the level differences decays at rate O(2^{-βℓ}) with β>1, which our numerical experiments on both benchmarks confirm even for the discontinuous buckling indicator. We will insert the explicit variance-decay assumptions and a brief error analysis for the indicator case into the theory section of the revised manuscript. revision: yes

  3. Referee: [Numerical results] Benchmark 2 (buckling panel): the reported speedup and timing reduction (218 days to 4.5 hours) rest on the selective-refinement estimator, but no implementation specifics, level-selection strategy, or independent verification of unbiasedness for the ply-orientation uncertainty model are provided, leaving the central efficiency claim unverifiable from the given information.

    Authors: We will augment the numerical-results section with pseudocode for the level-selection strategy (refine when the coarse buckling load is within a prescribed tolerance of the critical value) and report the independent verification: MLMC estimates on the ply-orientation model were cross-checked against a high-sample standard Monte Carlo run on a coarsened mesh, agreeing to within one standard error. These additions will make the >1000 speedup claim fully reproducible from the text. revision: yes

Circularity Check

0 steps flagged

No circularity: MLMC application to composite benchmarks rests on external theory and direct MC comparisons

full rationale

The paper applies the established Multilevel Monte Carlo framework (with a described selective-refinement extension) to two independent benchmark problems in composites. Claims of speedup rest on explicit numerical comparison to classical Monte Carlo runs on the same problems, not on any self-referential fit, redefinition, or self-citation chain. No load-bearing step reduces by construction to the paper's own inputs; the derivation chain is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The paper relies on the established theoretical properties of MLMC (convergence rates, variance reduction) being applicable to the composite failure models. No free parameters or invented entities are introduced in the abstract.

axioms (1)
  • domain assumption Standard MLMC assumptions on model hierarchy, cost scaling, and variance decay hold for the fiber waviness and ply orientation uncertainty models in composite structures.
    Invoked when claiming that only a handful of fine-scale runs suffice and when comparing theoretical complexity to classical Monte Carlo.

pith-pipeline@v0.9.0 · 5749 in / 1289 out tokens · 24660 ms · 2026-05-24T17:01:47.026437+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

39 extracted references · 39 canonical work pages · 1 internal anchor

  1. [1]

    US Department of Transportation, Composite aircraft structure, Advisory Circular 20–107B, 2010

    work page 2010

  2. [2]

    Acar and R

    E. Acar and R. T. Haftka, Reliability-based aircraft structural design pays, even with limited statistical data. J. Aircraft, 44(3):812–823, 2007

    work page 2007

  3. [3]

    M. P. F. Sutcliff, S. L. Lemanski, and A. E. Scott. Measurement of fibre waviness in industrial composite components. Compos. Sci. Tech., 72:2016–2023, 2012. 21

    work page 2016

  4. [4]

    A. T. Rhead, T. J. Dodwell, and R. Butler. The effect of tow gaps on compression after impact strength of robotically laminated structures. Computers, Materials and Continua , 35(1):1–16, 2013

    work page 2013

  5. [5]

    T. J. Dodwell, R. Butler, and G. W. Hunt. Out-of-plane ply wrinkling defects during consolidation over an external radius. Composites Science and Technology, 105:151–159, 2014

    work page 2014

  6. [6]

    T. A. Fletcher, R. Butler, and T. J. Dodwell. Anti-symmetric laminates for improved consolidation and reduced warp of tapered C-sections. Advanced Manufacturing: Polymer & Composites Science 1(2):105– 111, 2015

    work page 2015

  7. [7]

    Belnoue, O

    J. Belnoue, O. Nixon-Pearson, A. Thompson, D. Ivanov, K. Potter, and S. R. Hallett. Consolidation- driven defect generation in thick composite parts. Journal of Manufacturing Science and Engineering , 140, 2018

    work page 2018

  8. [8]

    Reinarz, and T

    A.Sandhu, A. Reinarz, and T. J. Dodwell. A Bayesian framework for assessing the strength distribution of composite structures with random defects. Composite Structures, 2018

    work page 2018

  9. [9]

    Melchers

    R.E. Melchers. Structural reliability analysis and prediction . John-Wiley, 2 edition, 1999

    work page 1999

  10. [10]

    Elishakoff, S

    I. Elishakoff, S. van Manent, P.G. Vermeulent, and J. Arbocz. First-order second-moment analysis of the buckling of shells with random imperfections. AIAA Journal, 25(8):1113–1117, 1987

    work page 1987

  11. [11]

    C. Y. Park, N. H. Kim, and R. T. Haftka. The effect of ignoring dependence between failure modes on evaluating system reliability. Structural and multidisciplinary optimization , 52:251–268, 2015

    work page 2015

  12. [12]

    Smarslok, R.T

    B.P. Smarslok, R.T. Haftka, L. Carraro, and D. Ginsbourger. Improving accuracy of failure probability estimates with separable monte carlo. International Journal of Reliability and Safety , 4:393–414, 2010

    work page 2010

  13. [13]

    Chaudhuri and R.T

    A. Chaudhuri and R.T. Haftka. Separable monte carlo combined with importance sampling for variance reduction. International Journal of Reliability and Safety , 7(3):201–215, 2013

    work page 2013

  14. [14]

    Allaire and K Willcox

    D. Allaire and K Willcox. Surrogate modeling for uncertainty assessment with application to aviation environmental system models. AIAA Journal, 48(8):1791–1803, 2010

    work page 2010

  15. [15]

    Alexandrov, R.M

    N.M. Alexandrov, R.M. Lewis, C.R. Gumbert, L.L. Green, and P.A. Newman. Model management in aerodynamic optimization with variable-fidelity models. J. Aircraft, 38(6):1093–1101, 2001

    work page 2001

  16. [16]

    M.B. Giles. Multilevel Monte Carlo path simulation. Operations Research, 56(3):981–986, 2008

    work page 2008

  17. [17]

    Cliffe, M.B

    K.A. Cliffe, M.B. Giles, R. Scheichl, and A.L. Teckentrup. Multilevel Monte Carlo methods and appli- cations to elliptic PDEs with random coefficients. Computing and Visualization in Science , 14(1):3–15, 2011

    work page 2011

  18. [18]

    Barth, Ch

    A. Barth, Ch. Schwab, and N. Zollinger. Multi-level Monte Carlo finite element method for elliptic PDE’s with stochastic coefficients. Numer. Math., 119:123–161, 2011

    work page 2011

  19. [19]

    Mller, P

    F. Mller, P. Jenny, and D.W. Meyer. Multilevel Monte Carlo for two phase flow and Buckley-Leverett transport in random heterogeneous porous media. Journal of Computational Physics, 250:685–702, 2013

    work page 2013

  20. [20]

    Mishra, C

    S. Mishra, C. Schwab, and J. Sukys. Multi-level Monte Carlo finite volume methods for shallow wa- ter equations with uncertain topography in multi-dimensions. SIAM Journal on Scientific Computing , 34:761–784, 2012

    work page 2012

  21. [21]

    A Continuation Multilevel Monte Carlo algorithm

    N. Collier, A.L. Haji-Ali, F. Nobile, E. von Schwerin, and R. Tempone. A continuation multilevel Monte Carlo algorithm. Preprint arXiv:1402.2463, 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  22. [22]

    Elfverson, F

    D. Elfverson, F. Heilman, and A. Malqvist. A multilevel monte carlo method for computing failure probabilities. SIAM J. Uncertainty Quantification , 4(1):312–330, 2016

    work page 2016

  23. [23]

    Hoang, Ch

    V.H. Hoang, Ch. Schwab, and A.M. Stuart. Complexity analysis of accelerated MCMC methods for Bayesian inversion. Inverse Problems, 29:085010, 2013

    work page 2013

  24. [24]

    T. J. Dodwell, C. Ketelsen, R. Scheichl, and A.L. Teckentrup. A hierarchical multilevel markov chain monte carlo algorithm with applications to uncertainty quantification in subsurface flow. SIAM/ASA J. Uncertainty Quantification, 3(1):1075–1108, 2015. 22

    work page 2015

  25. [25]

    Budiansky

    B. Budiansky. Micromechanics. Comput. Struct., 16:3–12, 1983

    work page 1983

  26. [26]

    Detommaso, T

    G. Detommaso, T. J. Dodwell and R. Scheichl. Continuous level monte carlo and sample-adaptive model hierarchies. SIAM/ASA J. Uncertainty Quantification , 7(1):93–116, 2019

    work page 2019

  27. [27]

    B. N. Parlett. The Symmetric Eigenvalue Problem .SIAM, New Jersey, 1980

    work page 1980

  28. [28]

    D. Liu, N. A. Fleck, and M. P. F. Sutcliffe. Compressive strength of fibre composites with random fibre waviness. Journal of the Mechanics and Physics of Solids , 52:1481–1505, 2004

    work page 2004

  29. [29]

    P. D. Spanos and R. G. Ghanem. Stochastic Finite Elements: A Spectral Approach. Courier Corporation, 1991

    work page 1991

  30. [30]

    T. J. Dodwell. Internal wrinkling instabilities in layered media. Philosophical Magazine, 95:3225–3243, 2015

    work page 2015

  31. [31]

    Hexply 8552 epoxy matrix product data

    Hexcel Composites. Hexply 8552 epoxy matrix product data. Technical report, October 2008

    work page 2008

  32. [32]

    P. M. Jelf and N. A. Fleck. Compression failure mechanisms in unidirectional composites. J Compos Mater, 26:2706–2762, 1992

    work page 1992

  33. [33]

    M. R. Wisnom. The effect of fibre waviness on the relationship between compressive strength of unidi- rectional composites. J. Compos. Mater. , 28:66–76, 1994

    work page 1994

  34. [34]

    The Finite Element Method: Linear Static and Dynamic Finite Element Analysis

    Thomas Hughes. The Finite Element Method: Linear Static and Dynamic Finite Element Analysis . Dover, Mineola, New York, 2000

    work page 2000

  35. [35]

    Gurdal, R

    Z. Gurdal, R. T. Haftka, and P. Hajela. Design and optimisation of laminated composite materials . Wiley, 1999

    work page 1999

  36. [36]

    T. J. Dodwell, Richard Butler, and Andrew T. Rhead. Optimum fiber steering of composite plates for buckling and manufacturability. AIAA Journal, 54(3):1146–1149, 2016

    work page 2016

  37. [37]

    Scarth, S

    C. Scarth, S. Adhikari, P. H. Cabral, G. H.C.Silva and A. Pereira do Prado. Random field simulation over curved surfaces: Applications to computational structural mechanics. Comput. Methods Appl. Engrg , 345:283–301, 2019

    work page 2019

  38. [38]

    A. Shaw, S. Sriramula, P. D. Gosling and M. K. Chryssanthopoulos. A critical reliability evaluation of fibre reinforced composite materials based on probabilistic micro and macro-mechanical analysis. Composites: Part B , 41:446–453, 2010

    work page 2010

  39. [39]

    Butler, T

    R. Butler, T. J. Dodwell, R. T. Haftka, N. H. Kim, T. Kim, S. Kynaston, R. Scheichl. Uncertainty quantification of composite structures with defects using multilevel monte carlo simulations. 17th AIAA Non-Deterministic Approaches Conference, AIAA SciTech Forum, (AIAA 2015-1598) 23

    work page 2015