pith. sign in

arxiv: 2409.18272 · v1 · pith:NMOXDXY2new · submitted 2024-09-26 · 💻 cs.LG

SLIDE: A machine-learning based method for forced dynamic response estimation of multibody systems

Peter Manzl , Alexander Humer , Qasim Khadim , Johannes Gerstmayr This is my paper

Pith reviewed 2026-05-23 20:37 UTC · model grok-4.3

classification 💻 cs.LG
keywords machine learningmultibody systemsdynamic responseneural networkssimulation accelerationdamped systemsforced excitationeigenvalue truncation
0
0 comments X

The pith

SLIDE estimates forced dynamic responses of multibody systems without full state by truncating transients via complex eigenvalues.

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

The paper presents SLIDE as a deep learning method to estimate output sequences from mechanical and multibody systems that experience forced excitation. It handles damped cases by truncating the initial part of each output window according to decay rates taken from the complex eigenvalues of the linearized equations. This removes the need to supply the complete system state at each step, which matters for flexible systems where state information is expensive or incomplete. A second network supplies an accompanying error estimate. Tests across the Duffing oscillator, a flexible slider-crank, and an industrial 6R manipulator report speedups reaching several million times real time.

Core claim

SLIDE is a deep learning-based method designed to estimate output sequences of mechanical or multibody systems with primarily forced excitation. A key advantage is its ability to estimate the dynamic response of damped systems without requiring the full system state. The method truncates the output window based on the decay of initial effects such as damping, which is approximated by the complex eigenvalues of the systems linearized equations. In addition a second neural network is trained to provide an error estimation. The method is applied to the Duffing oscillator, a flexible slider-crank system, and an industrial 6R manipulator mounted on a flexible socket, demonstrating speedups from a

What carries the argument

The SLiding-window Initially-truncated Dynamic-response Estimator (SLIDE) that approximates initial-effect decay from complex eigenvalues of the linearized system to truncate output windows without full state.

If this is right

  • The approach yields speedups of several million times and exceeds real-time performance on the tested systems.
  • A companion network supplies per-estimate error bounds that increase the method's practical use.
  • The same truncation strategy works across linear and nonlinear examples including the Duffing oscillator and flexible multibody mechanisms.
  • No full state vector is required at inference time, which removes a common bottleneck for flexible systems.

Where Pith is reading between the lines

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

  • The same truncation idea could be tested on other forced physical systems where only partial observations are available.
  • Real-time deployment on embedded hardware becomes feasible once the network is trained, opening closed-loop control uses.
  • Hybrid pipelines that switch between SLIDE for fast segments and full solvers only when error estimates rise could cut total compute in long design studies.

Load-bearing premise

The decay of initial effects such as damping can be approximated by the complex eigenvalues of the system's linearized equations to allow reliable truncation without full state information.

What would settle it

Run a full nonlinear simulation of one of the tested systems in which measured transient decay deviates markedly from the eigenvalue-based prediction; the truncated SLIDE estimates should then show large errors relative to the reference solution.

Figures

Figures reproduced from arXiv: 2409.18272 by Alexander Humer, Johannes Gerstmayr, Peter Manzl, Qasim Khadim.

Figure 1
Figure 1. Figure 1: The components of the explored surrogate models: using an original model the [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Structure of a general feedforward network. The layers are connected with weights [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The SLIDE method uses an input window of length [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The structure of the proposed error estimator. For better training performance, [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The spring damper model consists of the mass [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: a) The mean square error on the test set while training shown over the epochs. Lines [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visualizing of the weight matrix W′ as a heatmap. The first column shows the homogeneous solution for x0 = 1, which decays over time, represented by the output index. The solution for an arbitrary input vector can be calculated by superimposing the individual solutions - which is what the network’s matrix multiplication yˆ = W′xˆ does. the time-sequence. To construct training and test data, we sample from … view at source ↗
Figure 8
Figure 8. Figure 8: a) The MSE on the validation set of the nonlinear damper over the course of the [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: SLIDE applied to a longer input sequence of the nonlinear spring-damper system [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: The slider-crank model. The connecting rod is flexible and has a deflection [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: a) Examples for applied trajectories to the crankshaft, where angular velocities [PITH_FULL_IMAGE:figures/full_fig_p018_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Processing a longer time-segment for the slider-crank system using the proposed [PITH_FULL_IMAGE:figures/full_fig_p019_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: a) Simulation of the Puma560 manipulator standing on a flexible cylindrical [PITH_FULL_IMAGE:figures/full_fig_p020_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Decay times based on the eigenvalues of the systems, calculated over an exemplarily [PITH_FULL_IMAGE:figures/full_fig_p021_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Continuation by window shifting with error of the surrogate model to the refer [PITH_FULL_IMAGE:figures/full_fig_p022_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: The mean absolute error of the error estimator and the surrogate model on parts [PITH_FULL_IMAGE:figures/full_fig_p023_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: The error of the surrogate model e is plotted over the estimated error ˆe, where e95 marks the 95th percentile of the surrogate model error. Thus, for the training, the surrogate model has an accuracy of e95 = 0.227 mm or better on 95% of the dataset. In the test e95 = 0.273 mm. In the estimator’s training set the surrogate training set is contained. general, by increasing the amount of training data, the… view at source ↗
Figure 18
Figure 18. Figure 18: Speedup of neural network compared to multibody simulation of the shown ma [PITH_FULL_IMAGE:figures/full_fig_p025_18.png] view at source ↗
read the original abstract

In computational engineering, enhancing the simulation speed and efficiency is a perpetual goal. To fully take advantage of neural network techniques and hardware, we present the SLiding-window Initially-truncated Dynamic-response Estimator (SLIDE), a deep learning-based method designed to estimate output sequences of mechanical or multibody systems with primarily, but not exclusively, forced excitation. A key advantage of SLIDE is its ability to estimate the dynamic response of damped systems without requiring the full system state, making it particularly effective for flexible multibody systems. The method truncates the output window based on the decay of initial effects, such as damping, which is approximated by the complex eigenvalues of the systems linearized equations. In addition, a second neural network is trained to provide an error estimation, further enhancing the methods applicability. The method is applied to a diverse selection of systems, including the Duffing oscillator, a flexible slider-crank system, and an industrial 6R manipulator, mounted on a flexible socket. Our results demonstrate significant speedups from the simulation up to several millions, exceeding real-time performance substantially.

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

2 major / 1 minor

Summary. The paper introduces SLIDE, a neural-network method for estimating forced dynamic responses of multibody systems. It uses a sliding-window estimator whose output window is truncated according to decay rates taken from the complex eigenvalues of the linearized equations, thereby avoiding the need for full system state. A second network supplies per-prediction error estimates. The approach is applied to the Duffing oscillator, a flexible slider-crank, and a 6R manipulator on a flexible base, with reported speedups reaching several million times real time.

Significance. If the truncation rule and accuracy claims can be placed on a quantitative footing, the method would offer a practical route to real-time or faster-than-real-time simulation of flexible multibody systems where full-state integration remains expensive. The choice of three qualitatively distinct test cases is a positive feature that supports broader applicability claims.

major comments (2)
  1. [Abstract] Abstract (paragraph on truncation): The central claim that initial transients (including damping) can be reliably truncated using only the complex eigenvalues of the linearized equations, without full state, is load-bearing for the advertised 'state-free' advantage. For the Duffing oscillator the effective decay rate is amplitude- and forcing-dependent; the linearization supplies only the small-signal poles and supplies no a-priori bound on settling time under the nonlinear excursions actually simulated. No quantitative check (e.g., comparison of linear-predicted versus observed settling windows) is referenced.
  2. [Abstract] Abstract (results paragraph): The manuscript reports application to three systems and 'significant speedups' but supplies no numerical error metrics (RMSE, maximum absolute error, etc.), no baseline comparisons against full-order integration or alternative reduced-order models, and no description of the validation protocol or train/test split. Without these data the accuracy of the state-free estimator cannot be assessed.
minor comments (1)
  1. The abstract would be clearer if it stated the network architectures, loss functions, and training hyperparameters used for both the primary estimator and the error network.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We appreciate the referee's thorough review and constructive feedback. The two major comments highlight important aspects for strengthening the presentation of our method. We respond to each below and will make the necessary revisions to the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract (paragraph on truncation): The central claim that initial transients (including damping) can be reliably truncated using only the complex eigenvalues of the linearized equations, without full state, is load-bearing for the advertised 'state-free' advantage. For the Duffing oscillator the effective decay rate is amplitude- and forcing-dependent; the linearization supplies only the small-signal poles and supplies no a-priori bound on settling time under the nonlinear excursions actually simulated. No quantitative check (e.g., comparison of linear-predicted versus observed settling windows) is referenced.

    Authors: We agree that the truncation rule relies on an approximation derived from the eigenvalues of the linearized system and that this is central to the state-free claim. For the Duffing oscillator, nonlinear effects can indeed modulate the effective decay. The linear poles are intended to supply a conservative window that remains valid under the forcing amplitudes used in our experiments. To place this on a firmer quantitative footing, we will add a direct comparison of the linear-predicted settling window against the observed decay of initial transients in the nonlinear simulations, both in the main text and as a supplementary figure. revision: yes

  2. Referee: [Abstract] Abstract (results paragraph): The manuscript reports application to three systems and 'significant speedups' but supplies no numerical error metrics (RMSE, maximum absolute error, etc.), no baseline comparisons against full-order integration or alternative reduced-order models, and no description of the validation protocol or train/test split. Without these data the accuracy of the state-free estimator cannot be assessed.

    Authors: We acknowledge that the abstract emphasizes speedups without accompanying accuracy figures and that the referee could not locate explicit numerical error metrics, baseline comparisons, or validation details. The results sections of the manuscript do contain per-system RMSE and maximum-error values together with comparisons against full-order integration; however, these were not summarized in the abstract and the validation protocol was described only in the methods. We will revise the abstract to report the key error metrics and will ensure the validation protocol and train/test splits are stated clearly and concisely in both the abstract and the main text. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained

full rationale

The paper constructs SLIDE from standard neural-network training on simulation trajectories plus an external truncation rule that approximates initial transient decay via complex eigenvalues of the linearized equations. Neither the NN predictions nor the reported speedups reduce by the paper's own equations to quantities defined only in terms of fitted constants or prior self-citations; the truncation step is presented as a domain-derived modeling choice rather than a self-referential derivation. No load-bearing self-citation chains, self-definitional steps, or fitted-input-renamed-as-prediction patterns appear in the abstract or described method.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The approach rests on standard neural-network training assumptions plus one domain-specific truncation rule; no new physical entities are introduced.

free parameters (1)
  • neural network architecture and training hyperparameters
    Typical for any deep-learning method; specific choices of layers, learning rate, and window sizes are fitted to each example system.
axioms (1)
  • domain assumption Decay of initial transients can be approximated by complex eigenvalues of the linearized system equations
    Invoked to determine the truncation length of the output window.

pith-pipeline@v0.9.0 · 5730 in / 1382 out tokens · 30482 ms · 2026-05-23T20:37:36.877623+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

36 extracted references · 36 canonical work pages · 3 internal anchors

  1. [1]

    Multilayer feedforward networks are uni- versal approximators,

    K. Hornik, M. Stinchcombe, and H. White, “Multilayer feedforward networks are uni- versal approximators,” Neural networks, vol. 2, no. 5, pp. 359–366, 1989

    work page 1989

  2. [2]

    Imagenet classification with deep con- volutional neural networks,

    A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep con- volutional neural networks,” Advances in neural information processing systems, vol. 25, 2012

    work page 2012

  3. [3]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770– 778, 2016

    work page 2016

  4. [4]

    Playing Atari with Deep Reinforcement Learning

    V. Mnih, K. Kavukcuoglu, D. Silver, A. Graves, I. Antonoglou, D. Wierstra, and M. Ried- miller, “Playing atari with deep reinforcement learning,”arXiv preprint arXiv:1312.5602, 2013

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  5. [5]

    Mastering the game of go with deep neural networks and tree search,

    D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre, G. Van Den Driessche, J. Schrit- twieser, I. Antonoglou, V. Panneershelvam, M. Lanctot, et al., “Mastering the game of go with deep neural networks and tree search,” nature, vol. 529, no. 7587, pp. 484–489, 2016

    work page 2016

  6. [6]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural information processing systems, vol. 30, 2017

    work page 2017

  7. [7]

    Deep learning applied to computational mechanics: A comprehensive review, state of the art, and the classics

    L. Vu-Quoc and A. Humer, “Deep learning applied to computational mechanics: A com- prehensive review, state of the art, and the classics,” arXiv preprint arXiv:2212.08989 , 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  8. [8]

    Machine learning in modeling and simulation,

    T. Rabczuk and K.-J. Bathe, “Machine learning in modeling and simulation,” Springer Cham, Switzerland, vol. 10, pp. 978–3, 2023

    work page 2023

  9. [9]

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,

    M. Raissi, P. Perdikaris, and G. E. Karniadakis, “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,” Journal of Computational physics , vol. 378, pp. 686–707, 2019

    work page 2019

  10. [10]

    Hamiltonian neural networks,

    S. Greydanus, M. Dzamba, and J. Yosinski, “Hamiltonian neural networks,” Advances in neural information processing systems , vol. 32, 2019

    work page 2019

  11. [11]

    et al.Lagrangian Neural Networks

    M. Cranmer, S. Greydanus, S. Hoyer, P. Battaglia, D. Spergel, and S. Ho, “Lagrangian neural networks,” arXiv preprint arXiv:2003.04630 , 2020

    work page arXiv 2003

  12. [12]

    Physics-informed neural networks (PINNs) for fluid mechanics: A review,

    S. Cai, Z. Mao, Z. Wang, M. Yin, and G. E. Karniadakis, “Physics-informed neural networks (PINNs) for fluid mechanics: A review,” Acta Mechanica Sinica, vol. 37, no. 12, pp. 1727–1738, 2021. 27

    work page 2021

  13. [13]

    Physics-informed neural networks for heat transfer problems,

    S. Cai, Z. Wang, S. Wang, P. Perdikaris, and G. E. Karniadakis, “Physics-informed neural networks for heat transfer problems,” Journal of Heat Transfer , vol. 143, no. 6, p. 060801, 2021

    work page 2021

  14. [14]

    Data-driven simulation for general-purpose multibody dynamics using deep neural networks,

    H.-S. Choi, J. An, S. Han, J.-G. Kim, J.-Y. Jung, J. Choi, G. Orzechowski, A. Mikkola, and J. H. Choi, “Data-driven simulation for general-purpose multibody dynamics using deep neural networks,” Multibody System Dynamics, vol. 51, pp. 419–454, 2021

    work page 2021

  15. [15]

    A DNN-based data-driven modeling employing coarse sample data for real-time flexible multibody dynamics simu- lations,

    S. Han, H.-S. Choi, J. Choi, J. H. Choi, and J.-G. Kim, “A DNN-based data-driven modeling employing coarse sample data for real-time flexible multibody dynamics simu- lations,” Computer Methods in Applied Mechanics and Engineering , vol. 373, p. 113480, 2021

    work page 2021

  16. [16]

    Data-driven inverse dynamics modeling using neural-networks and regression-based techniques,

    M. Pikuli´ nski, P. Malczyk, and R. Aarts, “Data-driven inverse dynamics modeling using neural-networks and regression-based techniques,” Multibody System Dynamics , pp. 1– 26, 2024

    work page 2024

  17. [17]

    Deep learning for model order reduction of multi- body systems to minimal coordinates,

    A. Angeli, W. Desmet, and F. Naets, “Deep learning for model order reduction of multi- body systems to minimal coordinates,” Computer Methods in Applied Mechanics and Engineering, vol. 373, p. 113517, 2021

    work page 2021

  18. [18]

    Overview of design considerations for data-driven time-stepping schemes applied to nonlinear mechanical systems,

    T. Slimak, A. Zw¨ olfer, B. Todorov, and D. J. Rixen, “Overview of design considerations for data-driven time-stepping schemes applied to nonlinear mechanical systems,”Journal of Computational and Nonlinear Dynamics , vol. 19, no. 7, 2024

    work page 2024

  19. [19]

    A structure-preserving machine learning framework for accurate prediction of struc- tural dynamics for systems with isolated nonlinearities,

    D. A. Najera-Flores, D. D. Quinn, A. Garland, K. Vlachas, E. Chatzi, and M. D. Todd, “A structure-preserving machine learning framework for accurate prediction of struc- tural dynamics for systems with isolated nonlinearities,” Mechanical Systems and Signal Processing, vol. 213, p. 111340, 2024

    work page 2024

  20. [20]

    Multibody models generated from natural language,

    J. Gerstmayr, P. Manzl, and M. Pieber, “Multibody models generated from natural language,” Multibody System Dynamics, pp. 1–23, 2024

    work page 2024

  21. [21]

    Measuring github copilot’s impact on productivity,

    A. Ziegler, E. Kalliamvakou, X. A. Li, A. Rice, D. Rifkin, S. Simister, G. Sittampalam, and E. Aftandilian, “Measuring github copilot’s impact on productivity,” Communica- tions of the ACM , vol. 67, no. 3, pp. 54–63, 2024

    work page 2024

  22. [22]

    Data-driven friction force predic- tion model for hydraulic actuators using deep neural networks,

    S. Han, G. Orzechowski, J.-G. Kim, and A. Mikkola, “Data-driven friction force predic- tion model for hydraulic actuators using deep neural networks,”Mechanism and Machine Theory, vol. 192, p. 105545, 2024

    work page 2024

  23. [23]

    Mbd-node: Physics- informed data-driven modeling and simulation of constrained multibody systems,

    J. Wang, S. Wang, H. M. Unjhawala, J. Wu, and D. Negrut, “Mbd-node: Physics- informed data-driven modeling and simulation of constrained multibody systems,”Multi- body System Dynamics, pp. 1–43, 2024

    work page 2024

  24. [24]

    A survey of projection-based model reduction methods for parametric dynamical systems,

    P. Benner, S. Gugercin, and K. Willcox, “A survey of projection-based model reduction methods for parametric dynamical systems,” SIAM review, vol. 57, no. 4, pp. 483–531, 2015

    work page 2015

  25. [25]

    A. A. Shabana, Dynamics of multibody systems . Cambridge university press, 5th ed., 2020. 28

    work page 2020

  26. [26]

    Exudyn–a c++-based python package for flexible multibody systems,

    J. Gerstmayr, “Exudyn–a c++-based python package for flexible multibody systems,” Multibody System Dynamics, vol. 60, no. 4, pp. 533–561, 2024

    work page 2024

  27. [27]

    Goodfellow, Y

    I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. MIT Press, 2016

    work page 2016

  28. [28]

    Activation functions in deep learning: A comprehensive survey and benchmark,

    S. R. Dubey, S. K. Singh, and B. B. Chaudhuri, “Activation functions in deep learning: A comprehensive survey and benchmark,” Neurocomputing, vol. 503, pp. 92–108, 2022

    work page 2022

  29. [29]

    How neural networks extrapolate: From feedforward to graph neural networks,

    K. Xu, J. Li, M. Zhang, S. S. Du, K. Kawarabayashi, and S. Jegelka, “How neural networks extrapolate: From feedforward to graph neural networks,” CoRR, vol. abs/2009.11848, 2020

    work page arXiv 2009

  30. [30]

    Understanding the difficulty of training deep feedforward neural networks,

    X. Glorot and Y. Bengio, “Understanding the difficulty of training deep feedforward neural networks,” in Proceedings of the thirteenth international conference on artificial intelligence and statistics , pp. 249–256, JMLR Workshop and Conference Proceedings, 2010

    work page 2010

  31. [31]

    Adam: A Method for Stochastic Optimization

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  32. [32]

    Pytorch 2: Faster machine learning through dynamic python bytecode transformation and graph compilation,

    J. Ansel, E. Yang, H. He, N. Gimelshein, A. Jain, M. Voznesensky, B. Bao, P. Bell, D. Be- rard, E. Burovski, et al., “Pytorch 2: Faster machine learning through dynamic python bytecode transformation and graph compilation,” in Proceedings of the 29th ACM Inter- national Conference on Architectural Support for Programming Languages and Operating Systems, V...

    work page 2024

  33. [33]

    On the correct representation of bending and axial de- formation in the absolute nodal coordinate formulation with an elastic line approach,

    J. Gerstmayr and H. Irschik, “On the correct representation of bending and axial de- formation in the absolute nodal coordinate formulation with an elastic line approach,” Journal of Sound and Vibration , vol. 318, no. 3, pp. 461–487, 2008

    work page 2008

  34. [34]

    Not your grandmother’s toolbox–the robotics toolbox rein- vented for python,

    P. Corke and J. Haviland, “Not your grandmother’s toolbox–the robotics toolbox rein- vented for python,” in 2021 IEEE international conference on robotics and automation (ICRA), pp. 11357–11363, IEEE, 2021

    work page 2021

  35. [35]

    The explicit dynamic model and inertial parameters of the puma 560 arm,

    B. Armstrong, O. Khatib, and J. Burdick, “The explicit dynamic model and inertial parameters of the puma 560 arm,” in Proceedings. 1986 IEEE international conference on robotics and automation , vol. 3, pp. 510–518, IEEE, 1986

    work page 1986

  36. [36]

    The enhancement of the mechanical properties of a high-density polyethylene,

    G. Kalay, R. A. Sousa, R. L. Reis, A. M. Cunha, and M. J. Bevis, “The enhancement of the mechanical properties of a high-density polyethylene,” Journal of applied polymer science, vol. 73, no. 12, pp. 2473–2483, 1999. 29

    work page 1999