pith. sign in

arxiv: 2605.04229 · v1 · submitted 2026-05-05 · 💻 cs.LG · cond-mat.mtrl-sci

Capabilities of Auto-encoders and Principal Component Analysis of the Reduction of Microstructural Images; Application on the Acceleration of Phase-Field Simulations

Seifallah Fetni , Thinh Quy Duc Pham , Truong Vinh Hoang , Hoang Son Tran , Laurent Duch\^ene , Xuan-Van Tran , Anne Marie Habraken This is my paper

Pith reviewed 2026-05-08 17:00 UTC · model grok-4.3

classification 💻 cs.LG cond-mat.mtrl-sci
keywords auto-encoderprincipal component analysisdimensionality reductionphase-field simulationmicrostructural imagesLSTMtime-series predictionneural networks
0
0 comments X

The pith

Auto-encoders combined with principal component analysis reduce phase-field microstructural images by a factor of 196 while retaining over 80 percent accuracy, allowing LSTM networks to predict future frames and accelerate the simulations.

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

The paper constructs a dataset from high-fidelity phase-field simulations of microstructures and tests whether auto-encoders paired with principal component analysis can compress the images to 1/196 of their original size while preserving more than 80 percent accuracy. It further shows that long short-term memory networks can forecast the next frame from this compressed latent representation. If successful, the approach would let researchers perform time-series analysis and forward predictions in the low-dimensional space, removing the need for repeated full-resolution computations on high-performance hardware.

Core claim

The association of auto-encoder neural networks and principal component analyses leads to efficient and significant dimensionality reduction of microstructural images at a 1/196 ratio with more than 80 percent accuracy. Long short-term memory networks applied to the resulting latent space can generate next-frame predictions, which in turn enable acceleration of phase-field simulations without high computing resources.

What carries the argument

A data-driven pipeline that first compresses phase-field image sequences via auto-encoders and principal component analysis into a low-dimensional latent space, then applies LSTM (and GRU) networks for sequence prediction within that space.

If this is right

  • Analyses such as statistical or physical queries can be performed directly inside the latent space instead of the full image domain.
  • Phase-field simulations become feasible on modest hardware by replacing each expensive time step with a fast LSTM prediction.
  • The same reduction-plus-prediction pipeline can be applied to other time-evolving image datasets in materials research.
  • Alternative reduction methods (plain auto-encoders, PCA alone, or simple feed-forward networks) and alternative predictors (GRU) become directly comparable on the same latent space.

Where Pith is reading between the lines

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

  • The compressed latent trajectories could be used to train surrogate models that couple directly to experimental imaging data.
  • Long-horizon predictions in latent space might expose asymptotic steady states or coarsening laws without completing a full simulation run.
  • The framework suggests a route to hybrid models that embed the learned latent dynamics inside physics-informed constraints.
  • Systematic testing across different material systems and boundary conditions would reveal how universal the 80 percent accuracy threshold actually is.

Load-bearing premise

The information retained after 196-fold compression is still sufficient for LSTM predictions to stay accurate and physically consistent with the original phase-field evolution over many time steps.

What would settle it

Run the LSTM predictor for dozens of successive frames on a held-out test sequence and check whether the reconstructed microstructures reproduce the same grain-boundary motion and energy dissipation curves obtained from an independent full-resolution phase-field simulation.

Figures

Figures reproduced from arXiv: 2605.04229 by Anne Marie Habraken, Hoang Son Tran, Laurent Duch\^ene, Seifallah Fetni, Thinh Quy Duc Pham, Truong Vinh Hoang, Xuan-Van Tran.

Figure 1
Figure 1. Figure 1: Machine-learning based framework (surrogate model) for the dimensionality reduction of view at source ↗
Figure 2
Figure 2. Figure 2: Examples of the microstructure evolutions of binary alloys under spinodal decomposition for view at source ↗
Figure 3
Figure 3. Figure 3: Results of the training of auto-encoders with different code sizes (2000, 1000, 750, 500 and 250 view at source ↗
Figure 4
Figure 4. Figure 4: Reconstructed images obtained from different saved auto-encoders models. view at source ↗
Figure 5
Figure 5. Figure 5: Training results of the 2nd auto-encoder for various shapes and number of hidden layers (HL) ; (a) evolution of the MSE and validation MSE for 9 HL and different code sizes. (b) MSE and validation MSE for different number of HL and code sizes. 12 view at source ↗
Figure 6
Figure 6. Figure 6: Explained variance obtained by the application of PCA on view at source ↗
Figure 7
Figure 7. Figure 7: (a) Illustration of the flow of dimensionality reduction. Near net-shape images could be obtained view at source ↗
Figure 8
Figure 8. Figure 8: Results of training of time series algorithms; an emphasis was put on LSTM predictions. view at source ↗
read the original abstract

In this work, a data-driven framework based on Phase-Field simulations data is proposed to highlight the capabilities of neural networks to ensure accurate low dimensionality reduction of simulated microstructural images and to provide time-series analysis. The dataset was indeed constructed from high-fidelity Phase-Field simulations. Analyses demonstrated that the association of auto-encoder neural networks and principal component analyses leads to ensure efficient and significant dimensionality reduction: 1/196 of reduction ratio with more than 80% of accuracy. These findings give insight to apply analyses on data from the latent dimension. Application of Long Short Term Memory (LSTM) neural networks showed the possibility of making next frame predictions; that makes possible the acceleration of Phase-Field simulation without the need of high computing resources. We discussed the application of such a framework on various areas of research. Different methods are proposed from the conducted analyses, in order to ensure dimensionality reduction, including auto-encoders, principal component analysis and Artificial Neural Networks, and time-series analysis, including LSTM and Gated Recurrent Unit (GRU).

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 / 2 minor

Summary. The paper proposes a data-driven approach using phase-field simulation data to show that auto-encoder neural networks combined with principal component analysis can reduce microstructural image dimensionality by a factor of 1/196 while achieving more than 80% accuracy. It further applies LSTM (and mentions GRU) networks for next-frame prediction in the latent space, arguing this enables acceleration of phase-field simulations without high computational resources. Broader applications to other research areas are discussed.

Significance. If the central claims hold, the work could contribute to practical acceleration of expensive phase-field models in materials science by leveraging dimensionality reduction and time-series prediction. The empirical use of standard ML pipelines on simulation data is noted, but the manuscript provides no machine-checked proofs, reproducible code, parameter-free derivations, or falsifiable physics-based predictions.

major comments (2)
  1. [Abstract] Abstract: the claim of 'more than 80% of accuracy' for the 1/196 reduction provides no definition of the accuracy metric (e.g., reconstruction MSE, SSIM, or latent-space prediction error), no error bars, and no baseline comparisons (e.g., to PCA alone).
  2. [Abstract] Abstract: the acceleration claim via LSTM next-frame predictions requires evidence that multi-step rollouts in the reduced space, once decoded, preserve dynamical invariants of the phase-field model (interface curvature, phase fractions, free-energy dissipation); no such multi-step results or consistency checks with the underlying Allen-Cahn/Cahn-Hilliard dynamics are described.
minor comments (2)
  1. The reduction ratio '1/196' is stated without specifying the original image dimensions or the exact latent dimension size used.
  2. Network architectures, hyperparameters, and training details for the auto-encoder, PCA, and LSTM components should be reported explicitly to support reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major comment below and indicate the revisions we will make to improve clarity and address the concerns raised.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim of 'more than 80% of accuracy' for the 1/196 reduction provides no definition of the accuracy metric (e.g., reconstruction MSE, SSIM, or latent-space prediction error), no error bars, and no baseline comparisons (e.g., to PCA alone).

    Authors: We agree that the abstract should explicitly define the accuracy metric. The manuscript quantifies accuracy via reconstruction error in image space following the combined auto-encoder and PCA reduction. We will revise the abstract to state this definition, note that results are from single runs (hence no error bars are reported), and add a direct comparison to PCA alone, which is already analyzed in the main text. revision: yes

  2. Referee: [Abstract] Abstract: the acceleration claim via LSTM next-frame predictions requires evidence that multi-step rollouts in the reduced space, once decoded, preserve dynamical invariants of the phase-field model (interface curvature, phase fractions, free-energy dissipation); no such multi-step results or consistency checks with the underlying Allen-Cahn/Cahn-Hilliard dynamics are described.

    Authors: The manuscript presents single-step next-frame predictions in latent space to demonstrate the feasibility of the approach for potential acceleration. We acknowledge that multi-step rollouts and explicit verification of dynamical invariants (such as interface curvature, phase fractions, or free-energy dissipation) against the underlying Allen-Cahn/Cahn-Hilliard equations are not included. We will revise the abstract and add a limitations paragraph in the discussion to clarify that the acceleration is proposed as a promising direction based on the observed prediction capability, rather than a fully validated multi-step outcome, and identify this as an avenue for future work. revision: partial

Circularity Check

0 steps flagged

No circularity; standard empirical ML application on simulation data

full rationale

The paper applies auto-encoders combined with PCA to reduce phase-field microstructural images to 1/196 dimensionality while reporting >80% accuracy, then uses LSTM networks for next-frame latent predictions to accelerate simulations. No derivation chain, equations, or self-citations reduce the reported reduction ratio, accuracy metric, or multi-step prediction performance back to quantities defined or fitted inside the same work by construction. The results are presented as outcomes of training standard architectures on an external dataset generated from high-fidelity simulations, with no uniqueness theorems, ansatzes smuggled via prior self-work, or renaming of known patterns as new derivations. The framework is self-contained as an application study.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claims rest on the assumption that dimensionality reduction via AE+PCA preserves enough information for LSTM time-series prediction to remain useful, plus standard neural-network training assumptions.

free parameters (1)
  • latent dimension size
    Chosen to achieve the stated 1/196 reduction ratio while meeting the accuracy threshold; value not reported in abstract.
axioms (1)
  • domain assumption Microstructural images from phase-field simulations lie on a low-dimensional manifold that autoencoders and PCA can recover.
    Invoked when claiming that 1/196 compression retains >80% accuracy.

pith-pipeline@v0.9.0 · 5524 in / 1420 out tokens · 58195 ms · 2026-05-08T17:00:32.268808+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

37 extracted references · 34 canonical work pages

  1. [1]

    Y. Gu, X. He, D. Han, On the phase-field modeling of rapid solidification, Computational Materials Science 199 (2021) 110812.doi:https://doi.org/10.1016/ j.commatsci.2021.110812

    work page arXiv 2021

  2. [2]

    Lindroos, T

    M. Lindroos, T. Pinomaa, K. Ammar, A. Laukkanen, N. Provatas, S. Forest, Dislocation density in cellular rapid solidification using phase field modeling and crystal plasticity, International Journal of Plasticity 148 (2022) 103139.doi:https://doi.org/10.1016/ j.ijplas.2021.103139

    work page arXiv 2022

  3. [3]

    Boisse, N

    J. Boisse, N. Lecoq, R. Patte, H. Zapolsky, Phase-field simulation of coarsening ofγ precipitates in an orderedγ′ matrix, Acta Materialia 55 (18) (2007) 6151–6158.doi: https://doi.org/10.1016/j.actamat.2007.07.014

    work page doi:10.1016/j.actamat.2007.07.014 2007

  4. [4]

    X. Li, Y. Xu, Phase field modeling scheme with mesostructure for crack propagation in concrete composite, International Journal of Solids and Structures 234-235 (2022) 111259.doi:https://doi.org/10.1016/j.ijsolstr.2021.111259

    work page doi:10.1016/j.ijsolstr.2021.111259 2022

  5. [5]

    F. Wang, H. Huang, D. Zhang, M. Zhou, Cracking feature and mechanical behavior of shield tunnel lining simulated by a phase-field modeling method based on spectral decomposition, Tunnelling and Underground Space Technology 119 (2022) 104246.doi: https://doi.org/10.1016/j.tust.2021.104246

    work page doi:10.1016/j.tust.2021.104246 2022

  6. [6]

    L. K. Aagesen, D. Schwen, K. Ahmed, M. R. Tonks, Quantifying elastic energy effects on interfacial energy in the kim-kim-suzuki phase-field model with different interpolation schemes, Computational Materials Science 140 (2017) 10–21. 17

    2017

  7. [7]

    Kumbhar, P

    A. Kumbhar, P. G. Dhawale, S. Kumbhar, U. Patil, P. Magdum, A comprehensive review: Machine learning and its application in integrated power system, Energy Reports 7 (2021) 5467–5474.doi:https://doi.org/10.1016/j.egyr.2021.08.133

    work page doi:10.1016/j.egyr.2021.08.133 2021

  8. [8]

    Rabbani, G

    N. Rabbani, G. Y. Kim, C. J. Suarez, J. H. Chen, Applications of machine learning in routine laboratory medicine: Current state and future directions, Clinical Biochemistry (2022).doi:https://doi.org/10.1016/j.clinbiochem.2022.02.011

    work page doi:10.1016/j.clinbiochem.2022.02.011 2022

  9. [9]

    T. Q. D. Pham, T. V. Hoang, X. V. Tran, Q. T. Pham, S. Fetni, L. Duchêne, H. S. Tran, A.-M. Habraken, Fast and accurate prediction of temperature evolutions in additive manufacturing process using deep learnin, Journal of Intelligent Manufacturing 36 (2020) 101641.doi:https://doi.org/10.1007/s10845-021-01896-8

    work page doi:10.1007/s10845-021-01896-8 2020

  10. [10]

    Fetni, Q

    S. Fetni, Q. D. T. Pham, V. X. Tran, L. Duchêne, H. S. Tran, A. M. Habraken, Thermal field prediction in ded manufacturing process using artificial neural network, 24th International Conference on Material Forming (2021).doi:https://doi:10.25518/ esaform21.2812

    2021

  11. [11]

    T. Pham, T. Hoang, X. Tran, S. Fetni, L. Duchêne, H. Tran, A. Habraken, Characterization, propagation, and sensitivity analysis of uncertainties in the directed energy deposition process using a deep learning-based surrogate model, Probabilistic Engineering Mechanics 69 (2022) 103297.doi:https://doi.org/10. 1016/j.probengmech.2022.103297

    work page arXiv 2022

  12. [12]

    Ibragimova, A

    O. Ibragimova, A. Brahme, W. Muhammad, J. Lévesque, K. Inal, A new ann based crystal plasticity model for fcc materials and its application to non-monotonic strain paths, International Journal of Plasticity 144 (2021) 103059.doi:https://doi.org/ 10.1016/j.ijplas.2021.103059

    work page doi:10.1016/j.ijplas.2021.103059 2021

  13. [13]

    Materials Science and Engineering: A 590, 66–73

    T. Bhattacharyya, S. Brat Singh, S. Sikdar (Dey), S. Bhattacharyya, W. Bleck, D. Bhattacharjee, Microstructural prediction through artificial neural network (ann) for development of transformation induced plasticity (trip) aided steel, Materials Science and Engineering: A 565 (2013) 148–157.doi:https://doi.org/10.1016/j.msea. 2012.11.110

    work page doi:10.1016/j.msea 2013

  14. [14]

    J. Gola, J. Webel, D. Britz, A. Guitar, T. Staudt, M. Winter, F. Mücklich, Objective microstructure classification by support vector machine (svm) using a combination of morphological parameters and textural features for low carbon steels, Computational Materials Science 160 (2019) 186–196.doi:https://doi.org/10.1016/j.commatsci. 2019.01.006

    work page doi:10.1016/j.commatsci 2019

  15. [15]

    Kptr:Key point transformer for lidar-based 3d object detection

    S. Gupta, J. Sarkar, M. Kundu, N. Bandyopadhyay, S. Ganguly, Automatic recognition of sem microstructure and phases of steel using lbp and random decision forest operator, Measurement 151 (2020) 107224.doi:https://doi.org/10.1016/j.measurement. 2019.107224

    work page doi:10.1016/j.measurement 2020

  16. [16]

    Synthesis, properties, and applications of carbyne nanocrystals

    H. Fırat, M. E. Asker, D. Hanbay, Classification of hyperspectral remote sensing images using different dimension reduction methods with 3d/2d cnn, Remote Sensing 18 Applications: Society and Environment 25 (2022) 100694.doi:https://doi.org/10. 1016/j.rsase.2022.100694

    work page arXiv 2022

  17. [17]

    D. W. Abueidda, S. Koric, N. A. Sobh, H. Sehitoglu, Deep learning for plasticity and thermo-viscoplasticity, International Journal of Plasticity 136 (2021) 102852.doi: https://doi.org/10.1016/j.ijplas.2020.102852

    work page doi:10.1016/j.ijplas.2020.102852 2021

  18. [18]

    L. Wu, V. D. Nguyen, N. G. Kilingar, L. Noels, A recurrent neural network-accelerated multi-scale model for elasto-plastic heterogeneous materials subjected to random cyclic and non-proportional loading paths, Computer Methods in Applied Mechanics and Engineering 369 (2020) 113234.doi:https://doi.org/10.1016/j.cma.2020.113234

    work page doi:10.1016/j.cma.2020.113234 2020

  19. [19]

    Rehmer, A

    A. Rehmer, A. Kroll, On the vanishing and exploding gradient problem in gated recurrent units, IFAC-PapersOnLine 53 (2) (2020) 1243–1248, 21st IFAC World Congress.doi:https://doi.org/10.1016/j.ifacol.2020.12.1342

    work page doi:10.1016/j.ifacol.2020.12.1342 2020

  20. [20]

    R. Meka, A. Alaeddini, K. Bhaganagar, A robust deep learning framework for short- term wind power forecast of a full-scale wind farm using atmospheric variables, Energy 221 (2021) 119759.doi:https://doi.org/10.1016/j.energy.2021.119759

    work page doi:10.1016/j.energy.2021.119759 2021

  21. [21]

    T. Shi, S. Huang, L. Chen, Y. Heng, Z. Kuang, L. Xu, H. Mei, A molecular generative model of adam10 inhibitors by using gru-based deep neural network and transfer learning, Chemometrics and Intelligent Laboratory Systems 205 (2020) 104122. doi:https://doi.org/10.1016/j.chemolab.2020.104122

    work page doi:10.1016/j.chemolab.2020.104122 2020

  22. [22]

    Huang, X

    L. Huang, X. Hong, Z. Yang, Y. Liu, B. Zhang, Cnn-lstm network-based damage detection approach for copper pipeline using laser ultrasonic scanning, Ultrasonics 121 (2022) 106685.doi:https://doi.org/10.1016/j.ultras.2022.106685

    work page doi:10.1016/j.ultras.2022.106685 2022

  23. [23]

    de Oca Zapiain, J

    M. de Oca Zapiain, J. D., Stewart, R. Dingreville, Accelerating phase-field- based microstructure evolution predictions via surrogate models trained by machine learning methods, npj Comput Mater 7,3 (2021).doi:https://doi.org/10.1038/ s41524-020-00471-8

    2021

  24. [24]

    C. Xu, S. Gao, M. Li, A novel pca-based microstructure descriptor for heterogeneous material design, Computational Materials Science 130 (2017) 39–49.doi:https://doi. org/10.1016/j.commatsci.2016.12.031

    work page doi:10.1016/j.commatsci.2016.12.031 2017

  25. [26]

    C. Hu, S. Martin, R. Dingreville, Accelerating phase-field predictions via recurrent neural networks learning the microstructure evolution in latent space, Computer Methods in Applied Mechanics and Engineering 397 (2022) 115128.doi:https: //doi.org/10.1016/j.cma.2022.115128. 19

    work page doi:10.1016/j.cma.2022.115128 2022

  26. [28]

    Chevrot, A

    A. Chevrot, A. Vernotte, B. Legeard, Cae: Contextual auto-encoder for multivariate time-series anomaly detection in air transportation, Computers & Security 116 (2022) 102652.doi:https://doi.org/10.1016/j.cose.2022.102652

    work page doi:10.1016/j.cose.2022.102652 2022

  27. [29]

    Alahmadi, N

    A. Alahmadi, N. Alkhraan, W. BinSaeedan, Mpsautodetect: A malicious powershell script detection model based on stacked denoising auto-encoder, Computers & Security 116 (2022) 102658.doi:https://doi.org/10.1016/j.cose.2022.102658

    work page doi:10.1016/j.cose.2022.102658 2022

  28. [30]

    Fetni, J

    S. Fetni, J. Delahaye, L. Duchêne, A. Mertens, A. M. Habraken, Adaptive time stepping approach for phase-field modeling of phase separation and precipitates coarsening in additive manufacturing alloys, XVI International Conference on Computational Plasticity : FEM and Particle-Based Methods/Discrete Element Methods, Proceedings of COMPLAS 2021 (2021).doi:...

    work page doi:10.23967/complas.2021.009 2021

  29. [31]

    T. Kerr, K. Duncan, L. Myers, Post fire materials identification by micro-raman spectroscopy and principal components analysis, Journal of Analytical and Applied Pyrolysis 102 (2013) 103–113.doi:https://doi.org/10.1016/j.jaap.2013.03.008

    work page doi:10.1016/j.jaap.2013.03.008 2013

  30. [32]

    L. Sun, K. Wang, L. Xu, C. Zhang, T. Balezentis, A time-varying distance based interval-valued functional principal component analysis method – a case study of consumer price index, Information Sciences 589 (2022) 94–116.doi:https://doi. org/10.1016/j.ins.2021.12.113

    work page doi:10.1016/j.ins.2021.12.113 2022

  31. [33]

    Lozano, A

    K. ArunKumar, D. V. Kalaga, C. Mohan Sai Kumar, M. Kawaji, T. M. Brenza, Comparative analysis of gated recurrent units (gru), long short-term memory (lstm) cells, autoregressive integrated moving average (arima), seasonal autoregressive integrated moving average (sarima) for forecasting covid-19 trends, Alexandria Engineering Journal 61 (10) (2022) 7585–7...

    work page doi:10.1016/j 2022

  32. [34]

    S. Gao, Y. Huang, S. Zhang, J. Han, G. Wang, M. Zhang, Q. Lin, Short-term runoff prediction with gru and lstm networks without requiring time step optimization during sample generation, Journal of Hydrology 589 (2020) 125188.doi:https://doi.org/ 10.1016/j.jhydrol.2020.125188

    work page doi:10.1016/j.jhydrol.2020.125188 2020

  33. [35]

    J.-Y. Wu, Learning analytics on structured and unstructured heterogeneous data sources: Perspectives from procrastination, help-seeking, and machine-learning defined cognitive engagement, Computers & Education 163 (2021) 104066.doi:https://doi. org/10.1016/j.compedu.2020.104066

    work page doi:10.1016/j.compedu.2020.104066 2021

  34. [36]

    Pospelov, A

    N. Pospelov, A. Tetereva, O. Martynova, K. Anokhin, The laplacian eigenmaps dimensionality reduction of fmri data for discovering stimulus-induced changes in the 20 resting-state brain activity, Neuroimage: Reports 1 (3) (2021) 100035.doi:https: //doi.org/10.1016/j.ynirp.2021.100035

    work page doi:10.1016/j.ynirp.2021.100035 2021

  35. [37]

    N. Oune, R. Bostanabad, Latent map gaussian processes for mixed variable metamodeling, Computer Methods in Applied Mechanics and Engineering 387 (2021) 114128.doi:https://doi.org/10.1016/j.cma.2021.114128

    work page doi:10.1016/j.cma.2021.114128 2021

  36. [38]

    Bencheikh, M

    F. Bencheikh, M. Harkat, A. Kouadri, A. Bensmail, New reduced kernel pca for fault detection and diagnosis in cement rotary kiln, Chemometrics and Intelligent Laboratory Systems 204 (2020) 104091.doi:https://doi.org/10.1016/j.chemolab. 2020.104091

    work page doi:10.1016/j.chemolab 2020

  37. [39]

    Y. Wei, C. Wu, G. Li, H. Shi, Sequential transformer via an outside-in attention for image captioning, Engineering Applications of Artificial Intelligence 108 (2022) 104574. doi:https://doi.org/10.1016/j.engappai.2021.104574. 21

    work page doi:10.1016/j.engappai.2021.104574 2022