pith. machine review for the scientific record. sign in

arxiv: 2604.06935 · v1 · submitted 2026-04-08 · ⚛️ physics.app-ph

Recognition: unknown

Determination of Nanoparticle and Microdroplet Parameters in Levitating Microdroplets of Suspension by Speckle Image Analysis Using Convolutional Neural Networks

Daniel Jakubczyk, Kwasi Nyandey, Yaroslav Shopa

Pith reviewed 2026-05-10 16:48 UTC · model grok-4.3

classification ⚛️ physics.app-ph
keywords speckle image analysisconvolutional neural networkslevitating microdropletsnanoparticle suspensionsoptical diagnosticsparameter classificationelectrodynamic trap
0
0 comments X

The pith

Convolutional neural networks classify droplet diameter, nanoparticle concentration, and nanoparticle diameter from speckle images of levitating microdroplets.

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

The paper tests whether laser speckle images from single levitating microdroplets of nanoparticle suspensions can serve as input for convolutional neural networks to recognize droplet diameter, nanoparticle concentration, and nanoparticle diameter. This approach addresses the inverse problem of interpreting scattered light patterns, which normally depends on both the dispersed particles and the droplet's own size and structure. Experiments used slowly evaporating monodisperse TiO2 suspensions in diethylene glycol held in a linear electrodynamic quadrupole trap. Networks were trained first on separate classification tasks and then on combined two- and three-parameter tasks. The results indicate that droplet diameter can be recovered reliably while simultaneous classification across multiple parameters remains feasible under the tested conditions.

Core claim

Speckle images from levitating microdroplets of monodisperse TiO2 nanoparticle suspensions in diethylene glycol were recorded and fed to convolutional neural networks trained to classify droplet diameter, nanoparticle concentration, and nanoparticle diameter. Under the present experimental conditions, droplet diameter was identified with good reliability, with an estimated accuracy better than approximately 6% for the tested dataset. Nanoparticle concentration yielded useful discrimination when classes were sufficiently separated, and nanoparticle diameter was classified unambiguously in the selected cases. Simultaneous classification of up to three parameters across 27 classes was alsoach 0

What carries the argument

Convolutional neural network trained on speckle images for single-task and multi-task classification of droplet and nanoparticle parameters.

If this is right

  • Droplet diameter can be recovered with accuracy better than 6% under the reported trap and suspension conditions.
  • Nanoparticle concentration can be discriminated provided the concentration classes are spaced far enough apart.
  • Nanoparticle diameter classification succeeds unambiguously for the diameter values examined.
  • Simultaneous classification of three parameters across 27 classes remains possible with a single network.

Where Pith is reading between the lines

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

  • The same speckle-analysis pipeline could be tested on evaporating droplets containing other nanoparticle materials or carrier liquids to check transferability.
  • Integration with real-time imaging in electrodynamic traps might allow continuous tracking of all three parameters without separate instruments.
  • Extension to polydisperse or mixed suspensions would require retraining on correspondingly varied experimental images.

Load-bearing premise

The recorded speckle images contain distinguishable, learnable features for the chosen parameters and the limited experimental dataset is representative enough to support generalization.

What would settle it

A new collection of speckle images from droplets with independently measured parameters, processed by the same trained network, that produces classification errors substantially larger than 6% for droplet diameter or collapses multi-parameter accuracy.

read the original abstract

The optical response of a suspension microdroplet is governed not only by the properties of the dispersed phase, but also by the finite size and optical structure of the droplet itself. As a result, the interpretation of scattered-light patterns from such systems constitutes a non-trivial inverse problem. In this work, we examine whether laser speckle images recorded from single levitating microdroplets of suspension can be used for data-driven recognition of selected droplet and suspension parameters. Experiments were performed on slowly evaporating microdroplets of monodisperse TiO$_2$ nanoparticle suspensions in diethylene glycol confined in a linear electrodynamic quadrupole trap. Speckle images were analyzed with a convolutional neural network trained to classify droplet diameter, nanoparticle concentration, and nanoparticle diameter, first in separate tasks and then in combined two-parameter and three-parameter classifications. Under the present experimental conditions, droplet diameter was identified with good reliability, with an estimated accuracy better than approximately 6% for the tested dataset. Nanoparticle concentration was more difficult to resolve, but useful discrimination was obtained when concentration classes were sufficiently separated. Nanoparticle diameter was also classified unambiguously for the selected cases. In addition, simultaneous classification of up to three parameters across 27 classes was achieved. These results suggest that CNN-based analysis of speckle images may provide a viable route toward multi-parameter optical diagnostics of free suspension microdroplets and, potentially, more complex aerosol-like systems.

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 / 0 minor

Summary. The manuscript reports experiments on slowly evaporating monodisperse TiO2 nanoparticle suspensions in diethylene glycol levitated in a linear electrodynamic quadrupole trap. Laser speckle images are processed with convolutional neural networks to classify droplet diameter, nanoparticle concentration, and nanoparticle diameter, first separately and then in simultaneous two- and three-parameter tasks (up to 27 classes). The central claim is that droplet diameter is recovered with accuracy better than ~6% on the tested dataset, with useful concentration discrimination when classes are well-separated and unambiguous nanoparticle-diameter classification in selected cases.

Significance. If the reported classification accuracies prove robust, the work would demonstrate a practical data-driven route to multi-parameter optical diagnostics of suspension microdroplets without requiring explicit scattering models. This could be relevant for aerosol characterization and levitated-particle studies where inverse scattering problems are otherwise intractable.

major comments (2)
  1. [Abstract / Methods] Abstract and experimental-methods description: accuracies (droplet diameter <6%, multi-parameter 27-class results) are stated only “for the tested dataset,” yet no dataset size, image count per class, train/test split, cross-validation scheme, or hold-out strategy is supplied. Without these, it is impossible to determine whether the CNN performance reflects learnable scattering features or overfitting to the narrow experimental conditions (fixed trap geometry, slow evaporation, single suspension chemistry).
  2. [Results (multi-parameter classification)] Results on simultaneous three-parameter classification: evaporation couples droplet diameter and nanoparticle concentration in the slowly evaporating droplets, so speckle images may contain correlated artifacts. The manuscript does not report any test that decouples these variables (e.g., hold-out droplets with altered evaporation rates or trap voltages) or quantifies how much the network relies on such correlations versus intrinsic parameter-specific features.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments on our manuscript. We address each major comment point by point below, indicating the revisions made where appropriate. Our responses focus on clarifying the experimental and analytical procedures while acknowledging limitations in the current dataset.

read point-by-point responses

  1. Referee: [Abstract / Methods] Abstract and experimental-methods description: accuracies (droplet diameter <6%, multi-parameter 27-class results) are stated only “for the tested dataset,” yet no dataset size, image count per class, train/test split, cross-validation scheme, or hold-out strategy is supplied. Without these, it is impossible to determine whether the CNN performance reflects learnable scattering features or overfitting to the narrow experimental conditions (fixed trap geometry, slow evaporation, single suspension chemistry).

    Authors: We agree that the absence of these details in the original submission limits the ability to fully assess robustness and potential overfitting. In the revised manuscript we have expanded the Methods section with the requested information on total image count, per-class distribution, train/test split ratios, and cross-validation scheme. These additions document the procedures used to train and evaluate the CNN on the collected speckle images under the described experimental conditions. revision: yes

  2. Referee: [Results (multi-parameter classification)] Results on simultaneous three-parameter classification: evaporation couples droplet diameter and nanoparticle concentration in the slowly evaporating droplets, so speckle images may contain correlated artifacts. The manuscript does not report any test that decouples these variables (e.g., hold-out droplets with altered evaporation rates or trap voltages) or quantifies how much the network relies on such correlations versus intrinsic parameter-specific features.

    Authors: The referee correctly notes the physical coupling between droplet diameter and nanoparticle concentration during slow evaporation, which is inherent to the experimental setup described in the manuscript. The training images were acquired across multiple evaporation stages for each levitated droplet, so the network encountered a range of coupled states. While we did not perform dedicated hold-out experiments with deliberately altered evaporation rates or trap voltages, we have added a paragraph in the revised Discussion section that explicitly addresses this coupling, its possible contribution to the observed classification performance, and the implications for interpreting the multi-parameter results. We view this as a limitation of the present study rather than a flaw in the reported accuracies for the tested conditions. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical CNN classification on experimental data

full rationale

The manuscript presents a data-driven experimental study: speckle images from levitating TiO2/DEG microdroplets are recorded and used to train CNN classifiers for droplet diameter, nanoparticle concentration, and nanoparticle diameter (separately and jointly). Reported accuracies (e.g., <6% for diameter on the tested dataset) are empirical performance measures on held-out or cross-validated images from the same narrow experimental campaign. No derivation chain, governing equations, first-principles model, or ansatz is invoked whose output is shown to be identical to its inputs by construction. No self-citations are used to justify uniqueness or load-bearing premises. The approach is therefore self-contained as supervised learning on measured images; any generalization concerns (dataset narrowness, possible evaporation-induced correlations) belong to correctness or overfitting risk, not circularity.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that speckle patterns encode the target parameters in a manner learnable by CNNs from the specific experimental images; the CNN itself introduces a large number of fitted weights, but these are implicit in any supervised learning approach.

free parameters (2)
  • CNN network weights and hyperparameters
    Trained on the collected speckle images to minimize classification error; exact architecture and regularization choices are not specified in the abstract.
  • Concentration and diameter class boundaries
    Chosen to achieve 'useful discrimination' when classes are sufficiently separated.
axioms (1)
  • domain assumption Speckle images from the levitated droplets contain sufficient information to distinguish the chosen parameter classes
    Stated as the premise for data-driven recognition in the abstract.

pith-pipeline@v0.9.0 · 5568 in / 1328 out tokens · 46480 ms · 2026-05-10T16:48:45.867143+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

25 extracted references · 22 canonical work pages

  1. [1]

    Giorgini, S

    A. Giorgini, S. Avino, P. Malara, P. De Natale, G. Gagliardi, Liquid Droplet Microresonators, Sensors 19 (2019) 473. https://doi.org/10.3390/s19030473

    work page doi:10.3390/s19030473 2019

  2. [2]

    Leung, T.C

    C.H. Leung, T.C. She, W.K. Lee, Positions of low-order morphology-dependent resonances determined by elastic light scattering, J. Opt. Soc. Am. B 12 (1995) 1259. https://doi.org/10.1364/josab.12.001259

    work page doi:10.1364/josab.12.001259 1995

  3. [3]

    Kaur Kohli, R.D

    R. Kaur Kohli, R.D. Davis, J.F. Davies, Tutorial: Electrodynamic balance methods for single particle levitation and the physicochemical analysis of aerosol, J. Aerosol Sci. 174 (2023) 106255. https://doi.org/10.1016/j.jaerosci.2023.106255

    work page doi:10.1016/j.jaerosci.2023.106255 2023

  4. [4]

    Knoop, N

    M. Knoop, N. Madsen, R.C. Thompson, Chapter 1: Physics with Trapped Charged Particles, in: Phys. with Trapped Charg. Part., IMPERIAL COLLEGE PRESS, 2014: pp. 1–24. https://doi.org/10.1142/9781783264063_0001

    work page doi:10.1142/9781783264063_0001 2014

  5. [5]

    M.B. Hart, V. Sivaprakasam, J.D. Eversole, L.J. Johnson, J. Czege, Optical measurements from single levitated particles using a linear electrodynamic quadrupole trap, Appl. Opt. 54 (2015) F174. https://doi.org/10.1364/ao.54.00f174

    work page doi:10.1364/ao.54.00f174 2015

  6. [6]

    W. Wu, D.H. Panduro Vela, A. Bück, A.P. Fröba, Characterization of nanofluids in evaporating droplets by dynamic light scattering, Meas. J. Int. Meas. Confed. 220 (2023). https://doi.org/10.1016/j.measurement.2023.113321

    work page doi:10.1016/j.measurement.2023.113321 2023

  7. [7]

    Goodman, Speckle Phenomena in Optics: Theory and Applications, Second Edition, 2nd ed., SPIE,

    J.W. Goodman, Speckle Phenomena in Optics: Theory and Applications, Second Edition, 2nd ed., SPIE,

  8. [8]

    https://doi.org/10.1117/3.2548484

    work page doi:10.1117/3.2548484

  9. [9]

    Bendjus, U

    B. Bendjus, U. Cikalova, J. Schreiber, Material characterization by laser speckle photometry, in: Á.F. Doval, C. Trillo (Eds.), Speckle 2012 V Int. Conf. Speckle Metrol., 2012: p. 841315. https://doi.org/10.1117/12.978246

    work page doi:10.1117/12.978246 2012

  10. [10]

    Kulcsar, D.C

    C. Kulcsar, D.C. Louie, A. Vitkin, Laser speckle size and contrast investigation of volumetric scattering from controlled turbid phantoms and mouse skin tissues, IScience 28 (2025) 112433. https://doi.org/10.1016/j.isci.2025.112433

    work page doi:10.1016/j.isci.2025.112433 2025

  11. [11]

    LeCun, Y

    Y. Lecun, Y. Bengio, G. Hinton, Deep learning, Nat. 2015 5217553 521 (2015) 436–444. https://doi.org/10.1038/nature14539

    work page doi:10.1038/nature14539 2015

  12. [12]

    K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2016: pp. 770–778

    2016

  13. [13]

    L. Liu, J. Chen, P. Fieguth, G. Zhao, R. Chellappa, M. Pietikäinen, From BoW to CNN: Two Decades of Texture Representation for Texture Classification, Int. J. Comput. Vis. 127 (2019) 74–109. https://doi.org/10.1007/s11263-018-1125-z

    work page doi:10.1007/s11263-018-1125-z 2019

  14. [14]

    Nuzzi, S

    C. Nuzzi, S. Pasinetti, I. Bassi, V. Bello, On the applicability of speckle pattern imaging combined with AI for raw milk classification, Meas. J. Int. Meas. Confed. 258 (2026) 119246. https://doi.org/10.1016/j.measurement.2025.119246

    work page doi:10.1016/j.measurement.2025.119246 2026

  15. [15]

    Jakubczyk, D

    T. Jakubczyk, D. Jakubczyk, A. Stachurski, Assessing the properties of a colloidal suspension with the aid of deep learning, J. Quant. Spectrosc. Radiat. Transf. 261 (2021) 107496. https://doi.org/10.1016/j.jqsrt.2020.107496

    work page doi:10.1016/j.jqsrt.2020.107496 2021

  16. [16]

    Nyandey, D

    K. Nyandey, D. Jakubczyk, Combining transmission speckle photography and convolutional neural network for determination of fat content in cow milk: an exercise in classification of parameters of a complex suspension, Eur. Phys. J. Plus 139 (2024) 123. https://doi.org/10.1140/epjp/s13360-024-04900- 1

    work page doi:10.1140/epjp/s13360-024-04900- 2024

  17. [17]

    Nyandey, G

    K. Nyandey, G. Derkachov, D. Jakubczyk, A convolutional neural network technique for online tracking of the radius evolution of levitating evaporating microdroplets of pure liquids, liquid mixtures and suspensions, J. Quant. Spectrosc. Radiat. Transf. (2025) 109533. https://doi.org/10.1016/j.jqsrt.2025.109533

    work page doi:10.1016/j.jqsrt.2025.109533 2025

  18. [18]

    Shopa, M

    Y. Shopa, M. Kolwas, D. Jakubczyk, G. Derkachov, I. Kamińska, K. Fronc, T. Wojciechowski, Up- Converting Luminescent Nanoparticles as Probes of Surface Dynamics in Single Evaporating Microdroplets of Suspension, J. Phys. Chem. C 129 (2025) 8955–8965. https://doi.org/10.1021/acs.jpcc.5c00174

    work page doi:10.1021/acs.jpcc.5c00174 2025

  19. [19]

    Woźniak, J

    M. Woźniak, J. Archer, T. Wojciechowski, G. Derkachov, T. Jakubczyk, K. Kolwas, M. Kolwas, D. Jakubczyk, Application of a linear electrodynamic quadrupole trap for production of nanoparticle aggregates from drying microdroplets of colloidal suspension, J. Instrum. 14 (2019). https://doi.org/10.1088/1748-0221/14/12/P12007

    work page doi:10.1088/1748-0221/14/12/p12007 2019

  20. [20]

    Shopa, M

    Y. Shopa, M. Kolwas, I. Kamińska, G. Derkachov, K. Nyandey, T. Jakubczyk, T. Wojciechowski, A. Derkachova, D. Jakubczyk, Luminescent nanoparticles in a shrinking spherical cavity – probing the evaporating microdroplets of colloidal suspension – optical lattices and structural transitions, J. Quant. Spectrosc. Radiat. Transf. 296 (2023) 108439. https://doi...

    work page doi:10.1016/j.jqsrt.2022.108439 2023

  21. [21]

    Polyanskiy, Refractiveindex.info database of optical constants, Sci

    M.N. Polyanskiy, Refractiveindex.info database of optical constants, Sci. Data 11 (2024) 94. https://doi.org/10.1038/s41597-023-02898-2

    work page doi:10.1038/s41597-023-02898-2 2024

  22. [22]

    Jakubczyk, G

    D. Jakubczyk, G. Derkachov, K. Nyandey, S. Alikhanzadeh-Arani, A. Derkachova, Chromatic dispersion and thermal coefficients of hygroscopic liquids: 5 glycols and glycerol, Sci. Data 10 (2023)

    2023

  23. [23]

    https://doi.org/10.1038/s41597-023-02819-3

    work page doi:10.1038/s41597-023-02819-3

  24. [24]

    Derkachov, D

    G. Derkachov, D. Jakubczyk, K. Kolwas, K. Piekarski, Y. Shopa, M. Woźniak, Dynamic light scattering investigation of single levitated micrometre-sized droplets containing spherical nanoparticles, Measurement 158 (2020) 107681. https://doi.org/10.1016/j.measurement.2020.107681

    work page doi:10.1016/j.measurement.2020.107681 2020

  25. [25]

    Shopa, K

    I. Shopa, K. Nyandey, D. Jakubczyk, Determination of Nano-particle and Micro-droplet Parameters in Levitating Suspension Micro-droplets by Convolutional Neural Network-based Speckle Image Analysis, Mendeley Data (2026) V1. https://doi.org/10.17632/bcrpv8zpn3.2

    work page doi:10.17632/bcrpv8zpn3.2 2026