pith. sign in

arxiv: 2605.08109 · v1 · submitted 2026-04-27 · 💻 cs.LG · cond-mat.mtrl-sci· physics.flu-dyn

Geometry-free prediction of inertial lift forces in microfluidic devices using deep learning

Jesse Ward-Bond , Ali Mashadian , Timothy C. Y. Chan , Edmond W. K. Young This is my paper

Pith reviewed 2026-05-12 01:44 UTC · model grok-4.3

classification 💻 cs.LG cond-mat.mtrl-sciphysics.flu-dyn
keywords inertial microfluidicslift force predictiondeep learningneural networksparticle migrationmicrofluidic devicesflow simulationgeneralization
0
0 comments X

The pith

A neural network predicts inertial lift forces on particles in microfluidic channels without any explicit geometric inputs.

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

The authors train a neural network to forecast how particles experience lift forces inside microfluidic channels. They replace direct descriptions of channel shape with a new set of parameters that describe the surrounding flow. The resulting model matches the accuracy of earlier geometry-specific networks on familiar channel types while performing substantially better on cross-sections never seen in training. This removes the need to retrain a separate model for every new device design and allows the lift-force predictions to plug directly into particle-tracing simulators.

Core claim

We develop a novel approach for predicting particle lift forces that contains no explicit geometric parameters. We train a neural network model using a new parameter set and show that while it performs comparably to existing models on channel geometries in the training set, it is able to generalize to unseen channel geometries far more effectively. We show that the lift force model developed herein can be easily transferred to particle tracing simulation software, where it is capable of predicting particle migration patterns consistent with the literature across a variety of channel designs.

What carries the argument

A neural network whose inputs are a geometry-free parameter set that encodes flow-field information sufficient to compute lift forces.

If this is right

  • A single model can serve rectangular, triangular, and other channel types without retraining for each shape.
  • Lift-force predictions integrate directly into existing particle-tracing codes to produce migration patterns.
  • Simulated particle paths remain consistent with published experimental and numerical results for multiple designs.
  • The computational cost of exploring new channel geometries drops because no additional model training is required.

Where Pith is reading between the lines

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

  • Designers could iterate channel shapes in simulation without repeatedly training new networks for each candidate geometry.
  • The same parameter encoding might be tested on three-dimensional or time-varying flows to see whether generalization holds beyond two-dimensional steady cases.
  • If the model remains accurate on experimental data, it could support closed-loop optimization in which desired particle paths guide automatic channel design.

Load-bearing premise

The new parameter set supplies enough information about the flow field for the network to predict lift forces accurately on channel cross-sections never encountered during training.

What would settle it

Run the trained model on a channel cross-section absent from the training data, such as a circular or trapezoidal shape, and compare its lift-force values and resulting particle trajectories against independent high-fidelity numerical simulations or laboratory measurements.

Figures

Figures reproduced from arXiv: 2605.08109 by Ali Mashadian, Edmond W. K. Young, Jesse Ward-Bond, Timothy C. Y. Chan.

Figure 1
Figure 1. Figure 1: Overview of development and downstream integration of a geometry-free lift force model. (a) Direct numerical [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The transitive relationship between the undisturbed velocity profile and the lift force [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Summary of raw datasets used for model training and testing. From left to right: rectangular ( [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Model architecture. Input nodes are on the left, output nodes on the right. Hidden layers are represented in [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison between our model and the model developed in Su [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Model performance on seen and unseen geometries as new geometries are added to the training data. Models [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of DNS- and FCNN-predicted CLx and CLy profiles in straight channels with different cross￾sections. The model was trained only on rectangular cross sections. Channel dimensions are not to scale. 9 [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Particle trajectories from COMSOL simulations employing our lift-force model trained on straight channels with rectangular, semicircular, triangular, and trapezoidal cross-sections. (a) A straight square channel containing 10 µm (blue), 15 µm (green), and 20 µm (red) particles. (b) A straight triangular channel containing 10 µm particles. (c) Twenty-five chambers of an expansion-contraction channel contain… view at source ↗
Figure 9
Figure 9. Figure 9: Shapley values for CLx (top) and CLy (bottom), calculated using the SHAP library for python on 500 samples from the rectangular test data Rts . are from the symmetry of the velocity profile. At the intermediate distance, they are due to a balance between shear gradient lift forces and wall-induced lift and it is at these lateral equilibrium positions where particles ultimately focus during inertial migrati… view at source ↗
Figure 10
Figure 10. Figure 10: (a) Predicted lift coefficient vector components plotted against actual lift coefficient components. (b) [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗
read the original abstract

Inertial microfluidic devices (IMDs) offer low-cost, high-throughput alternative techniques for many traditional particle- (or cell-) manipulation tasks, but simulating them requires being able to predict particle migration, and thus particle lift forces, under a variety of possible channel geometries. Recent work has demonstrated that machine learning models can be used to drastically speed up these numerical simulations, but doing so required training individual models for every unique channel cross-section type (e.g., rectangular, triangular) -- shifting the burden from the simulation step to the training step. In this paper, we develop a novel approach for predicting particle lift forces that contains no explicit geometric parameters. We train a neural network model using a new parameter set and show that while it performs comparably to existing models on channel geometries in the training set, it is able to generalize to unseen channel geometries far more effectively. We show that the lift force model developed herein can be easily transferred to particle tracing simulation software, where it is capable of predicting particle migration patterns consistent with the literature across a variety of channel designs.

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 claims to introduce a deep learning model for predicting inertial lift forces on particles in microfluidic channels that uses a novel parameter set containing no explicit geometric descriptors. Trained on simulation data, the model achieves performance comparable to prior geometry-specific networks on in-distribution channel cross-sections (e.g., rectangular and triangular) while generalizing more effectively to unseen geometries; the resulting lift-force predictor is shown to integrate directly into particle-tracing software and reproduce literature-consistent migration patterns across multiple channel designs.

Significance. If the generalization result holds, the work would meaningfully reduce the per-geometry retraining burden that currently limits ML-accelerated inertial microfluidics simulations, enabling faster exploration of arbitrary cross-sections. The demonstrated transfer to existing particle-tracing codes is a practical strength that could accelerate device prototyping.

major comments (2)
  1. [Abstract and §3] Abstract and §3 (method): the central claim that the new parameter set enables superior generalization to unseen geometries is not supported by any quantitative metrics, error bars, or explicit description of the parameter definition, training/test split across cross-section types, or out-of-distribution test geometries. Without these details it is impossible to assess whether the reported improvement reflects true flow-invariant encoding or residual distribution leakage.
  2. [Results] Results section: the assertion of 'comparable in-distribution performance and superior out-of-distribution behavior' is stated without tabulated error values, confidence intervals, or direct comparison tables against the geometry-specific baselines on the held-out shapes, leaving the load-bearing generalization claim unverified.
minor comments (2)
  1. Figure captions and legends should explicitly distinguish in-distribution versus out-of-distribution test cases and include error bars or shaded regions for all reported lift-force predictions.
  2. The manuscript should add a short reproducibility subsection listing the exact composition of the training set (number of rectangular, triangular, and other cross-sections) and the precise mathematical definition of the new parameter vector.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive feedback, which identifies areas where additional quantitative detail will strengthen the presentation of our generalization results. We address each major comment below and will revise the manuscript to incorporate the requested metrics, definitions, and tables.

read point-by-point responses

  1. Referee: [Abstract and §3] Abstract and §3 (method): the central claim that the new parameter set enables superior generalization to unseen geometries is not supported by any quantitative metrics, error bars, or explicit description of the parameter definition, training/test split across cross-section types, or out-of-distribution test geometries. Without these details it is impossible to assess whether the reported improvement reflects true flow-invariant encoding or residual distribution leakage.

    Authors: We agree that the current manuscript does not provide sufficient explicit quantitative support or detailed descriptions in the abstract and §3. In the revision we will: (1) add a precise mathematical definition of the novel parameter set in §3, (2) specify the exact training/test splits with the cross-section types assigned to each, (3) list the out-of-distribution test geometries, and (4) report quantitative metrics (e.g., mean absolute error with standard-error bars) comparing in-distribution and out-of-distribution performance. These additions will allow readers to evaluate the nature of the observed generalization. revision: yes

  2. Referee: [Results] Results section: the assertion of 'comparable in-distribution performance and superior out-of-distribution behavior' is stated without tabulated error values, confidence intervals, or direct comparison tables against the geometry-specific baselines on the held-out shapes, leaving the load-bearing generalization claim unverified.

    Authors: We acknowledge that the results section would be clearer with explicit tabulated comparisons. In the revised manuscript we will insert tables that report error metrics (with confidence intervals) for the geometry-free model versus the geometry-specific baselines on both in-distribution and held-out channel shapes. This will directly substantiate the claims of comparable in-distribution accuracy and improved out-of-distribution generalization. revision: yes

Circularity Check

0 steps flagged

No circularity in ML-based lift force prediction

full rationale

The paper trains a neural network on simulation data using a new (undisclosed) parameter set to predict inertial lift forces without explicit geometric inputs. The generalization claim to unseen channel cross-sections is presented as an empirical result of supervised learning and held-out testing, not as a first-principles derivation. No equations, self-citations, or ansatzes are shown that reduce the output to a redefinition or fit of the inputs. The approach remains falsifiable against independent particle-tracing simulations and does not exhibit any of the enumerated circular patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on the assumption that a fixed set of non-geometric parameters can fully substitute for channel shape in lift-force prediction. No explicit free parameters, axioms, or invented entities are named in the abstract.

pith-pipeline@v0.9.0 · 5504 in / 1096 out tokens · 20957 ms · 2026-05-12T01:44:00.050100+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

47 extracted references · 47 canonical work pages

  1. [1]

    Zhang, S

    J. Zhang, S. Yan, D. Yuan, G. Alici, N.-T. Nguyen, M. E. Warkiani and W. Li,Lab Chip, 2016,16, 10–34

    work page 2016

  2. [2]

    Xiang and Z

    N. Xiang and Z. Ni,Lab Chip, 2022,22, 4792–4804

    work page 2022

  3. [3]

    M. G. Lee, S. Choi and J.-K. Park,J. Chromatogr. A, 2011,1218, 4138–4143

    work page 2011

  4. [4]

    Zhou and I

    J. Zhou and I. Papautsky,Lab Chip, 2013,13, 1121–1132

    work page 2013

  5. [5]

    S. C. Hur, T. Z. Brinckerhoff, C. M. Walthers, J. C. Dunn and D. Di Carlo,PLOS ONE, 2012, 46550

    work page 2012

  6. [6]

    D. H. Yoon, J. B. Ha, Y . K. Bahk, T. Arakawa, S. Shoji and J. S. Go,Lab Chip, 2009,9, 87–90

    work page 2009

  7. [7]

    G. Guan, L. Wu, A. A. Bhagat, Z. Li, P. C. Chen, S. Chao, C. J. Ong and J. Han,Sci. Rep., 2013,3, 1475

    work page 2013

  8. [8]

    Park, S.-H

    J.-S. Park, S.-H. Song and H.-I. Jung,Lab Chip, 2009,9, 939–948

    work page 2009

  9. [9]

    Zhang, S

    J. Zhang, S. Yan, R. Sluyter, W. Li, G. Alici and N.-T. Nguyen,Sci. Rep., 2014,4, 4527

    work page 2014

  10. [10]

    S. R. Bazaz, A. Mashhadian, A. Ehsani, S. C. Saha, T. Krüger and M. E. Warkiani,Lab Chip, 2020,20, 1023–1048

    work page 2020

  11. [11]

    Jiang and N

    F. Jiang and N. Xiang,Anal. Chem., 2022,94, 1859–1866

    work page 2022

  12. [12]

    H. Jeon, T. Kwon, J. Yoon and J. Han,Lab Chip, 2022,22, 272–285

    work page 2022

  13. [13]

    V olpe, P

    A. V olpe, P. Paiè, A. Ancona and R. Osellame,Microfluid. Nanofluid., 2019,23, 37

    work page 2019

  14. [14]

    S. K. Sia and G. M. Whitesides,Electrophoresis, 2003,24, 3563–3576

    work page 2003

  15. [15]

    D. J. Guckenberger, T. E. De Groot, A. M. Wan, D. J. Beebe and E. W. Young,Lab Chip, 2015,15, 2364–2378. 12 APREPRINT- MAY12, 2026

    work page 2015

  16. [16]

    Hwang, Y

    J. Hwang, Y . H. Cho, M. S. Park and B. H. Kim,International Journal of Precision Engineering and Manufacturing, 2019,20, 479–495

    work page 2019

  17. [17]

    Grimes, D

    A. Grimes, D. N. Breslauer, M. Long, J. Pegan, L. P. Lee and M. Khine,Lab Chip, 2008,8, 170–172

    work page 2008

  18. [18]

    W. Tang, S. Zhu, D. Jiang, L. Zhu, J. Yang and N. Xiang,Lab Chip, 2020,20, 3485–3502

    work page 2020

  19. [19]

    Segré and A

    G. Segré and A. Silberberg,Nature, 1961,189, 209–210

    work page 1961

  20. [20]

    Di Carlo, D

    D. Di Carlo, D. Irimia, R. G. Tompkins and M. Toner,Proc. Natl. Acad. Sci. U. S. A., 2007,104, 18892–18897

    work page 2007

  21. [21]

    K. Hood, S. Lee and M. Roper,J. Fluid Mech., 2015,765, 452–479

    work page 2015

  22. [22]

    E. S. Asmolov, A. L. Dubov, T. V . Nizkaya, J. Harting and O. I. Vinogradova,J. Fluid Mech., 2018,840, 613–630

    work page 2018

  23. [23]

    J. Su, X. Zheng and G. Hu,Phys. Fluids, 2023,35, 092010

    work page 2023

  24. [24]

    Di Carlo, J

    D. Di Carlo, J. F. Edd, K. J. Humphry, H. A. Stone and M. Toner,Phys. Rev. Lett., 2009,102, 094503

    work page 2009

  25. [25]

    G. Li, G. H. McKinley and A. M. Ardekani,J. Fluid Mech., 2015,785, 486–505

    work page 2015

  26. [26]

    C. Liu, G. Hu, X. Jiang and J. Sun,Lab Chip, 2015,15, 1168–1177

    work page 2015

  27. [27]

    Shamloo and A

    A. Shamloo and A. Mashhadian,Phys. Fluids, 2018,30, 012002

    work page 2018

  28. [28]

    McIntyre, A

    D. McIntyre, A. Lashkaripour, P. Fordyce and D. Densmore,Lab Chip, 2022,22, 2925–2937

    work page 2022

  29. [29]

    Owen,Front

    B. Owen,Front. Lab Chip Technol., 2024,3, 1328004

    work page 2024

  30. [30]

    J. Su, X. Chen, Y . Zhu and G. Hu,Lab Chip, 2021,21, 2544–2556

    work page 2021

  31. [31]

    Hamdi, R

    E. Hamdi, R. Dezhkam, A. Shamloo and A. Mashhadian,microAI: A machine learning tool for fast calculation of lift coefficients in microchannels, 2022,https://arxiv.org/abs/2210.11591

    work page arXiv 2022

  32. [32]

    Safari, P

    M. Safari, P. Abbasi, S. Momeni, M. S. Farahani and H. Safari,Chem. Eng. Sci., 2024,297, 120301

    work page 2024

  33. [33]

    Akhbari and M

    M. Akhbari and M. Nasr Esfahany,Indian Chem. Eng., 2024, 1–17

    work page 2024

  34. [34]

    Buckingham,Phys

    E. Buckingham,Phys. Rev., 1914,4, 345

    work page 1914

  35. [35]

    Nesterov,Proc

    Y . Nesterov,Proc. USSR Acad. Sci., 1983,269, 3

    work page 1983

  36. [36]

    Sutskever, J

    I. Sutskever, J. Martens, G. Dahl and G. Hinton, Proceedings of the 30th International Conference on Machine Learning, Atlanta, Georgia, USA, 2013, pp. 1139–1147

    work page 2013

  37. [37]

    Godbole, G

    V . Godbole, G. E. Dahl, J. Gilmer, C. J. Shallue and Z. Nado,Deep Learning Tuning Playbook, 2023, http: //github.com/google-research/tuning_playbook, Version 1.0

    work page 2023

  38. [38]

    S. M. Lundberg and S.-I. Lee, Proceedings of the 31st International Conference on Neural Information Processing Systems, Red Hook, NY , USA, 2017, p. 4768–4777

    work page 2017

  39. [39]

    J. Bai, F. Lu, K. Zhanget al.,ONNX: Open Neural Network Exchange, https://github.com/onnx/onnx, 2019

    work page 2019

  40. [40]

    Y . Zhou, Z. Ma and Y . Ai,Microsyst. Nanoeng., 2018,4, 5

    work page 2018

  41. [41]

    J. S. Dai,Mech. Mach. Theory, 2015,92, 144–152

    work page 2015

  42. [42]

    Z. Wu, Y . Chen, M. Wang and A. J. Chung,Lab Chip, 2016,16, 532–542

    work page 2016

  43. [43]

    Nivedita, P

    N. Nivedita, P. Ligrani and I. Papautsky,Sci. Rep., 2017,7, 44072

    work page 2017

  44. [44]

    Botchkarev,Interdisciplinary Journal of Information, Knowledge, and Management, 2019,14, 045–076

    A. Botchkarev,Interdisciplinary Journal of Information, Knowledge, and Management, 2019,14, 045–076

    work page 2019

  45. [45]

    Moloudi, S

    R. Moloudi, S. Oh, C. Yang, M. Ebrahimi Warkiani and M. W. Naing,Microfluid. Nanofluid., 2018,22, 1–14

    work page 2018

  46. [46]

    Kim, J.-Y

    U. Kim, J.-Y . Kwon, T. Kim and Y . Cho,Micromachines, 2022,13, 151

    work page 2022

  47. [47]

    imaginary

    M. E. Warkiani, B. L. Khoo, L. Wu, A. K. P. Tay, A. A. S. Bhagat, J. Han and C. T. Lim,Nat. Protoc., 2016,11, 134–148. 13 APREPRINT- MAY12, 2026 Supplementary Information S1 Taylor Series Approximation Figure S1 shows that a second-order Taylor series expansion can almost completely recover the velocity profile around the circumference of an “imaginary” (...

    work page 2016