Pith. sign in

REVIEW 4 major objections 6 minor 71 references

A machine-learned constitutive model trained only on shear flows predicts elastic instabilities in complex geometries.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 00:36 UTC pith:CADPEYOD

load-bearing objection A genuine stability fix for TBNN-based viscoelastic CFD, with honest limits; the abstract oversells the 3D generality but the core contribution is solid. the 4 major comments →

arxiv 2607.14944 v1 pith:CADPEYOD submitted 2026-07-16 physics.flu-dyn

Harnessing Machine Learning for Hybrid Constitutive Modelling of Viscoelastic Fluid Flows in Computational Rheology

J.L. Cummings , C. Fernandes , F. Dong , M.A. Alves , M.S.N. Oliveira This is my paper
classification physics.flu-dyn PACS 47.50.-a47.57.Qk47.11.-j
keywords viscoelastic flowmachine learningconstitutive modellingtensor basis neural networkuniversal differential equationscomputational rheologyelastic instabilitiesreduced integrity basis
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper claims that a neural-network correction attached to a standard viscoelastic base model, trained solely on oscillatory shear data, can be embedded in a CFD solver and produce qualitatively correct predictions in strong extensional flows—including the onset of elastic instabilities—in both 2D and 3D geometries. The key move is replacing the previously used 8- or 9-term tensor basis with a reduced 4-term basis, which the authors argue is sufficient for most common constitutive laws and drastically improves numerical stability at high Deborah numbers. If true, this means data-driven constitutive models need not be trained on every flow type to be useful in simulation, and shear-only rheometry can constrain models that extrapolate to extension-dominated flows, at least within a limited regime. The paper also shows that adding first normal stress difference information during training extends quantitative accuracy to higher Deborah numbers, and that the learned models often do not recover the true underlying equation even when their predictions are good.

Core claim

The authors establish that a Tensor-Basis Neural Network forming a Universal Differential Equation, wrapped around the upper-convected Maxwell model and trained only on small- and large-amplitude oscillatory shear data, can be deployed inside a finite-volume CFD solver and remain stable at high Deborah numbers when the tensor basis is reduced to the four terms {σ, γ̇, σ·σ, σ·γ̇ + γ̇·σ}. In this reduced formulation, the UDEs reproduce benchmark flow features—corner vortex length in 4:1 contractions and the supercritical symmetry-breaking bifurcation in cross-slot flow—with near-quantitative agreement at low De and qualitative agreement at higher De, despite the neural network not recovering t

What carries the argument

The central object is the reduced two-dimensional integrity basis for symmetric tensors, T = {σ, γ̇, σ·σ, σ·γ̇ + γ̇·σ}, used as the projection space for the TBNN output. It carries frame-invariance—any linear combination with invariant-dependent coefficients is objective—while deliberately excluding higher-order 3D tensor interactions that the authors argue are unnecessary for common constitutive models and that cause numerical instability in CFD. The UDE structure (base model plus neural correction) and the trace-based scalar invariants ρ_i as inputs complete the machinery: the network learns coefficient functions g_i(ρ) that multiply the basis tensors.

Load-bearing premise

The load-bearing premise is that the four-term two-dimensional tensor basis {σ, γ̇, σ·σ, σ·γ̇ + γ̇·σ} captures all stress–strain-rate interactions that matter in fully three-dimensional viscoelastic flows; if a real 3D flow activates higher-order tensor interactions that this basis cannot express, the learned model is structurally incomplete no matter how much data it is trained on.

What would settle it

Train the same UDE with the 4-basis formulation on the same LAOS shear data, then deploy it in a 3D flow where the third invariant tr(γ̇·γ̇·γ̇) is non-zero and significant (e.g., a 4:1 square-square contraction at De=10 or a twisted channel), and compare the stress field and stability against the ground-truth classical model. If the UDE fails well before the classical model's stability limit, the sufficiency of the 2D basis for 3D flows is falsified. Alternatively, for a real fluid, train on LAOS data, simulate a cross-slot above the measured critical Deborah number, and check whether the pred

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Shear-only LAOS measurements on a real fluid could be used to build a constitutive model for CFD simulation of complex flows, bypassing the need to select and fit a classical model form a priori.
  • The stability gain means data-driven constitutive models can be pushed to high Deborah numbers (e.g., De=20 in a planar contraction), enabling simulation of strongly elastic flows that previously were inaccessible to ML-based constitutive closures.
  • The extrapolation-region diagnostic gives practitioners a way to know where a learned model is being trusted beyond its training data, suggesting that simulation results outside that envelope should be treated as qualitative only.
  • Adding the first normal stress difference to the training loss is a cheap way to extend predictive fidelity to higher Deborah numbers and improve extensional-viscosity estimates, without needing extensional flow data.

Where Pith is reading between the lines

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

  • If the 4-term basis is indeed sufficient, then any rheological behaviour that cannot be represented by this basis in 3D lies outside the scope of this modelling approach; a testable extension is to train on planar extensional data and see whether the current shear-only UDEs' shortfall in extensional hardening overshoot at high strain rates is corrected.
  • The authors' observation that the TBNN rarely recovers the parsimonious ground-truth equation suggests that model identifiability is not what makes these UDEs useful; what matters is whether the learned coefficient functions produce the right flow-level observables, so validation should target benchmark observables (vortex length, bifurcation points) rather than coefficient recovery.
  • The instability mechanism identified—noise in the coefficient of the quartic tensor term being amplified at inlet singularities—likely applies to other data-driven constitutive models using higher-order integrity bases, so the reduced basis may become a standard design choice for CFD-coupled learning.
  • A direct extension would be to test whether the same reduced-basis UDE can extrapolate from shear training to a genuinely 3D flow with a non-trivial third invariant (e.g., a square-square contraction at higher De), since the paper's only full-3D evidence is at De=4.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper presents a hybrid Tensor Basis Neural Network (TBNN) Universal Differential Equation (UDE) framework for viscoelastic constitutive modelling, integrated into an OpenFOAM/rheoTool finite-volume solver. A reduced four-tensor integrity basis — {σ, γdot, σ·σ, σ·γdot+γdot·σ} — is introduced, motivated by 2D rheometric training data and improved numerical stability. The UDEs are trained on synthetic LAOS data from Giesekus, Johnson-Segalman, sPTT/LPTT, and ePTT models, and then deployed in start-up shear/extension, 2D planar contraction, 3D square-square contraction, and planar cross-slot flows. The central claims are that the reduced basis improves stability, enables high-Deborah-number CFD simulations, and that the trained UDEs generalise beyond their shear-only training to capture extensional-flow features and elastic instabilities, with added first-normal-stress-difference information improving fidelity.

Significance. If the claims hold, the paper is a valuable contribution to hybrid ML constitutive modelling: it provides an open-source CFD implementation, a controlled bias-injection experiment isolating a numerical-stability mechanism, and an honest analysis of model identifiability and extrapolation regimes. The Giesekus and Johnson-Segalman recoveries are clean and the vortex-length agreement with Alves et al. (within 0.4%) is strong evidence of a working pipeline. The cross-slot and contraction tests, while not perfect, demonstrate that a shear-trained UDE can qualitatively capture extensional-flow and instability features. The main risk is that the central generalization claim rests on the 2D-motivated basis being sufficient for full 3D flows, an assumption tested by only a single 3D benchmark.

major comments (4)
  1. [§3, Eqs. (6)–(10); §6.4.2] The central claim that the reduced 2D integrity basis {σ, γdot, σ·σ, σ·γdot+γdot·σ} is sufficient for full 3D constitutive modelling is not established. The argument in §3 that 'most commonly used constitutive models' do not use the higher-order tensors of Eqs. (7)–(8) is a statement about a few benchmark models, not a completeness proof. The single 3D test in §6.4.2 is a square-square contraction at De=4 for an sPTT-trained UDE; sPTT's correction term lies entirely in the 2D subspace, so the test cannot detect the absence of the missing 3D invariants (e.g., tr(γdot^3), tr(σ·γdot^2), tr(σ^2·γdot)). Consequently, the abstract's statement that the framework 'generalise[s] beyond ... training regime' is broader than the evidence. Please either add a 3D deployment with a ground-truth model whose nonlinearity involves γdot·γdot or another excluded tensor, or restrict the claim to the tested m
  2. [§6.3, §6.4.1, §7] The numerical-stability advantage is convincingly demonstrated for the Giesekus-trained UDE, where the learned network degenerates to the exact Giesekus model (g3≈α, other coefficients ≈0) and the bias-injection test pinpoints the quartic term as the instability source. However, the same stability margin is not shown for the PTT/ePTT UDEs, whose coefficients are non-constant: the ePTT UDE in §6.4.1 still diverges at De=100 and no comparison to the 8/9-basis ePTT UDE is given. The conclusion that the reduced basis 'allows simulation of flows at high Deborah numbers' should be scoped to the near-exact-recovery case; for complex learned coefficients, the stability gain is plausible but quantitatively unquantified.
  3. [§6.1.2, §6.5, Abstract] The abstract's phrase 'data-efficient discovery of ... constitutive models' overstates the PTT results. For sPTT/ePTT, the cross-component validation loss plateaus orders of magnitude above the training loss (Fig. 7b) and the learned coefficient maps do not resemble the parsimonious ground truth (Figs. 9–11). In the cross-slot, the shear-only UDE delays the bifurcation (onset between De=0.52 and 0.54 vs ground-truth ~0.508) and the N1-trained UDE predicts a small asymmetry already at De=0.50, with centreline stress 10% high/low. 'Capturing the onset and growth of elastic instabilities' is therefore only qualitatively true. Please revise the abstract and conclusion to state that the instability is reproduced qualitatively, with quantitative accuracy dependent on training information.
  4. [§3, Eqs. (5)–(6)] The replacement of the identity tensor by σ·σ in the basis deserves a caveat. From the 2D Cayley–Hamilton relation, σ·σ = tr(σ)σ − det(σ)I, so an isotropic contribution c I can only be represented by a coefficient g3 = −c/det(σ) on σ·σ. This is singular when det(σ)=0, a state that can occur in general 3D flows (e.g., at stagnation points or in purely extensional kinematics with one zero principal stress). The paper argues that det(σ)=0 is 'not problematic,' but this is only true as long as no isotropic term is needed. If the goal is a general frame-invariant constitutive model, the basis should either retain I with a quiescence penalty or the limitation must be explicit.
minor comments (6)
  1. [Eq. (3)] Equation (3) contains corrupted control characters ('⌟⟨⟨⟪rl⟫l⟩⟩...') that must be cleaned for publication.
  2. [§6.5] Typo: 'the ground truth constitutive modelare almost indistinguishable' should read 'the ground-truth constitutive model is almost indistinguishable' or similar.
  3. [Data Availability] The statement 'available upon reasonable request' is weak for a reproducibility-focused ML paper; consider releasing trained network weights, training scripts, and the UDE model class for rheoTool.
  4. [§6.3] Wording: 'the basis tensor in σ⋅σ⋅γdot⋅γdot+...' should be 'the basis tensor σ⋅σ⋅γdot⋅γdot+...'.
  5. [§3] The quiescence penalty is described as anticipated but never implemented or tested; clarify whether it is used in the final training pipeline or only motivates the choice to exclude I.
  6. [§4.2] The hyperparameters κ=0.001 and w_cyc=1e-6 are fixed without sensitivity analysis; a short robustness statement would strengthen the training methodology.

Circularity Check

0 steps flagged

No significant circularity: LAOS-trained UDEs are genuinely extrapolated to contraction/cross-slot flows; Giesekus recovery is explicitly acknowledged to be by construction and is not load-bearing.

full rationale

The paper's central claim is that a TBNN-UDE trained only on oscillatory shear data generalises to unseen flow types (extension, contraction, cross-slot) when embedded in CFD. The derivation chain is not circular: the network is fitted to transient shear stress signals (Eq. 12) and optionally N1 (Eq. 13), while the evaluation targets—extensional viscosity, corner vortex lengths, and the cross-slot symmetry-breaking bifurcation—are emergent outputs of the learned constitutive model in new kinematics, not refits of the training observables. The paper is transparent that Giesekus recovery is partly by construction: it states 'the Giesekus model is a particular case in which the TBNN can effectively recover the true constitutive model' and later 'this level of model recovery is an exception rather than the rule.' Because the basis includes sigma·sigma (Eq. 6), the Giesekus target alpha*sigma·sigma coincides with a single basis coefficient (Eq. 11); the paper explicitly treats this as an expected control, not as the central evidence. The load-bearing tests are the sPTT and ePTT UDEs, for which the network does NOT recover the parsimonious ground-truth (Sec. 6.1.2) yet still produces qualitatively and semi-quantitatively correct contraction and cross-slot behaviour—a genuine extrapolation test. Validation against external benchmarks [8] and [70] (Alves et al.; Cruz et al.) provides independent support, and no derivation step relies on a self-citation chain. The restricted 2D basis in 3D is an explicitly stated modelling assumption ('this choice restricts the model class to a subset of all 3D frame invariant constitutive relations'), and the single 3D test (Sec. 6.4.2) is acknowledged as a narrow check; this is a correctness/generality limitation, not a circular reduction of the prediction to its input. No fitted parameter is renamed as a prediction, and no uniqueness theorem is imported from the authors' own prior work. Therefore no specific circular step can be exhibited.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

No new physical entities are introduced. The four tensor basis functions are mathematical constructs from classical invariant theory, and the extrapolation diagnostic n_in is a post-processing metric rather than a new entity. The main free parameters are the trained network weights and a few hand-chosen regularisation weights. The key axiomatic load is the 2D-to-3D basis-sufficiency assumption, which is supported by only one 3D test case.

free parameters (4)
  • TBNN weights and biases = not reported (trained weights not shipped)
    The central learned object, optimised with Adam then L-BFGS on LAOS stress signals; the entire constitutive correction depends on these fitted parameters.
  • L1 regularisation weight κ = 0.001
    Chosen by hand in the total loss Eq. (14); affects sparsity and the learned model, with no sensitivity analysis reported.
  • Cyclic penalty weight w_cyc = 1e-6
    Chosen by hand in Eq. (14) to suppress long-time drift; no sensitivity analysis reported.
  • Hidden-layer architecture = 2 hidden layers × 32 neurons
    Selected for consistency with Lennon et al. [21]; no systematic architecture search is reported.
axioms (6)
  • standard math Cayley-Hamilton theorem in 2D gives a minimal basis {I, σ, γdot}.
    Invoked in Section 3, Eq. (5), as the foundation for replacing the minimal basis with the chosen four-term set.
  • domain assumption Stress and strain-rate are the only tensorial flow quantities influencing stress evolution.
    Section 3 states the integrity basis is a complete description of all interactions of stress and strain-rate; this completeness is what justifies the TBNN projection.
  • domain assumption The quartic basis term σ·σ·γdot·γdot + γdot·γdot·σ·σ is not required for a valid representation.
    The paper cites Kamrin and Govindjee [45] and uses this to justify dropping the term; this is also central to the numerical-stability claim in Section 6.3.
  • domain assumption A 2D-reduced basis remains adequate for fully 3D flows because common constitutive models rarely use higher-order tensor products.
    Stated in Section 3 and tested only in one 3D square-square contraction at De=4 in Section 6.4.2; if false, the 2D-to-3D transfer claim fails.
  • domain assumption The UCM model is an adequate base model for the UDE correction.
    Used throughout as the prior analytic term in Eq. (3); the TBNN must supply all nonlinear corrections relative to UCM, and this becomes harder for strongly nonlinear fluids.
  • standard math Neural networks are universal approximators and can represent the required correction term.
    Implicit in the UDE/TBNN framework; not proved here, but standard background for the method.

pith-pipeline@v1.3.0-alltime-deepseek · 32128 in / 12431 out tokens · 120375 ms · 2026-08-02T00:36:59.083696+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Harnessing Machine Learning for Hybrid Constitutive Modelling of Viscoelastic Fluid Flows in Computational Rheology." pith.science (2026). https://pith.science/paper/CADPEYOD

@misc{pith2026260714944,
  author       = {Pith},
  title        = {Pith review of: Harnessing Machine Learning for Hybrid Constitutive Modelling of Viscoelastic Fluid Flows in Computational Rheology},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CADPEYOD}},
  note         = {Machine review of arXiv:2607.14944}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Recent advances in data-driven modelling have highlighted the potential of hybrid approaches which combine Tensor Basis Neural Networks (TBNN) with Universal Differential Equations (UDE) to discover frame-invariant, non-linear viscoelastic constitutive models. These hybrid models enable the creation of digital twins for complex viscoelastic fluids, offering direct transferability to computational fluid dynamics simulations. In this work, we introduce a reduced dimensional tensor basis formulation that enhances both the physical consistency of the learned representations with respect to the training data and the numerical stability of subsequent simulations. The UDE architecture is embedded into an open-source finite volume solver in which the constitutive response is generated dynamically at runtime based on local fluid flow conditions. Training on synthetic datasets generated using a range of well established viscoelastic models in oscillatory shear flows alone, the performance of the resulting UDEs is evaluated under extrapolation to unseen conditions and flow-types. These include deploying the UDEs in viscometric extensional flows as well as 2D and 3D benchmark flows, such as the 4:1 sudden contraction and cross-slot, providing a quantitative analysis of their capabilities, limitations and failure modes. The proposed reduced-basis framework enables data-efficient discovery of frame-invariant constitutive models that generalise beyond their training regime, capturing key flow features such as the onset and growth of flow-induced elastic instabilities in strong extensional flows even though trained solely on shear data. Quantitative accuracy decreases as extrapolation increases, but incorporating first normal stress difference information further improves quantitative accuracy and extends predictive fidelity to higher Deborah numbers.

Figures

Figures reproduced from arXiv: 2607.14944 by C. Fernandes, F. Dong, J.L. Cummings, M.A. Alves, M.S.N. Oliveira.

Figure 1
Figure 1. Figure 1: The integration sequence which generates the UDE stress responses employed in the calculation of the loss [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: A schematic representation of the embedded TBNN in a finite-volume solver. The stress-strain-rate state [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Variation of TBNN outputs g1 to g4 with each of the scalar invariants for a UDE trained on the Giesekus fluid (α = 0.4). The points corresponding to g1, g2 overlap with g4 visible in the graphs. Similarly, the TBNN trained on a Johnson-Segalman fluid (not shown) captures the analytical form of the constitu￾tive model. In this case g4 ≈ ξ 2 and g1, g2, g3 ≈ 0 for all of the input configurations, and direct … view at source ↗
Figure 4
Figure 4. Figure 4: Histograms showing the distribution of the weights (left) and biases (right) of the neural network trained [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Variation of the 8 basis TBNN coefficient outputs [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Variation of the 9 basis TBNN coefficient outputs [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Typical evolution of the total loss used for training, [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of the stress response of the trained UDE and corresponding ground-truth fluid (sPTT [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Variation of neural network coefficient outputs [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Variation of neural network coefficient outputs [PITH_FULL_IMAGE:figures/full_fig_p016_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Variation of neural network coefficient outputs [PITH_FULL_IMAGE:figures/full_fig_p016_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Comparison of the stress responses of the trained UDEs and the corresponding ground-truth fluid (ePTT [PITH_FULL_IMAGE:figures/full_fig_p017_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Comparison between the trained UDE and the corresponding Giesekus fluid ( [PITH_FULL_IMAGE:figures/full_fig_p017_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Material properties for the trained UDEs and the corresponding ground truth, considering sPTT ( [PITH_FULL_IMAGE:figures/full_fig_p018_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Material properties for the trained UDEs and the corresponding ground truth, considering ePTT fluids [PITH_FULL_IMAGE:figures/full_fig_p018_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Comparison between the solution of a LAOS protocol at [PITH_FULL_IMAGE:figures/full_fig_p019_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Comparison between the trained UDE and the corresponding ground truth Giesekus fluid [PITH_FULL_IMAGE:figures/full_fig_p021_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Illustration of the spatial distribution of the inputs, [PITH_FULL_IMAGE:figures/full_fig_p022_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Illustration of the spatial distribution of inputs, [PITH_FULL_IMAGE:figures/full_fig_p022_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Flow patterns and contour plots of the trace of the stress tensor [PITH_FULL_IMAGE:figures/full_fig_p023_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Comparison of the variation of the normalised corner vortex length, [PITH_FULL_IMAGE:figures/full_fig_p024_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Comparison of the behaviour of the ground-truth (ePTT, [PITH_FULL_IMAGE:figures/full_fig_p025_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Isometric view of a representative computational domain of the 4:1 square-square sudden contraction [PITH_FULL_IMAGE:figures/full_fig_p026_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: Comparison of the ground-truth fluid (sPTT [PITH_FULL_IMAGE:figures/full_fig_p027_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: Comparison of the ground-truth (sPTT ϵ = 0.25, β = 1/9) and the UDE trained on shear stress only (σ12) at De = 4 in terms of the: (a) contour plot of the streamwise velocity at a central plane (y = 0); and (b) contour plot of the trace of the normalised stress tensor at a central plane (y = 0). the UDE ( [PITH_FULL_IMAGE:figures/full_fig_p027_25.png] view at source ↗
Figure 26
Figure 26. Figure 26: Streamlines and contour plots of the trace of the normalised viscoelastic extra-stress tensor in a planar [PITH_FULL_IMAGE:figures/full_fig_p028_26.png] view at source ↗
Figure 27
Figure 27. Figure 27: Variation in the magnitude of the flow asymmetry parameter, [PITH_FULL_IMAGE:figures/full_fig_p029_27.png] view at source ↗

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

71 extracted references · 4 linked inside Pith

  1. [1]

    On the formulation of rheological equations of state

    Oldroyd J. G. “On the formulation of rheological equations of state”.Proceedings of the Royal Society of London. Series A. Mathematical and Physical Sciences200 (1950), pp. 523–541

    1950

  2. [2]

    A simple constitutive equation for polymer fluids based on the concept of deformation-dependent tensorial mobility

    Giesekus H. “A simple constitutive equation for polymer fluids based on the concept of deformation-dependent tensorial mobility”.Journal of Non-Newtonian Fluid Mechanics11 (1982), pp. 69–109

    1982

  3. [3]

    A new constitutive equation derived from network theory

    Phan-Thien N., Tanner R. I. “A new constitutive equation derived from network theory”.Journal of Non- Newtonian Fluid Mechanics2 (1977), pp. 353–365

    1977

  4. [4]

    Generalized viscoelastic models: their fractional equa- tions with solutions

    Schiessel H., Metzler R., Blumen A., Nonnenmacher T. “Generalized viscoelastic models: their fractional equa- tions with solutions”.Journal of physics A: Mathematical and General28 (1995), pp. 6567–6584

    1995

  5. [5]

    Numerical simulation of non-linear elastic flows with a general collocated finite-volume method

    Oliveira P. J., Pinho F. T., Pinto G. “Numerical simulation of non-linear elastic flows with a general collocated finite-volume method”.Journal of Non-Newtonian Fluid Mechanics79 (1998), pp. 1–43

    1998

  6. [6]

    Stabilization of an open-source finite-volume solver for viscoelastic fluid flows

    Pimenta F., Alves M. “Stabilization of an open-source finite-volume solver for viscoelastic fluid flows”.Journal of Non-Newtonian Fluid Mechanics239 (2017), pp. 85–104

    2017

  7. [7]

    Popinet S.Basilisk.url:https://basilisk.fr/

  8. [8]

    Benchmark solutions for the flow of Oldroyd-B and PTT fluids in planar contractions

    Alves M. A., Oliveira P. J., Pinho F. T. “Benchmark solutions for the flow of Oldroyd-B and PTT fluids in planar contractions”.Journal of Non-Newtonian Fluid Mechanics110 (2003), pp. 45–75

    2003

  9. [9]

    nn-PINNs: Non-Newtonian physics-informed neural networks for complex fluid modeling

    Mahmoudabadbozchelou M., Karniadakis G. E., Jamali S. “nn-PINNs: Non-Newtonian physics-informed neural networks for complex fluid modeling”.Soft Matter18 (2022), pp. 172–185

    2022

  10. [10]

    RheologyNet: A physics-informed neural network solution to evaluate the thixotropic properties of cementitious materials

    Zhang T., Wang D., Lu Y. “RheologyNet: A physics-informed neural network solution to evaluate the thixotropic properties of cementitious materials”.Cement and concrete research168 (2023), p. 107157

    2023

  11. [11]

    ViscoelasticNet: A physics informed neural network framework for stress discovery and model selection

    Thakur S., Raissi M., Ardekani A. M. “ViscoelasticNet: A physics informed neural network framework for stress discovery and model selection”.Journal of Non-Newtonian Fluid Mechanics330 (2024), p. 105265

    2024

  12. [12]

    Data-driven selection of constitutive models via rheology- informed neural networks (RhINNs)

    Saadat M., Mahmoudabadbozchelou M., Jamali S. “Data-driven selection of constitutive models via rheology- informed neural networks (RhINNs)”.Rheologica Acta61 (2022), pp. 721–732

    2022

  13. [13]

    Data-driven constitutive model of complex fluids using recurrent neural networks

    Jin H., Yoon S., Park F. C., Ahn K. H. “Data-driven constitutive model of complex fluids using recurrent neural networks”.Rheologica Acta62 (2023), pp. 569–586

    2023

  14. [14]

    Recurrent neural networks (RNNs) learn the constitutive law of viscoelasticity

    Chen G. “Recurrent neural networks (RNNs) learn the constitutive law of viscoelasticity”.Computational Me- chanics67 (2021), pp. 1009–1019

    2021

  15. [15]

    One test to predict them all: Rheological characterization of complex fluids via artificial neural network

    Mishra A. A., Ghai V., Matovic V., Arlov D., K´ ad´ ar R. “One test to predict them all: Rheological characterization of complex fluids via artificial neural network”.Engineering Applications of Artificial Intelligence139 (2025), p. 109598

    2025

  16. [16]

    RheOFormer: A generative transformer model for simulation of complex fluids and flows

    Saberi M., Farimani A. B., Jamali S. “RheOFormer: A generative transformer model for simulation of complex fluids and flows”.arXiv preprint arXiv:2510.01365(2025)

    arXiv 2025

  17. [17]

    Rheo-SINDy: Finding a constitutive model from rheological data for complex fluids using sparse identification for nonlinear dynamics

    Sato T., Miyamoto S., Kato S. “Rheo-SINDy: Finding a constitutive model from rheological data for complex fluids using sparse identification for nonlinear dynamics”.Journal of Rheology69 (2025), pp. 15–34

    2025

  18. [18]

    Sparse regression for discovery of constitutive models from oscillatory shear mea- surements

    Shanbhag S., Erlebacher G. “Sparse regression for discovery of constitutive models from oscillatory shear mea- surements”.Physics of Fluids36 (2024)

    2024

  19. [19]

    Hammering at the entropy: a GENERIC- guided approach to learning polymeric rheological constitutive equations using PINNs

    Simavilla D. N., Bonfanti A., Garc ´ ıa-Beristain I., Espa˜ nol P., Ellero M. “Hammering at the entropy: a GENERIC- guided approach to learning polymeric rheological constitutive equations using PINNs”.Journal of Fluid Me- chanics1016 (2025), A11

    2025

  20. [20]

    Cfd-Nn Coupling for the Simulation of Complex Fluids

    Darves-Blanc Q., Hippolyte A., Martin S., Macqueron C., William J., Kauric G., Abibatou N., Bonnefoy O. “Cfd-Nn Coupling for the Simulation of Complex Fluids”.Available at SSRN 4748606(2024)

    2024

  21. [21]

    Scientific machine learning for modeling and simulating complex fluids

    Lennon K. R., McKinley G. H., Swan J. W. “Scientific machine learning for modeling and simulating complex fluids”.Proceedings of the National Academy of Sciences120 (2023), e2304669120. 31

    2023

  22. [22]

    Learning constitutive models and rheology from partial flow measurements

    Sunol A. M., Roggeveen J. V., Alhashim M. G., Bae H. S., Brenner M. P. “Learning constitutive models and rheology from partial flow measurements”.arXiv preprint arXiv:2510.24673(2025)

    arXiv 2025

  23. [23]

    Unbiased construction of constitutive relations for soft materials from experiments via rheology-informed neural networks

    Mahmoudabadbozchelou M., Kamani K. M., Rogers S. A., Jamali S. “Unbiased construction of constitutive relations for soft materials from experiments via rheology-informed neural networks”.Proceedings of the National Academy of Sciences121 (2024), e2313658121

    2024

  24. [24]

    Learning a family of rheological constitutive models using neural operators

    Mangal D., Saadat M., Jamali S. “Learning a family of rheological constitutive models using neural operators”. Journal of Rheology69 (2025), pp. 55–67

    2025

  25. [25]

    Data- driven physics-informed constitutive metamodeling of complex fluids: A multifidelity neural network (MFNN) framework

    Mahmoudabadbozchelou M., Caggioni M., Shahsavari S., Hartt W. H., Karniadakis G. E., Jamali S. “Data- driven physics-informed constitutive metamodeling of complex fluids: A multifidelity neural network (MFNN) framework”.Journal of Rheology65 (2021), pp. 179–198

    2021

  26. [26]

    Machine learning for viscoelastic constitutive model identification and parameterisation using large amplitude oscillatory shear

    John T. P., Mowbray M., Alalwyat A., Vousvoukis M., Martin P., Kowalski A., Fonte C. P. “Machine learning for viscoelastic constitutive model identification and parameterisation using large amplitude oscillatory shear”. Chemical Engineering Science294 (2024), p. 120075

    2024

  27. [27]

    Digital rheometer twins: Learning the hidden rheology of complex fluids through rheology-informed graph neural networks

    Mahmoudabadbozchelou M., Kamani K. M., Rogers S. A., Jamali S. “Digital rheometer twins: Learning the hidden rheology of complex fluids through rheology-informed graph neural networks”.Proceedings of the National Academy of Sciences119 (2022), e2202234119

    2022

  28. [28]

    Pinnsformer: A transformer-based framework for physics-informed neural networks

    Zhao Z., Ding X., Prakash B. A. “Pinnsformer: A transformer-based framework for physics-informed neural networks”.arXiv preprint arXiv:2307.11833(2023)

    Pith/arXiv arXiv 2023

  29. [29]

    Attention is all you need

    Vaswani A., Shazeer N., Parmar N., Uszkoreit J., Jones L., Gomez A. N., Kaiser L., Polosukhin I. “Attention is all you need”.Advances in neural information processing systems30 (2017)

    2017

  30. [30]

    Fourier neural operator for parametric partial differential equations

    Li Z., Kovachki N., Azizzadenesheli K., Liu B., Bhattacharya K., Stuart A., Anandkumar A. “Fourier neural operator for parametric partial differential equations”.arXiv preprint arXiv:2010.08895(2020)

    Pith/arXiv arXiv 2010

  31. [31]

    A composite neural network that learns from multi-fidelity data: Application to function approximation and inverse PDE problems

    Meng X., Karniadakis G. E. “A composite neural network that learns from multi-fidelity data: Application to function approximation and inverse PDE problems”.Journal of Computational Physics401 (2020), p. 109020

    2020

  32. [32]

    Multiscale simulations for viscoelastic fluids with approximate constitutive models derived by a sparse identification method

    Sato T., Miyamoto S. “Multiscale simulations for viscoelastic fluids with approximate constitutive models derived by a sparse identification method”.Rheologica Acta(2025), pp. 1–17

    2025

  33. [33]

    A survey on the application of machine learning in turbulent flow simulations

    Majchrzak M., Marciniak-Lukasiak K., Lukasiak P. “A survey on the application of machine learning in turbulent flow simulations”.Energies16 (2023), p. 1755

    2023

  34. [34]

    Reynolds averaged turbulence modelling using deep neural networks with embedded invariance

    Ling J., Kurzawski A., Templeton J. “Reynolds averaged turbulence modelling using deep neural networks with embedded invariance”.Journal of Fluid Mechanics807 (2016), pp. 155–166

    2016

  35. [35]

    Finding the underlying viscoelastic constitutive equation via universal differential equations and differentiable physics

    Rodrigues E. C., Thompson R. L., Oliveira D. A., Ausas R. F. “Finding the underlying viscoelastic constitutive equation via universal differential equations and differentiable physics”.Engineering Applications of Artificial Intelligence160 (2025), p. 111788

    2025

  36. [36]

    Discovering governing equations from data by sparse identification of nonlinear dynamical systems

    Brunton S. L., Proctor J. L., Kutz J. N. “Discovering governing equations from data by sparse identification of nonlinear dynamical systems”.Proceedings of the national academy of sciences113 (2016), pp. 3932–3937

    2016

  37. [37]

    On the high Weissenberg number problem

    Keunings R. “On the high Weissenberg number problem”.Journal of Non-Newtonian Fluid Mechanics20 (1986), pp. 209–226

    1986

  38. [38]

    The log-conformation tensor approach in the finite-volume method framework

    Afonso A., Oliveira P. J., Pinho F. T., Alves M. A. “The log-conformation tensor approach in the finite-volume method framework”.Journal of Non-Newtonian Fluid Mechanics157 (2009), pp. 55–65

    2009

  39. [39]

    Numerical methods for viscoelastic fluid flows

    Alves M. A., Oliveira P. J., Pinho F. T. “Numerical methods for viscoelastic fluid flows”.Annual Review of Fluid Mechanics53 (2021), pp. 509–541

    2021

  40. [40]

    Universal differential equations for scientific machine learning

    Rackauckas C., Ma Y., Martensen J., Warner C., Zubov K., Supekar R., Skinner D., Ramadhan A., Edelman A. “Universal differential equations for scientific machine learning”.arXiv preprint arXiv:2001.04385(2020)

    Pith/arXiv arXiv 2001

  41. [41]

    Neural ordinary differential equations

    Chen R. T. Q., Rubanova Y., Bettencourt J., Duvenaud D. K. “Neural ordinary differential equations”.Advances in neural information processing systems31 (2018)

    2018

  42. [42]

    Multilayer feedforward networks are universal approximators

    Hornik K., Stinchcombe M., White H. “Multilayer feedforward networks are universal approximators”.Neural networks2 (1989), pp. 359–366

    1989

  43. [43]

    The theory of matrix polynomials and its application to the mechanics of isotropic continua

    Spencer A. J. M., Rivlin R. S. “The theory of matrix polynomials and its application to the mechanics of isotropic continua”.Archive for rational mechanics and analysis2 (1958), pp. 309–336

    1958

  44. [44]

    On isotropic functions of symmetric tensors, skew-symmetric tensors and vectors

    Smith G. “On isotropic functions of symmetric tensors, skew-symmetric tensors and vectors”.International journal of engineering science9 (1971), pp. 899–916

    1971

  45. [45]

    Clarifying the representation of isotropic symmetric tensor-valued functions of two symmetric tensors

    Kamrin K., Govindjee S. “Clarifying the representation of isotropic symmetric tensor-valued functions of two symmetric tensors”.Mathematics and Mechanics of Solids31 (2026), pp. 172–182. 32

    2026

  46. [46]

    Enforcing Dirichlet boundary conditions in physics-informed neural networks and variational physics-informed neural networks

    Berrone S., Canuto C., Pintore M., Sukumar N. “Enforcing Dirichlet boundary conditions in physics-informed neural networks and variational physics-informed neural networks”.Heliyon9 (2023)

    2023

  47. [47]

    F.The theory of Polymer Dynamics

    Doi M., Edwards S. F.The theory of Polymer Dynamics. 1986

    1986

  48. [48]

    Pytorch: An imperative style, high-performance deep learning library

    Paszke A., Gross S., Massa F., Lerer A., Bradbury J., Chanan G., Killeen T., Lin Z., Gimelshein N., Antiga L., et al. “Pytorch: An imperative style, high-performance deep learning library”.Advances in neural information processing systems32 (2019)

    2019

  49. [49]

    Compiling machine learning programs via high-level tracing

    Frostig R., Johnson M. J., Leary C. “Compiling machine learning programs via high-level tracing”.SysML conference 2018. 2019

    2018

  50. [50]

    Jakob W.pybind11 Documentation. 2024

    2024

  51. [51]

    A general approach for running Python codes in OpenFOAM using an embedded Pybind11 Python interpreter

    Rodriguez S., Cardiff P. “A general approach for running Python codes in OpenFOAM using an embedded Pybind11 Python interpreter”.arXiv preprint arXiv:2203.16394(2022)

    arXiv 2022

  52. [52]

    Understanding the difficulty of training deep feedforward neural networks

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

    2010

  53. [53]

    Batch normalization: Accelerating deep network training by reducing internal covariate shift

    Ioffe S., Szegedy C. “Batch normalization: Accelerating deep network training by reducing internal covariate shift”.International conference on machine learning. pmlr. 2015, pp. 448–456

    2015

  54. [54]

    A review of nonlinear oscillatory shear tests: Analysis and application of large amplitude oscillatory shear (LAOS)

    Hyun K., Wilhelm M., Klein C. O., Cho K. S., Nam J. G., Ahn K. H., Lee S. J., Ewoldt R. H., McKinley G. H. “A review of nonlinear oscillatory shear tests: Analysis and application of large amplitude oscillatory shear (LAOS)”.Progress in polymer science36 (2011), pp. 1697–1753

    2011

  55. [55]

    Self-consistent Fourier–Tschebyshev rep- resentations of the first normal stress difference in large amplitude oscillatory shear

    King N., Pashkovski E., Patterson R., Rockwell P., McKinley G. H. “Self-consistent Fourier–Tschebyshev rep- resentations of the first normal stress difference in large amplitude oscillatory shear”.Journal of Rheology70 (2026), pp. 307–327

    2026

  56. [56]

    Tuor A., Drgona J., Vrabie D.Constrained neural ordinary differential equations with stability guarantees. 2020

    2020

  57. [57]

    Thummerer T., Mikelsons L.Eigen-informed neural ordinary differential equations: Incorporating knowledge of system properties. 2025

    2025

  58. [58]

    Adam: A method for stochastic optimization

    Kingma D. P. “Adam: A method for stochastic optimization”.arXiv preprint arXiv:1412.6980(2014)

    Pith/arXiv arXiv 2014

  59. [59]

    On the limited memory BFGS method for large scale optimization

    Liu D. C., Nocedal J. “On the limited memory BFGS method for large scale optimization”.Mathematical programming45 (1989), pp. 503–528

    1989

  60. [60]

    An overview of overfitting and its solutions

    Ying X. “An overview of overfitting and its solutions”.Journal of physics: Conference series. Vol. 1168. IOP Publishing. 2019, p. 022022

    2019

  61. [61]

    A nonlinear network viscoelastic model

    Phan-Thien N. “A nonlinear network viscoelastic model”.Journal of Rheology22 (1978), pp. 259–283

    1978

  62. [62]

    Working group on numerical techniques

    Hassager O. “Working group on numerical techniques”.J. of Non-Newtonian Fluid Mech.29 (1988), pp. 2–5

    1988

  63. [63]

    Dynamics of high-Deborah-number entry flows: a numerical study

    Afonso A. M., Oliveira P. J., Pinho F. T., Alves M. A. “Dynamics of high-Deborah-number entry flows: a numerical study”.Journal of Fluid Mechanics677 (2011), pp. 272–304

    2011

  64. [64]

    Numerical simulation of the planar contraction flow of a Giesekus fluid

    Choi H. C., Song J. H., Yoo J. Y. “Numerical simulation of the planar contraction flow of a Giesekus fluid”. Journal of Non-Newtonian Fluid Mechanics29 (1988), pp. 347–379

    1988

  65. [65]

    Effect of the contraction ratio upon viscoelastic fluid flow in three-dimensional square–square contractions

    Sousa P. C., Coelho P. M., Oliveira M. S. N., Alves M. A. “Effect of the contraction ratio upon viscoelastic fluid flow in three-dimensional square–square contractions”.Chemical Engineering Science66 (2011), pp. 998–1009

    2011

  66. [66]

    Extrapolation limitations of multilayer feedforward neural networks

    Haley P. J., Soloway D. “Extrapolation limitations of multilayer feedforward neural networks”.[Proceedings 1992] IJCNN international joint conference on neural networks. Vol. 4. IEEE. 1992, pp. 25–30

    1992

  67. [67]

    Large amplitude oscilla- tory extension (LAOE) of dilute polymer solutions

    Recktenwald S. M., John T. P., Shen A. Q., Poole R. J., Fonte C. P., Haward S. J. “Large amplitude oscilla- tory extension (LAOE) of dilute polymer solutions”.Journal of Non-Newtonian Fluid Mechanics342 (2025), p. 105421

    2025

  68. [68]

    Optimized cross-slot flow geometry for microfluidic extensional rheometry

    Haward S. J., Oliveira M. S. N., Alves M. A., McKinley G. H. “Optimized cross-slot flow geometry for microfluidic extensional rheometry”.Physical Review Letters109 (2012), p. 128301

    2012

  69. [69]

    Purely elastic flow asymmetries

    Poole R. J., Alves M. A., Oliveira P. J. “Purely elastic flow asymmetries”.Physical review letters99 (2007), p. 164503

    2007

  70. [70]

    A new viscoelastic benchmark flow: Stationary bifurcation in a cross-slot

    Cruz F. A., Poole R. J., Afonso A. M., Pinho F. T., Oliveira P. J., Alves M. A. “A new viscoelastic benchmark flow: Stationary bifurcation in a cross-slot”.Journal of Non-Newtonian Fluid Mechanics214 (2014), pp. 57–68

    2014

  71. [71]

    On extensibility effects in the cross-slot flow bifurcation

    Rocha G. N., Poole R. J., Alves M. A., Oliveira P. J. “On extensibility effects in the cross-slot flow bifurcation”. Journal of Non-Newtonian Fluid Mechanics156 (2009), pp. 58–69. 33

    2009