Pith. sign in

REVIEW 4 major objections 5 minor 38 references

Hybrid Adaptive Modeling in Process Monitoring: Leveraging Sequence Encoders and Physics-Informed Neural Networks

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper introduces a physics-informed neural network that uses Deep Sets to encode sensor data, allowing one model to adapt to new parameters and boundary conditions without retraining, with tests on a chaotic ODE, flow past a cylinder, and real composite plate heating.

desk verdict A plausible and useful architecture that combines Deep Sets with PINNs, with a strong real-data test, but the headline claim about identifying hidden BC parameters from sparse pressure sensors is undersupported because the identification blocks are supervised and identifiability is never analyzed. read the letter →

arxiv 2505.14252 v2 pith:I4LS52K5 submitted 2025-05-20 cs.LG cs.AI

Mouad Elaarabi , Domenico Borzacchiello , Philippe Le Bot , Nathan Lauzeral , Sebastien Comas-Cardona This is my paper
classification cs.LGcs.AI
keywords modelconditionsparametersboundaryinitialsequencevelocityapproach
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Physics-informed neural networks (PINNs) are neural networks that learn to solve equations like differential equations by being penalized when they violate the known physics. They usually work for one fixed set of conditions: if the material property, the boundary condition, or the initial state changes, the network has to be trained again from scratch. This paper attaches a sequence encoder, a Deep Sets network, at the front of the PINN. The encoder watches a short time series of sensor measurements, for example temperatures at two surfaces of a plate or pressures at a few points around a cylinder, and condenses them into a small set of numbers. Those numbers are then fed into the PINN along with the coordinates where the prediction is wanted. Because the encoder summarizes the current state of the system, the same PINN can be used for many different parameter values and boundary conditions, without retraining.

The authors test the idea on three problems. The first is the Rössler ODE, which can behave chaotically. The model is given short, noisy segments of the three state variables and must predict the unknown parameter and continue the trajectory into the future. The second is 2D flow around a cylinder. The model sees pressure readings from 12 sensors and must deduce the amplitude and frequency of the sinusoidal inlet velocity, then use the Navier-Stokes equations to reconstruct velocity and pressure everywhere. The third is a 1D heat conduction problem using real data from an infrared oven heating glass-fiber and thermoplastic plates. The model is trained only on simulation data and then tested on real measurements, including cases where the bottom heater is switched off halfway through the process.

The reported results are mostly good, but there are caveats.

Extended reading notes

Core claim

From the abstract: the architecture 'employs Deep Sets (Sequence Encoders) to encode dynamic parameters, boundary conditions, and initial conditions, using these encoded features as inputs for the PINN, enabling the model to adapt to changes in parameters, BCs, and ICs.' The strongest load-bearing assertion, stated in Section 5.3.3, is that a single trained model 'successfully identified the boundary condition's parameters and use them to predict velocity, even when they are not explicitly provided as input,' i.e., the model can, without retraining, identify hidden problem parameters from a short sensor stream and use the physics loss to extrapolate the solution over the whole domain. If true, the paper demonstrates a parametric operator learning method: one PINN solves a family of ODE/PDE problems across a range of parameters and BCs/ICs.

Load-bearing premise

The load-bearing assumption is identifiability and sensor sufficiency: the sensor time series supplied to the Deep Set encoder must contain enough information to uniquely determine the hidden parameters and boundary conditions. For the Navier-Stokes case, 12 pressure point readings are assumed to determine (Umax, f) of the inlet velocity (Eq. 19); for the heat case, two surface temperatures plus the heater power percentages and the thickness must determine the heat flux and internal state (Section 5.4.5). The paper selects sensor locations (Fig. 21) and uses synthetic training data with known parameters, but it never proves or tests the non-identifiability boundary, e.g., pressure sensors placed in a low-sensitivity region, or pyrometer readings corrupted by emissivity variation between PP plates of different surface quality (Section 5.4.1).

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes PINN-SE, a hybrid architecture that combines sequence encoders (Deep Sets) with physics-informed neural networks to solve parametric ODE/PDE problems without retraining for each parameter instance. The encoder maps short sensor time series, together with known boundary/initial condition information, to a feature vector that conditions a PINN; additional identification blocks predict dynamic parameters and enable cross-validation between the identification and PINN blocks. The method is demonstrated on three problems: the Rössler ODE with variable initial conditions and parameter c, a 2D incompressible Navier–Stokes flow past a cylinder with a parametric sinusoidal inlet velocity, and a 1D heat conduction problem in an infrared oven using both synthetic and real composite-plate data. The paper reports accurate parameter regression and field predictions, including extrapolation to times beyond the encoder window, and a real-data transfer test.

Significance. If the claims hold, PINN-SE is a useful parametric operator-learning framework: a single trained model can accommodate variable parameters, boundary conditions, initial conditions, variable sampling frequencies, and noisy data, while the multi-block design offers a practical cross-validation mechanism. The paper has concrete strengths: it validates on real experimental data in the heat application, it explicitly handles variable-length and variable-frequency sequences, it studies robustness to 10% noise, and it reports quantitative errors over entire domains. The main limitations are that the identification blocks are trained with supervised labels of the true parameters and that no identifiability or sensitivity analysis is provided for the sensor-to-parameter map; as a result, the strongest claim of hidden-parameter identification is not fully established. The absence of baseline comparisons also makes it difficult to calibrate the reported accuracy against existing operator-learning or PINN approaches.

major comments (4)
  1. [Section 5.2.2, Eq. (15)] The Rössler physics loss LODE uses the true parameter c in the third equation, not the predicted parameter ec produced by the identification block. The PINN is therefore trained with ground-truth dynamics, and the experiment does not demonstrate that the physics loss can operate on an inferred parameter. The cross-validation study in Section 5.2.3 uses ec only in a diagnostic residual, not in the training objective. To support the claim that the model adapts to changing parameters through identified parameters, the authors should either train and evaluate with the predicted parameter in the ODE residual or explicitly state that the Rössler experiment validates parameter-conditioned PINN regression rather than identified-parameter adaptation.
  2. [Section 5.3.1, Eq. (22), and Section 5.3.3] The Navier–Stokes identification block is trained by supervised regression on the true parameters Umax and f (Eq. 22, LP), so the high R2 values in Fig. 23 certify interpolation on the sampled training distribution. They do not establish that the 12 pressure sensors determine (Umax, f) uniquely and in a well-conditioned manner over the claimed parameter domain. For large f, U(t) in Eq. (19) is nearly constant over the input window t in [0, 0.6T], making f nearly unobservable from pressure signals and creating near-degeneracy with Umax; the unexplained red-cross cases in Fig. 24 are consistent with such a non-identifiability region. The authors should add an identifiability/sensitivity analysis, or demonstrate correct inference in the low-sensitivity part of the domain, before claiming that the model 'successfully identified the boundary condition's parameters' in general.
  3. [Section 5.3 and Section 5.4] No baseline comparisons are reported. The paper does not compare against a retrained standard PINN, a pure Deep Set regression model, DeepONet/DeepOSets, or PINN-SR. Without such comparisons, the reported error levels in Tables 3, 7, and 9 are difficult to interpret, especially for a paper whose contribution is an architectural combination. At minimum, a comparison against a supervised Deep Set baseline and a per-instance retrained PINN would clarify whether the physics block and the cross-validation scheme add value beyond the encoder.
  4. [Section 5.4.5, Eq. (27), and Section 5.4.6] In the heat application, the Deep Set input explicitly includes the measured surface temperatures T0 and TL, while the supervised loss LSUP trains the model to reproduce these same temperatures at x=0 and x=L. Surface accuracy is therefore partly a copy or regression task, and the internal extrapolation is only tested on synthetic data; for real data, Table 9 reports only x=0 and x=L. To support the claim that the physics loss extrapolates inside the plate, the authors should provide an ablation without LPDE on the synthetic internal-point test, and ideally an indirect real-data check of internal temperatures, or state clearly that the real-data claim is limited to surface monitoring.
minor comments (5)
  1. [Eq. (15)] The parameter loss LP is written as ||P||_2^2, which does not indicate a comparison with the true parameter c; if this is a typographical artifact, the intended supervised parameter loss should be written out explicitly.
  2. [Algorithm 1 and Section 5.3.2] Several training thresholds are left unnamed ('threshold', 'threshold2'), and the text states both 'λ = 10' and 'at the beginning λ = 0' for the Navier–Stokes loss; these details should be reconciled to make the training procedure reproducible.
  3. [Section 5.3.1, Eq. (21)] The notation 'i ∈ [2, 11]' for the 12 pressure sensors is confusing; the index range and the sensor-location grouping should be clarified.
  4. [Section 5.3.3, Fig. 24] The red-cross cases where cross-validation fails are mentioned but not analyzed; a short discussion of these outliers would help readers judge whether they correspond to identifiable parameter regions.
  5. [Section 5.4.2] The paper states that pyrometer readings are susceptible to surface-quality and emissivity variations, but this sensitivity is not quantified; a brief analysis of how measurement noise propagates to the heat-flux inference would strengthen the real-data claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the parameter-identification blocks are supervised on true labels and tested on held-out data; the physics loss is a training signal, not a self-referential prediction.

full rationale

The paper's derivation chain is a supervised architecture: sequence encoders map sensor time series to feature vectors, which are then used by PINN blocks and identification blocks. The identification targets are training labels, not quantities derived from the same fitted values. In the Rössler case, LODE uses the true parameter c (Eq. 15), but only as a physics training signal for the PINN conditioned on the encoder feature vector; at test time c is not provided. In the Navier-Stokes case, LP = ||Umax − eUmax||2 + ||f − ef||2 (Eq. 22) is a supervised regression loss on the true inlet parameters, and the reported R2 > 0.9 is evaluated on held-out test datasets. This is standard supervised operator learning, not a case where the prediction reduces to the input by construction. The cross-validation between the identification block and the PINN is explicitly framed as an internal consistency check that provides 'some level of confidence,' not as an external first-principles derivation. Self-citations [18] and [37] support prior architectural choices and adaptive refinement, but they are not the load-bearing justification for the central results. The identifiability and sensor-sufficiency concerns (e.g., possible near-degeneracy of (Umax, f) for large f in Eq. 19, variable surface emissivity in Sec. 5.4.1) are legitimate validation and generalization issues, but they are not circularity: no equation in the paper makes the claimed prediction equivalent to its own fitted inputs. Therefore no circular step can be exhibited, and the score is 0.

Assumptions & free parameters 8 free parameters · 6 assumptions · 0 invented entities

The method relies on known physics, sensor sufficiency, and a set of hand-tuned hyperparameters and training schedules. There are no newly postulated physical entities; all components are established ML blocks. The main load-bearing assumptions are problem identifiability and simulation-to-real transfer.

free parameters (8)
  • Loss weights lambda and beta for Rossler (Eq. 15) = Scheduled from 0/1 with hand-tuned thresholds (Algo. 1)
    Chosen by hand on validation data; not derived.
  • Loss weights alpha, lambda, eta for Navier-Stokes (Eq. 22) = alpha=10, lambda scheduled from 0, eta=0.1
    Hand-set; lambda is increased when supervised loss drops below a threshold.
  • Loss weights beta1-beta4 for heat (Eq. 27) = beta1=0, beta2=5, beta3=1, beta4=10 initially
    Hand-set; beta1 gradually increased during training.
  • Deep Set time step Delta_TDS for Rossler = 0.5 s
    Chosen for training; affects input resolution.
  • Number of snapshots NDS for Navier-Stokes = 120 snapshots (t in [0, 3.6])
    Fixed window length; chosen by tuning.
  • Aggregation functions for Deep Sets = L2 norm (Rossler), sum (NS), sum of abs (heat)
    Selected by validation performance.
  • Network widths, depths, activations = Tables 1, 2, 6
    Manually tuned; not guaranteed optimal.
  • Learning rate schedule thresholds = lr from 6e-3 with per-epoch decay; thresholds unnamed
    Hand-tuned; algorithm describes thresholds without values.
assumptions (6)
  • domain assumption The governing PDE/ODE forms are exact (Rossler Eq. 9, NS Eq. 10, heat Eq. 11).
    The physics loss is built on these equations; if the governing physics is mis-specified, the constraints are wrong.
  • domain assumption Sensor time series contain enough information to identify the hidden parameters (identifiability).
    The encoder must produce a sufficient statistic for the parameters from the available sensor data; this is not proven.
  • domain assumption Training on synthetic data with additive Gaussian noise transfers to real sensor data with domain shift.
    The heat application relies on this transfer; surface quality and emissivity variations are not modeled in the synthetic data.
  • standard math Neural networks can approximate the solution operator (universal approximation).
    The PINN and encoder are assumed able to represent the solution operator; no convergence guarantee is given.
  • ad hoc to paper The 1D heat simplification with h=8 W/(m^2 K) and the chosen parameter ranges are representative.
    The model is trained on these ranges; if real plates fall outside, the encoder must extrapolate.
  • ad hoc to paper The hand-crafted training schedule in Algo. 1 with unnamed thresholds is necessary for convergence.
    No theoretical basis; the model's success depends on the specific schedule.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hybrid Adaptive Modeling in Process Monitoring: Leveraging Sequence Encoders and Physics-Informed Neural Networks." pith.science (2026). https://pith.science/paper/I4LS52K5

@misc{pith2026250514252,
  author       = {Pith},
  title        = {Pith review of: Hybrid Adaptive Modeling in Process Monitoring: Leveraging Sequence Encoders and Physics-Informed Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I4LS52K5}},
  note         = {Machine review of arXiv:2505.14252}
}
read the original abstract

In this work, we explore the integration of Sequence Encoding for Online Parameter Identification with Physics-Informed Neural Networks to create a model that, once trained, can be utilized for real time applications with variable parameters, boundary conditions, and initial conditions. Recently, the combination of PINNs with Sparse Regression has emerged as a method for performing dynamical system identification through supervised learning and sparse regression optimization, while also solving the dynamics using PINNs. However, this approach can be limited by variations in parameters or boundary and initial conditions, requiring retraining of the model whenever changes occur. In this work, we introduce an architecture that employs Deep Sets or Sequence Encoders to encode dynamic parameters, boundary conditions, and initial conditions, using these encoded features as inputs for the PINN, enabling the model to adapt to changes in parameters, BCs, and ICs. We apply this approach to three different problems. First, we analyze the Rossler ODE system, demonstrating the robustness of the model with respect to noise and its ability to generalize. Next, we explore the model's capability in a 2D Navier-Stokes PDE problem involving flow past a cylinder with a parametric sinusoidal inlet velocity function, showing that the model can encode pressure data from a few points to identify the inlet velocity profile and utilize physics to compute velocity and pressure throughout the domain. Finally, we address a 1D heat monitoring problem using real data from the heating of glass fiber and thermoplastic composite plates.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 30 canonical work pages

  1. [18]

    Machine Learning 114(1), 1–36 (2025)

    Elaarabi, M., Borzacchiello, D., Le Bot, P., LE Guennec, Y., Comas-Cardona, S.: Adaptive parameters identification for nonlinear dynamics using deep permuta- tion invariant networks. Machine Learning 114(1), 1–36 (2025)

    work page 2025

  2. [1]

    Nature machine intelligence 1(5), 206–215 (2019)

    Rudin, C.: Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature machine intelligence 1(5), 206–215 (2019)

    work page 2019

  3. [2]

    Advances in neural information processing systems 31 (2018)

    Jacot, A., Gabriel, F., Hongler, C.: Neural tangent kernel: Convergence and gen- eralization in neural networks. Advances in neural information processing systems 31 (2018)

    work page 2018

  4. [3]

    Nature Reviews Physics3(6), 422–440 (2021) 35

    Karniadakis, G.E., Kevrekidis, I.G., Lu, L., Perdikaris, P., Wang, S., Yang, L.: Physics-informed machine learning. Nature Reviews Physics3(6), 422–440 (2021) 35

    work page 2021

  5. [4]

    Reliability Engineering & System Safety 217, 107961 (2022)

    Chao, M.A., Kulkarni, C., Goebel, K., Fink, O.: Fusing physics-based and deep learning models for prognostics. Reliability Engineering & System Safety 217, 107961 (2022)

    work page 2022

  6. [5]

    arXiv preprint arXiv:2003.04919 1(1), 1–34 (2020)

    Willard, J., Jia, X., Xu, S., Steinbach, M., Kumar, V.: Integrating physics-based modeling with machine learning: A survey. arXiv preprint arXiv:2003.04919 1(1), 1–34 (2020)

    arXiv 2020

  7. [6]

    International Journal of Heat and Mass Transfer 217, 124671 (2023)

    Xu, J., Wei, H., Bao, H.: Physics-informed neural networks for studying heat transfer in porous media. International Journal of Heat and Mass Transfer 217, 124671 (2023)

    work page 2023

  8. [7]

    Computers & Fluids 248, 105632 (2022)

    Yang, Y., Mesri, Y.: Learning by neural networks under physical constraints for simulation in fluid mechanics. Computers & Fluids 248, 105632 (2022)

    work page 2022

Show all 38 references
  1. [8]

    Algorithms 15(2), 53 (2022)

    Ma, Y., Xu, X., Yan, S., Ren, Z.: A preliminary study on the resolution of electro- thermal multi-physics coupling problem using physics-informed neural network (pinn). Algorithms 15(2), 53 (2022)

    2022

  2. [9]

    PhD thesis, Universit´ e Cˆ ote d’Azur, Inria, CNRS, LJAD (2023)

    Coulaud, G., Duvigneau, R.: Physics-informed neural networks for multiphysics coupling: application to conjugate heat transfer. PhD thesis, Universit´ e Cˆ ote d’Azur, Inria, CNRS, LJAD (2023)

    2023

  3. [10]

    Journal of Computational Physics451, 110844 (2022)

    Penwarden, M., Zhe, S., Narayan, A., Kirby, R.M.: Multifidelity modeling for physics-informed neural networks (pinns). Journal of Computational Physics451, 110844 (2022)

    2022

  4. [11]

    Journal of Computational physics 378, 686–707 (2019)

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

    2019

  5. [12]

    Elsevier (2024)

    Hanna, J.M., Aguado, J.V., Comas-Cardona, S., Le Guennec, Y., Borzacchiello, D.: A self-supervised learning framework based on physics-informed and convo- lutional neural networks to identify local anisotropic permeability tensor from textiles 2D images for filling pattern pre...

    2024

  6. [13]

    Inverse Problems 41(3), 035006 (2025)

    Boulenc, H., Bouclier, R., Garambois, P.-A., Monnier, J.: Spatially-distributed parameter identification by physics-informed neural networks illustrated on the 2d shallow-water equations. Inverse Problems 41(3), 035006 (2025)

    2025

  7. [14]

    arXiv preprint arXiv:1907.04502 (2019)

    LU, L., MENG, X., MAO, Z., KARNIADAKIS, G.E.: Deepxde: A deep learning library for solving differential equations. arXiv preprint arXiv:1907.04502 (2019)

    2019 arXiv

  8. [15]

    Proceedings of the Royal Society A 474(2219), 20180335 (2018) 36

    Kaiser, E., Kutz, J.N., Brunton, S.L.: Sparse identification of nonlinear dynam- ics for model predictive control in the low-data limit. Proceedings of the Royal Society A 474(2219), 20180335 (2018) 36

    2018

  9. [16]

    Proceedings of the national academy of sciences 113(15), 3932–3937 (2016)

    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 sciences 113(15), 3932–3937 (2016)

    2016

  10. [17]

    Nature communications 12(1), 6136 (2021)

    Chen, Z., Liu, Y., Sun, H.: Physics-informed learning of governing equations from scarce data. Nature communications 12(1), 6136 (2021)

    2021

  11. [19]

    Advances in neural information processing systems 30 (2017)

    Zaheer, M., Kottur, S., Ravanbakhsh, S., Poczos, B., Salakhutdinov, R.R., Smola, A.J.: Deep sets. Advances in neural information processing systems 30 (2017)

    2017

  12. [20]

    In: International Conference on Machine Learning, pp

    Lee, J., Lee, Y., Kim, J., Kosiorek, A., Choi, S., Teh, Y.W.: Set transformer: A framework for attention-based permutation-invariant neural networks. In: International Conference on Machine Learning, pp. 3744–3753 (2019). PMLR

    2019

  13. [21]

    Advances in neural information processing systems 30 (2017)

    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 systems 30 (2017)

    2017

  14. [22]

    Parameters 22(72K), 84

    Chiu, S.-T., Hong, J., Braga-Neto, U.: Deeposets: Non-autoregressive in-context learning of operators. Parameters 22(72K), 84

  15. [23]

    arXiv preprint arXiv:1910.03193 (2019)

    Lu, L., Jin, P., Karniadakis, G.E.: Deeponet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators. arXiv preprint arXiv:1910.03193 (2019)

    2019 arXiv

  16. [24]

    Physical Review Research 4(2), 023174 (2022)

    Chen, Y., Luo, Y., Liu, Q., Xu, H., Zhang, D.: Symbolic genetic algorithm for discovering open-form partial differential equations (sga-pde). Physical Review Research 4(2), 023174 (2022)

    2022

  17. [25]

    Journal of Computational Physics 399, 108925 (2019)

    Long, Z., Lu, Y., Dong, B.: Pde-net 2.0: Learning pdes from data with a numeric- symbolic hybrid deep network. Journal of Computational Physics 399, 108925 (2019)

    2019

  18. [26]

    arXiv preprint arXiv:2004.08424 (2020)

    Silva, B.M., Champion, K., Quade, M., Loiseau, J.-C., Kutz, J.N., Brunton, S.L.: Pysindy: a python package for the sparse identification of nonlinear dynamics from data. arXiv preprint arXiv:2004.08424 (2020)

    2020 arXiv

  19. [27]

    Journal of the Royal Statistical Society Series B: Statistical Methodology 58(1), 267–288 (1996)

    Tibshirani, R.: Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society Series B: Statistical Methodology 58(1), 267–288 (1996)

    1996

  20. [28]

    IEEE Access 7, 1404–1423 (2018) 37

    Zheng, P., Askham, T., Brunton, S.L., Kutz, J.N., Aravkin, A.Y.: A unified frame- work for sparse relaxed regularized regression: Sr3. IEEE Access 7, 1404–1423 (2018) 37

    2018

  21. [29]

    Science advances 3(4), 1602614 (2017)

    Rudy, S.H., Brunton, S.L., Proctor, J.L., Kutz, J.N.: Data-driven discovery of partial differential equations. Science advances 3(4), 1602614 (2017)

    2017

  22. [30]

    Nature methods 17(3), 261–272 (2020)

    Virtanen, P., Gommers, R., Oliphant, T.E., Haberland, M., Reddy, T., Cour- napeau, D., Burovski, E., Peterson, P., Weckesser, W., Bright, J., et al.: Scipy 1.0: fundamental algorithms for scientific computing in python. Nature methods 17(3), 261–272 (2020)

    2020

  23. [31]

    Nature Methods 17, 261–272 (2020) https://doi.org/10.1038/ s41592-019-0686-2

    Virtanen, P., Gommers, R., Oliphant, T.E., Haberland, M., Reddy, T., Courna- peau, D., Burovski, E., Peterson, P., Weckesser, W., Bright, J., van der Walt, S.J., Brett, M., Wilson, J., Millman, K.J., Mayorov, N., Nelson, A.R.J., Jones, E., Kern, R., Larson, E., Carey, C.J., Po...

    2020

  24. [32]

    Sobol, I.: The distribution of points in a cube and the accurate evaluation of integrals (in russian) zh. Vychisl. Mat. i Mater. Phys 7, 784–802 (1967)

    1967

  25. [33]

    Peerj computer science 7, 623 (2021)

    Chicco, D., Warrens, M.J., Jurman, G.: The coefficient of determination r-squared is more informative than smape, mae, mape, mse and rmse in regression analysis evaluation. Peerj computer science 7, 623 (2021)

    2021

  26. [34]

    Data Mining and Knowledge Discovery 37(2), 788–832 (2023)

    Hewamalage, H., Ackermann, K., Bergmeir, C.: Forecast evaluation for data scien- tists: common pitfalls and best practices. Data Mining and Knowledge Discovery 37(2), 788–832 (2023)

    2023

  27. [35]

    Available at https://jsdokken

    Dokken, J.S.: Navier-Stokes Solver using FEniCSx. Available at https://jsdokken. com/dolfinx-tutorial/chapter2/ns code2.html. Accessed: 27-Mar-2025 (2023)

    2023

  28. [36]

    Available at https://wwwold.mathematik

    DFG Benchmark Problem. Available at https://wwwold.mathematik. tu-dortmund.de/∼featflow/en/benchmarks/cfdbenchmarking/flow/dfg benchmark1 re20.html. Accessed: 27-Mar-2025

    2025

  29. [37]

    Computer Methods in Applied Mechanics and Engineering 396, 115100 (2022) https://doi.org/10.1016/j.cma.2022.115100

    Hanna, J.M., Aguado, J.V., Comas-Cardona, S., Askri, R., Borzacchiello, D.: Residual-based adaptivity for two-phase flow simulation in porous media using physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering 396, 115100 (2022) https://doi.org/...

    2022

  30. [38]

    https://arxiv.org/abs/2006.10739 38

    Tancik, M., Srinivasan, P.P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, J.T., Ng, R.: Fourier Features Let Net- works Learn High Frequency Functions in Low Dimensional Domains (2020). https://arxiv.org/abs/2006.10739 38

    2020 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.