pith. sign in

arxiv: 2604.14201 · v1 · submitted 2026-04-03 · ⚛️ physics.comp-ph · physics.flu-dyn

LSTM-PINN for Steady-State Electrothermal Transport: Preserving Multi-Field Consis tency in Strongly Coupled Heat and Fluid Flow

Yuqing Zhou , Ze Tao , Hanxuan Wang , Fujun Liu This is my paper

Pith reviewed 2026-05-13 18:28 UTC · model grok-4.3

classification ⚛️ physics.comp-ph physics.flu-dyn
keywords LSTM-PINNelectrothermal transportphysics-informed neural networksmultiphysics couplingheat transferfluid flowconvective regimescross-field consistency
0
0 comments X

The pith

LSTM memory in PINNs preserves long-range dependencies to enforce cross-field consistency in strongly coupled electrothermal transport.

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

Standard physics-informed neural networks struggle to solve steady-state electrothermal problems because heat, fluid, and electric fields have very different gradient scales and stiff residuals. The paper proposes an LSTM-PINN that adds a depth-recursive memory mechanism to keep spatial features linked across the entire domain and across all fields at once. Tested on four regimes including buoyancy-driven convection and porous-media drag, the architecture reduces non-physical artifacts while producing the lowest global errors. A sympathetic reader would see this as a practical way to make multiphysics neural solvers reliable enough for engineering design of energy systems. The claim rests on quantitative comparisons showing consistent gains over both ordinary and attention-based PINNs.

Core claim

The LSTM-PINN framework uses a depth-recursive memory mechanism to preserve long-range spatial feature dependencies and enforce strict cross-field consistency within a unified five-field formulation of electrothermal transport. In the Boussinesq electrothermal flow, drift-potential gauge-constrained transport, strong buoyancy-coupled convection, and Brinkman-Forchheimer drift regimes, it suppresses non-physical artifacts and structural distortions, achieving the highest thermodynamic fidelity and the lowest global error metrics compared with conventional and attention-based networks.

What carries the argument

The depth-recursive memory mechanism borrowed from LSTM layers, which retains information across network depth to link spatial features and enforce consistency among fields whose residuals differ sharply in scale and stiffness.

If this is right

  • Non-physical artifacts and structural distortions are suppressed across all four tested convective and drag regimes.
  • Thermodynamic fidelity reaches its highest recorded value for the given five-field formulation.
  • Global error metrics improve consistently over both standard PINNs and attention-based variants.
  • Localized boundary layers and energy-momentum feedback loops are captured more accurately.
  • The same architecture supplies a stable computational baseline for simulating advanced electrothermal energy systems.

Where Pith is reading between the lines

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

  • The same memory mechanism could be applied to time-dependent or three-dimensional versions of the same multiphysics problem without redesigning the loss terms.
  • Other stiff coupled systems such as plasma transport or reacting flows might benefit from identical depth-recursive memory to reduce manual loss weighting.
  • If the approach generalizes, it could lower the computational cost of retraining by allowing reuse of memory states across similar geometries.
  • Direct comparison against experimental temperature and velocity fields from a real electrothermal device would test whether the numerical gains translate to physical accuracy.

Load-bearing premise

The LSTM memory can keep long-range spatial dependencies intact and force all physical fields to remain consistent even when their gradient magnitudes and residual stiffnesses differ by orders of magnitude.

What would settle it

A new test case in which the LSTM-PINN produces higher global error or visible non-physical temperature or velocity fields than a standard PINN under identical training budgets would falsify the claim that the memory mechanism reliably enforces cross-field consistency.

Figures

Figures reproduced from arXiv: 2604.14201 by Fujun Liu, Hanxuan Wang, Yuqing Zhou, Ze Tao.

Figure 1
Figure 1. Figure 1: The proposed LSTM-PINN framework for steady-state electrothermal systems. The diagram details the depth￾recursive architecture, spatial coordinate encoding, composite physical residual evaluation, and gated memory transport, integrating them into a cohesive workflow tailored to resolve the strongly coupled five-field interactions. Y.Zhou et al.: Preprint submitted to Elsevier Page 3 of 18 [PITH_FULL_IMAGE… view at source ↗
Figure 2
Figure 2. Figure 2: Visual comparison of the predicted electrothermal fields for Case 1. The top strip reports the convergence history, while the subsequent rows display the spatial distributions of 𝑢, 𝑣, 𝑝, 𝜙, and 𝑇 . The left block contrasts the exact benchmark against the network predictions, and the right block visualizes the corresponding absolute-error maps. Y.Zhou et al.: Preprint submitted to Elsevier Page 9 of 18 [P… view at source ↗
Figure 3
Figure 3. Figure 3: Visual comparison of the predicted fields for the drift-potential gauge configuration (Case 2). This scenario replaces direct pressure anchoring with a global zero-mean gauge constraint and introduces drift-potential coupling, specifically evaluating the capacity of each network to reconstruct the interactive multiphysics fields under indirect macroscopic pressure constraints. Y.Zhou et al.: Preprint submi… view at source ↗
Figure 4
Figure 4. Figure 4: Visual comparison of the predicted fields for the buoyancy-coupled electrothermal configuration (Case 3). This physical regime emphasizes advection-dominated transport and strong thermal-to-momentum feedback, explicitly evaluating the capacity of each architecture to resolve complex convective interactions without inducing numerical artifacts. Y.Zhou et al.: Preprint submitted to Elsevier Page 13 of 18 [P… view at source ↗
Figure 5
Figure 5. Figure 5: Visual comparison of the predicted fields for the Brinkman-Forchheimer drift configuration (Case 4). This regime combines intense non-linear interfacial drag with drift-potential coupling, specifically highlighting the rigorous competition between the memory-enhanced and residual-attention architectures by displaying the absolute-error maps alongside the reconstructed multiphysics distributions. Y.Zhou et … view at source ↗
read the original abstract

Steady-state electrothermal systems involve strongly coupled heat transfer, fluid flow, and electric-potential transport, creating severe numerical challenges for standard physics-informed neural networks (PINNs) due to stark disparities in gradient scales and residual stiffnesses across the physical fields. To resolve these multiphysics bottlenecks, we introduce a Long Short-Term Memory PINN (LSTM-PINN) framework that utilizes a depth-recursive memory mechanism to preserve long-range spatial feature dependencies and maintain strict cross-field consistency. The proposed architecture is rigorously evaluated against conventional and attention-based networks across a unified five-field formulation encompassing four complex convective and drag regimes: Boussinesq electrothermal flow, drift-potential gauge-constrained transport, strong buoyancy-coupled convection, and Brinkman--Forchheimer drift. Quantitative and visual analyses demonstrate that LSTM-PINN successfully suppresses non-physical artifacts and structural distortions, yielding the highest thermodynamic fidelity and consistently outperforming state-of-the-art baselines in global error metrics. Ultimately, this memory-enhanced approach provides a highly robust and accurate computational baseline for capturing localized boundary layers and complex energy-momentum feedback in advanced electrothermal energy systems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 3 minor

Summary. The manuscript introduces an LSTM-PINN framework that augments standard physics-informed neural networks with depth-recursive LSTM memory to solve steady-state electrothermal transport problems. It targets strongly coupled five-field systems (heat, fluid flow, electric potential) where gradient scale disparities cause stiffness in conventional PINNs. The architecture is evaluated on four regimes—Boussinesq electrothermal flow, drift-potential gauge-constrained transport, strong buoyancy-coupled convection, and Brinkman–Forchheimer drift—claiming lower global L2 errors, suppression of non-physical artifacts, and higher thermodynamic fidelity than conventional and attention-based baselines.

Significance. If the reported error reductions and artifact suppression hold under the provided training details and tables, the work supplies a practical, empirically validated route to enforcing cross-field consistency in multiphysics PINNs without explicit scale-balancing terms. The inclusion of architecture diagrams, five-field loss formulation, and quantitative comparisons across regimes strengthens its utility as a computational baseline for electrothermal energy systems.

major comments (2)
  1. [§4.2 and Table 2] §4.2 and Table 2: the central claim that the LSTM depth-recursive mechanism alone enforces strict cross-field consistency despite disparate gradient scales is not fully supported by the reported global L2 errors; an ablation isolating memory depth from parameter count or explicit residual weighting is needed to attribute the gains to the architecture rather than capacity.
  2. [§3.1] §3.1: the five-field residual loss is described but the weighting coefficients for each field's PDE residual (velocity, temperature, electric potential, etc.) are not stated; without these or an automatic balancing scheme, it remains unclear whether the observed consistency arises from the LSTM memory or from manual loss tuning.
minor comments (3)
  1. [Abstract] Abstract: the phrase 'unified five-field formulation' appears without naming the fields; this should be stated explicitly in the first paragraph of the introduction.
  2. [Figure 3] Figure 3: visual field comparisons lack quantitative colorbar ranges, making it difficult to assess the magnitude of artifact suppression relative to the baselines.
  3. [§5] §5: the discussion of thermodynamic fidelity would be strengthened by reporting at least one derived integral quantity (e.g., total heat flux or energy balance error) in addition to pointwise L2 norms.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which help clarify the attribution of our results and improve reproducibility. We address each major comment below and will incorporate revisions in the next version of the manuscript.

read point-by-point responses

  1. Referee: [§4.2 and Table 2] §4.2 and Table 2: the central claim that the LSTM depth-recursive mechanism alone enforces strict cross-field consistency despite disparate gradient scales is not fully supported by the reported global L2 errors; an ablation isolating memory depth from parameter count or explicit residual weighting is needed to attribute the gains to the architecture rather than capacity.

    Authors: We agree that an ablation isolating memory depth from parameter count would strengthen the attribution of gains to the recursive LSTM mechanism. In the revised manuscript we will add a controlled ablation: LSTM-PINN variants with increasing memory depth (1, 2, 3 layers) while keeping total trainable parameters approximately constant by proportionally reducing hidden-layer width. Global L2 errors and cross-field consistency metrics will be reported for these variants alongside the original results. This will allow readers to separate architectural effect from capacity. revision: yes

  2. Referee: [§3.1] §3.1: the five-field residual loss is described but the weighting coefficients for each field's PDE residual (velocity, temperature, electric potential, etc.) are not stated; without these or an automatic balancing scheme, it remains unclear whether the observed consistency arises from the LSTM memory or from manual loss tuning.

    Authors: We acknowledge that the exact weighting coefficients were omitted. In the revised §3.1 we will explicitly list the five-field loss weights (λ_u, λ_T, λ_φ, λ_p, λ_ρ) used in all experiments, together with the scaling rationale based on characteristic gradient magnitudes. We will also state that these weights are fixed after a short grid search on a single representative case and are not adjusted during training; the LSTM memory is therefore the only mechanism that can adaptively maintain consistency across fields once the weights are set. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The manuscript advances an empirical claim: an LSTM-PINN architecture with depth-recursive memory outperforms baselines on global L2 errors and visual fidelity across four electrothermal regimes. No derivation chain is presented that reduces a target quantity to a fitted parameter or self-citation by construction. The five-field formulation, loss terms, and architecture diagrams are described as inputs to training; performance is then measured against external baselines rather than recovered from those inputs. No self-definitional equations, fitted-input predictions, or load-bearing self-citations appear in the abstract or supporting context. The result is therefore self-contained against the reported experimental benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The approach rests on the assumption that standard neural network training can enforce physics residuals while the LSTM memory handles long-range dependencies; no new physical axioms or entities are introduced beyond the known governing equations.

free parameters (1)
  • LSTM hidden dimensions and memory depth
    Chosen to capture spatial dependencies but specific values and selection process not reported in abstract.
axioms (1)
  • domain assumption The five-field formulation (heat, fluid, electric potential plus two auxiliary fields) correctly captures the steady-state electrothermal physics including Boussinesq, Brinkman-Forchheimer, and gauge constraints.
    Invoked when defining the unified problem that the network is trained to satisfy.

pith-pipeline@v0.9.0 · 5513 in / 1147 out tokens · 27306 ms · 2026-05-13T18:28:06.557840+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

41 extracted references · 41 canonical work pages · 1 internal anchor

  1. [1]

    On Sta- bility and Instability of Standing Waves for 2d-Nonlinear Schr¨ odinger Equations with Point Interaction

    Zhuang Zhao, Ye Wang, Weijian Zhang, Zhenggang Ba, and Lin Sun. Physics-informed neural networks in heat transfer-dominated multiphysics systems: A comprehensive review.Engineering Applications of Artificial Intelligence, 157:111098, 2025. doi: 10.1016/j. engappai.2025.111098

    work page doi:10.1016/j 2025

  2. [2]

    Physics-informed neural network for real-time thermal modeling of large-scale borehole thermal energy storage systems.Energy, 315:134344, 2025

    Pengchao Li, Fang Guo, Yongfei Li, Xuejing Yang, and Xudong Yang. Physics-informed neural network for real-time thermal modeling of large-scale borehole thermal energy storage systems.Energy, 315:134344, 2025. doi: 10.1016/j.energy.2024.134344

    work page doi:10.1016/j.energy.2024.134344 2025

  3. [3]

    Physics-informed neural operators for generalizable and label-free inference of temperature-dependent thermoelectric properties.npj Computational Materials, 11:272, 2025

    Hyeonbin Moon et al. Physics-informed neural operators for generalizable and label-free inference of temperature-dependent thermoelectric properties.npj Computational Materials, 11:272, 2025. doi: 10.1038/s41524-025-01769-1

    work page doi:10.1038/s41524-025-01769-1 2025

  4. [4]

    Physics-informed neural network digital twin for real-time energy-state tracking in multivariate thermochemical heat storage.Journal of Energy Storage, 139:118686, 2025

    Gwangwoo Han, Sung Kook Hong, and Joo Hyun Moon. Physics-informed neural network digital twin for real-time energy-state tracking in multivariate thermochemical heat storage.Journal of Energy Storage, 139:118686, 2025. doi: 10.1016/j.est.2025.118686

    work page doi:10.1016/j.est.2025.118686 2025

  5. [5]

    PISA-Net: A physics-informed structure-aware neural network for multiphysics field reconstruction in liquid cooling systems.Applied Thermal Engineering, 2026

    Chaobo Zhang et al. PISA-Net: A physics-informed structure-aware neural network for multiphysics field reconstruction in liquid cooling systems.Applied Thermal Engineering, 2026. doi: 10.1016/j.applthermaleng.2025.129347. Online 2025

    work page doi:10.1016/j.applthermaleng.2025.129347 2026

  6. [6]

    Amethodforestimatinglithium-ionbatterystateofhealthbasedonphysics-informedhybridneuralnetwork.Electrochimica Acta, 525:146110, 2025

    YufuLuoetal. Amethodforestimatinglithium-ionbatterystateofhealthbasedonphysics-informedhybridneuralnetwork.Electrochimica Acta, 525:146110, 2025. doi: 10.1016/j.electacta.2025.146110

    work page doi:10.1016/j.electacta.2025.146110 2025

  7. [7]

    Physics-informedmachinelearningmethodforlithium-ionbatteriesstateofhealthestimationusingsimulatedvoltagedata

    RunzeWangetal. Physics-informedmachinelearningmethodforlithium-ionbatteriesstateofhealthestimationusingsimulatedvoltagedata. Applied Energy, 2026. In press

    work page 2026

  8. [8]

    Physics-informedneuralnetworkforco-estimationofstateofhealth,remaining usefullife,andshort-termdegradationpathinlithium-ionbatteries.AppliedEnergy,398:126427,2025

    LiYang,MingjianHe,YataoRen,BaohaiGao,andHongQi. Physics-informedneuralnetworkforco-estimationofstateofhealth,remaining usefullife,andshort-termdegradationpathinlithium-ionbatteries.AppliedEnergy,398:126427,2025. doi:10.1016/j.apenergy.2025.126427

    work page doi:10.1016/j.apenergy.2025.126427 2025

  9. [9]

    Physics-informeddual-stagenetworkforlithium-ionbatterystate-of- chargeestimationundervariousagingandtemperatureconditions.AppliedEnergy,401:126770,2025

    DongheeSon,ShinaPark,JunseokOh,TaehanLee,andSangWooKim. Physics-informeddual-stagenetworkforlithium-ionbatterystate-of- chargeestimationundervariousagingandtemperatureconditions.AppliedEnergy,401:126770,2025. doi:10.1016/j.apenergy.2025.126770

    work page doi:10.1016/j.apenergy.2025.126770 2025

  10. [10]

    Leveraging the capabilities of physics-informed neural networks for channel optimization in PEM fuel cells.Applied Energy, 409:127478, 2026

    Julian Nicolas Toussaint, Sebastian Pieper, Max Paul Mally, and Stefan Pischinger. Leveraging the capabilities of physics-informed neural networks for channel optimization in PEM fuel cells.Applied Energy, 409:127478, 2026. doi: 10.1016/j.apenergy.2026.127478

    work page doi:10.1016/j.apenergy.2026.127478 2026

  11. [11]

    Physics-informedneuralnetworkforthin-filmevaporationin hierarchical structures.International Journal of Heat and Mass Transfer, 224:125296, 2024

    AmirmohammadJahanbakhsh,RojanFiruznia,SinaNazifi,andHadiGhasemi. Physics-informedneuralnetworkforthin-filmevaporationin hierarchical structures.International Journal of Heat and Mass Transfer, 224:125296, 2024. doi: 10.1016/j.ijheatmasstransfer.2024.125296

    work page doi:10.1016/j.ijheatmasstransfer.2024.125296 2024

  12. [12]

    T-phPINN: Physics-informed neural networks for solving 2d non-fourier heat conduction equations.International Journal of Heat and Mass Transfer, 235:126216, 2024

    Jinglai Zheng, Fan Li, and Haiming Huang. T-phPINN: Physics-informed neural networks for solving 2d non-fourier heat conduction equations.International Journal of Heat and Mass Transfer, 235:126216, 2024. doi: 10.1016/j.ijheatmasstransfer.2024.126216

    work page doi:10.1016/j.ijheatmasstransfer.2024.126216 2024

  13. [13]

    Transfer learning throughphysics-informedneuralnetworksforbubblegrowthinsuperheatedliquiddomains.InternationalJournalofHeatandMassTransfer, 232:125940, 2024

    Darioush Jalili, Mohammad Jadidi, Amir Keshmiri, Bhaskar Chakraborty, Anastasios Georgoulas, and Yasser Mahmoudi. Transfer learning throughphysics-informedneuralnetworksforbubblegrowthinsuperheatedliquiddomains.InternationalJournalofHeatandMassTransfer, 232:125940, 2024. doi: 10.1016/j.ijheatmasstransfer.2024.125940

    work page doi:10.1016/j.ijheatmasstransfer.2024.125940 2024

  14. [14]

    Physics-informed neural networks for effective diffusion characteristics inversion in packed bed at low flow rates.International Journal of Heat and Mass Transfer, 244:126970, 2025

    Mouhao Wang, Shanshan Bu, Bing Zhou, Baoping Gong, Zhenzhong Li, and Deqi Chen. Physics-informed neural networks for effective diffusion characteristics inversion in packed bed at low flow rates.International Journal of Heat and Mass Transfer, 244:126970, 2025. doi: 10.1016/j.ijheatmasstransfer.2025.126970

    work page doi:10.1016/j.ijheatmasstransfer.2025.126970 2025

  15. [15]

    Physics-informed neural networks for two-phase film boiling heat transfer.International Journal of Heat and Mass Transfer, 241:126680, 2025

    Darioush Jalili and Yasser Mahmoudi. Physics-informed neural networks for two-phase film boiling heat transfer.International Journal of Heat and Mass Transfer, 241:126680, 2025. doi: 10.1016/j.ijheatmasstransfer.2025.126680

    work page doi:10.1016/j.ijheatmasstransfer.2025.126680 2025

  16. [16]

    Physics informed neural networks for solving inverse thermal wave coupled boundary-value problems.International Journal of Heat and Mass Transfer, 245:126985, 2025

    Hong Tang, Alexander Melnikov, MingRui Liu, Stefano Sfarra, Hai Zhang, and Andreas Mandelis. Physics informed neural networks for solving inverse thermal wave coupled boundary-value problems.International Journal of Heat and Mass Transfer, 245:126985, 2025. doi: 10.1016/j.ijheatmasstransfer.2025.126985

    work page doi:10.1016/j.ijheatmasstransfer.2025.126985 2025

  17. [17]

    Physics-informed neural network for estimating surface heat flux from surface temperature measurements.International Journal of Heat and Mass Transfer, 253:127580, 2025

    Zemin Cai, Xiangqi Lin, and Tianshu Liu. Physics-informed neural network for estimating surface heat flux from surface temperature measurements.International Journal of Heat and Mass Transfer, 253:127580, 2025. doi: 10.1016/j.ijheatmasstransfer.2025.127580

    work page doi:10.1016/j.ijheatmasstransfer.2025.127580 2025

  18. [18]

    KangchengGao,LinQiu,GefeiLi,FajieWang,JiLin,andYanGu. Aphysics-informedframeworkbasedonradialbasisfunctionsforforward and inverse heat conduction problems in anisotropic materials.International Journal of Heat and Mass Transfer, 252:127517, 2025. doi: 10.1016/j.ijheatmasstransfer.2025.127517

    work page doi:10.1016/j.ijheatmasstransfer.2025.127517 2025

  19. [19]

    S. J. Anagnostopoulos et al. Residual-based attention in physics-informed neural networks.Computer Methods in Applied Mechanics and Engineering, 2024

    work page 2024

  20. [20]

    Loss-attentional physics-informed neural networks.Journal of Computational Physics, 2024

    Yuan Song et al. Loss-attentional physics-informed neural networks.Journal of Computational Physics, 2024

    work page 2024

  21. [21]

    Self-adaptiveweightsbasedonbalancedresidualdecayratesinphysics-informedneuralnetworks.JournalofComputational Physics, 2025

    WenChenetal. Self-adaptiveweightsbasedonbalancedresidualdecayratesinphysics-informedneuralnetworks.JournalofComputational Physics, 2025. Y.Zhou et al.:Preprint submitted to ElsevierPage 17 of 18 LSTM-PINN for Steady Electrothermal Multiphysics

    work page 2025

  22. [22]

    Adaptive deep physics-informed neural network with dual adaptivity for complex partial differential equations.Computer Methods in Applied Mechanics and Engineering, 2025

    Ting Wang et al. Adaptive deep physics-informed neural network with dual adaptivity for complex partial differential equations.Computer Methods in Applied Mechanics and Engineering, 2025

    work page 2025

  23. [23]

    Martín-Guerrero, Roberto Ruiz de Austri-Bazan, Alejandro Torres-Forné, and José A

    Antonio Ferrer-Sánchez, José D. Martín-Guerrero, Roberto Ruiz de Austri-Bazan, Alejandro Torres-Forné, and José A. Font. Gradient- annihilated PINNs for solving Riemann problems: Application to relativistic hydrodynamics.Computer Methods in Applied Mechanics and Engineering, 424:116906, 2024. doi: 10.1016/j.cma.2024.116906

    work page doi:10.1016/j.cma.2024.116906 2024

  24. [24]

    Mixed-depth physics-informed neural network with nested finite basis functions for complex partial differential equations

    Ting Wang et al. Mixed-depth physics-informed neural network with nested finite basis functions for complex partial differential equations. Computer Methods in Applied Mechanics and Engineering, 2025

    work page 2025

  25. [25]

    doi:10.1016/j.cma.2024.117116

    VictoritaDolean,AlexanderHeinlein,SiddharthaMishra,andBenMoseley.Multileveldomaindecomposition-basedarchitecturesforphysics- informedneuralnetworks.ComputerMethodsinAppliedMechanicsandEngineering,page117116,2024. doi:10.1016/j.cma.2024.117116

    work page doi:10.1016/j.cma.2024.117116 2024

  26. [26]

    Adaptivefiniteelementinterpolatedneuralnetworks.ComputerMethodsinAppliedMechanics and Engineering, 2025

    SantiagoBadia,WeiLi,andAlbertoF.Martín. Adaptivefiniteelementinterpolatedneuralnetworks.ComputerMethodsinAppliedMechanics and Engineering, 2025

    work page 2025

  27. [27]

    Santiago Badia, Wei Li, and Alberto F. Martín. Compatible finite element interpolated neural networks.Computer Methods in Applied Mechanics and Engineering, page 117889, 2025. doi: 10.1016/j.cma.2025.117889

    work page doi:10.1016/j.cma.2025.117889 2025

  28. [28]

    Wang et al

    Y. Wang et al. A practical PINN framework for multi-scale problems with reconstructing the loss function and associating it with specialized neural network architectures.Journal of Computational Physics, 2024

    work page 2024

  29. [29]

    Wang et al

    Y. Wang et al. NAS-PINN: Neural architecture search-guided physics-informed neural network for solving partial differential equations. Journal of Computational Physics, 2024

    work page 2024

  30. [30]

    Learnable activation functions in physics-informed neural networks for solving partial differential equations.Computer Physics Communications, page 109753, 2025

    Afrah Farea and Mustafa Serdar Celebi. Learnable activation functions in physics-informed neural networks for solving partial differential equations.Computer Physics Communications, page 109753, 2025

    work page 2025

  31. [31]

    Khademi et al

    A. Khademi et al. Physics-informed neural networks with trainable sinusoidal activation for solving Navier–Stokes equations.Computer Physics Communications, 2025

    work page 2025

  32. [32]

    Subramani

    Sumanth Kumar Boya and Deepak N. Subramani. PINTO: Physics-informed transformer neural operator for learning generalized solutions of partial differential equations for any initial and boundary condition.Computer Physics Communications, page 109702, 2025. doi: 10.1016/j.cpc.2025.109702

    work page doi:10.1016/j.cpc.2025.109702 2025

  33. [33]

    Physics-informed neural operators for efficient modeling of infiltration in porous media.Journal of Computational Physics, 538:114156, 2025

    Hamza Kamil, Azzeddine Soulaïmani, and Abdelaziz Beljadid. Physics-informed neural operators for efficient modeling of infiltration in porous media.Journal of Computational Physics, 538:114156, 2025. doi: 10.1016/j.jcp.2025.114156

    work page doi:10.1016/j.jcp.2025.114156 2025

  34. [34]

    Hu et al

    Z. Hu et al. Tackling the curse of dimensionality with physics-informed neural networks.Neural Networks, 2024

    work page 2024

  35. [35]

    Lstm-pinn: An hybrid method for prediction of steady-state electrohydrodynamic flow.Journal of Computational Physics, page 114586, 2025

    Ze Tao, Ke Xu, and Fujun Liu. Lstm-pinn: An hybrid method for prediction of steady-state electrohydrodynamic flow.Journal of Computational Physics, page 114586, 2025

    work page 2025

  36. [36]

    Physics-informedneuralnetworkswithcontinuous-timethermalmodellingformodelpredictivecontrol of building energy systems in demand response.Energy and Buildings, 352:116864, 2026

    DanWang,WenjuHu,andWeiWang. Physics-informedneuralnetworkswithcontinuous-timethermalmodellingformodelpredictivecontrol of building energy systems in demand response.Energy and Buildings, 352:116864, 2026

    work page 2026

  37. [37]

    Modeling dynamic gas–liquid interfaces in underwater explosions using interval- constrained physics-informed neural networks.Physics of Fluids, 37(11), 2025

    Fulin Xing, Junjie Li, Ze Tao, Fujun Liu, and Yong Tan. Modeling dynamic gas–liquid interfaces in underwater explosions using interval- constrained physics-informed neural networks.Physics of Fluids, 37(11), 2025

    work page 2025

  38. [38]

    Analyticalandneural network approaches for solving two-dimensional nonlinear transient heat conduction.arXiv preprint arXiv:2504.02845, 2025

    Ze Tao, Fujun Liu, Jinhua Li, and Guibo Chen. Analytical and neural network approaches for solving two-dimensional nonlinear transient heat conduction.arXiv preprint arXiv:2504.02845, 2025

    work page arXiv 2025

  39. [39]

    An lstm-pinn hybrid method to the specific problem of population forecasting.arXiv preprint arXiv:2505.01819, 2025

    Ze Tao. An lstm-pinn hybrid method to the specific problem of population forecasting.arXiv preprint arXiv:2505.01819, 2025

    work page arXiv 2025

  40. [40]

    LNN-PINN: A Unified Physics-Only Training Framework with Liquid Residual Blocks

    Ze Tao, Hanxuan Wang, and Fujun Liu. LNN-PINN: A unified physics-only training framework with liquid residual blocks.arXiv preprint arXiv:2508.08935, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  41. [41]

    Residual attention physics-informed neural networks for robust multiphysics simulation of steady-state electrothermal energy systems.arXiv preprint arXiv:2603.23578, 2026

    Yuqing Zhou, Ze Tao, and Fujun Liu. Residual attention physics-informed neural networks for robust multiphysics simulation of steady-state electrothermal energy systems.arXiv preprint arXiv:2603.23578, 2026. Y.Zhou et al.:Preprint submitted to ElsevierPage 18 of 18

    work page arXiv 2026