pith. sign in

arxiv: 2604.02663 · v1 · submitted 2026-04-03 · 💻 cs.LG · cs.NA· math.NA

A Numerical Method for Coupling Parameterized Physics-Informed Neural Networks and FDM for Advanced Thermal-Hydraulic System Simulation

Jeesuk Shin , Donggyun Seo , Sihyeong Yu , Joongoo Jeon This is my paper

Pith reviewed 2026-05-13 20:39 UTC · model grok-4.3

classification 💻 cs.LG cs.NAmath.NA
keywords physics-informed neural networksfinite difference methodthermal-hydraulic simulationnuclear safety analysisparameterized PINNhybrid numerical methodmomentum conservationdata-free surrogate
0
0 comments X

The pith

A parameterized neural network trained only on momentum conservation, coupled with finite difference mass updates, simulates thermal-hydraulic systems accurately across time steps and initial conditions without retraining or simulation data

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

The paper develops the P2F method to speed up severe accident analysis in nuclear system codes by avoiding the need for large simulation datasets and repeated network training. It introduces a node-assigned parameterized PINN that accepts water-level difference, initial velocity, and time as inputs to learn the solution manifold for the momentum equation across flow paths. This network is coupled at each time step to a finite difference method that advances the mass conservation equation, guaranteeing exact discrete mass conservation while replacing the usual iterative momentum solve with one fast forward pass. Tests on a six-tank gravity-driven draining scenario produce water-level mean absolute error of 7.85e-5 m and velocity mean absolute error of 3.21e-3 m/s under nominal conditions, with accuracy holding for time steps from 0.2 to 1.0 s and five different initial conditions.

Core claim

The P2F framework couples a parameterized Node-Assigned PINN (NA-PINN) with a finite difference method (FDM) for the CVH/FP module in MELCOR. The NA-PINN takes water-level difference, initial velocity, and time as inputs and learns the momentum conservation solution manifold so that one trained network serves as a data-free surrogate for all flow paths. The FDM then updates mass conservation at each time step using the network-predicted velocities, ensuring exact discrete mass conservation. Verification on a six-tank gravity-driven draining scenario yields the stated mean absolute errors and shows consistent accuracy across the tested time steps and initial conditions without retraining.

What carries the argument

Parameterized Node-Assigned PINN (NA-PINN) that accepts water-level difference, initial velocity, and time to learn the momentum conservation solution manifold, coupled with FDM for mass conservation updates

If this is right

  • Parametric studies and uncertainty quantification for nuclear safety become feasible at lower computational cost because no simulation data or retraining is required for new conditions
  • The momentum solve is replaced by a single network forward pass at each time step while exact discrete mass conservation is preserved by the FDM update
  • Accuracy holds for time steps from 0.2 s to 1.0 s and for five distinct initial conditions in the tested six-tank draining scenario
  • The same trained network works across multiple flow paths without modification

Where Pith is reading between the lines

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

  • The same parameterization approach could be tested on other conservation laws by adding parameters for energy or species transport to the network inputs
  • Embedding the hybrid solver inside real-time monitoring systems for nuclear plants would be a direct next step if latency remains low
  • The method could transfer to non-nuclear fluid systems that solve coupled mass-momentum equations at system scale

Load-bearing premise

The velocities produced by the NA-PINN trained only on the momentum equation remain accurate enough that the FDM mass updates stay stable and physically correct when the network is embedded in a full system code that includes additional physics modules.

What would settle it

A new simulation run with a time step outside the 0.2-1.0 s range or an unseen initial condition in which the network velocities cause mass to become negative or the FDM update to diverge

Figures

Figures reproduced from arXiv: 2604.02663 by Donggyun Seo, Jeesuk Shin, Joongoo Jeon, Sihyeong Yu.

Figure 1
Figure 1. Figure 1: Architecture of the original NA-PINN [24]. Each nodal variable ( [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Nodalization for the scenario model. 14 [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Architecture of the reformulated NA-PINN for a single FP node. The FCN [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Standalone PINN predictions versus reference FDM solutions for three input [PITH_FULL_IMAGE:figures/full_fig_p022_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Architecture of the P2F framework. At each time step, the water level in each CV [PITH_FULL_IMAGE:figures/full_fig_p023_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Verification under the nominal initial condition ( [PITH_FULL_IMAGE:figures/full_fig_p027_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Generalization to unseen initial conditions ( [PITH_FULL_IMAGE:figures/full_fig_p030_7.png] view at source ↗
read the original abstract

Severe accident analysis using system-level codes such as MELCOR is indispensable for nuclear safety assessment, yet the computational cost of repeated simulations poses a significant bottleneck for parametric studies and uncertainty quantification. Existing surrogate models accelerate these analyses but depend on large volumes of simulation data, while physics-informed neural networks (PINNs) enable data-free training but must be retrained for every change in problem parameters. This study addresses both limitations by developing the Parameterized PINNs coupled with FDM (P2F) method, a node-assigned hybrid framework for MELCOR's Control Volume Hydrodynamics/Flow Path (CVH/FP) module. In the P2F method, a parameterized Node-Assigned PINN (NA-PINN) accepts the water-level difference, initial velocity, and time as inputs, learning a solution manifold so that a single trained network serves as a data-free surrogate for the momentum conservation equation across all flow paths without retraining. This PINN is coupled with a finite difference method (FDM) solver that advances the mass conservation equation at each time step, ensuring exact discrete mass conservation while replacing the iterative nonlinear momentum solve with a single forward pass. Verification on a six-tank gravity-driven draining scenario yields a water level mean absolute error of $7.85 \times 10^{-5}$ m and a velocity mean absolute error of $3.21 \times 10^{-3}$ m/s under the nominal condition with $\Delta t = 1.0$ s. The framework maintains consistent accuracy across time steps ranging from 0.2 to 1.0 s and generalizes to five distinct initial conditions, all without retraining or simulation data. This work introduces a numerical coupling methodology for integrating parameterized PINNs with FDM within a nuclear thermal-hydraulic system code framework.

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

3 major / 2 minor

Summary. The manuscript introduces the P2F framework, which couples a parameterized Node-Assigned PINN (NA-PINN) surrogate for the momentum equation with an FDM solver for mass conservation in MELCOR CVH/FP thermal-hydraulic simulations. The NA-PINN takes water-level difference, initial velocity, and time as inputs to learn a solution manifold, enabling a single trained network to replace iterative momentum solves across flow paths without retraining or simulation data. Verification on a six-tank gravity-driven draining scenario reports water-level MAE of 7.85e-5 m and velocity MAE of 3.21e-3 m/s at nominal Δt=1.0 s, with consistent accuracy for time steps 0.2-1.0 s and generalization to five initial conditions.

Significance. If the coupling stability and accuracy hold under additional physics (friction, form losses, coupled modules), the method could reduce computational cost for parametric studies and UQ in nuclear safety analysis while preserving exact discrete mass conservation, offering a data-free alternative to data-intensive surrogates.

major comments (3)
  1. [Verification / Results] Verification section: the central claim of generalization without retraining for the target MELCOR application rests on a single gravity-driven draining scenario at fixed geometry; the NA-PINN inputs exclude friction and form-loss parameters, so stability of the explicit FDM mass update under altered driving forces is unverified and load-bearing for the no-retraining assertion.
  2. [Method / Coupling] Coupling description (abstract and §3): no analysis or numerical tests address accumulation of small velocity errors in the FDM mass update over many steps or when additional modules modify effective forces, leaving the claimed physical consistency and stability unproven beyond the simple case.
  3. [Abstract / Results] Abstract and results: reported MAEs lack error bars, network-size convergence studies, or direct comparison against the original MELCOR iterative solver, weakening quantitative support for the accuracy claims across time steps and initial conditions.
minor comments (2)
  1. [Method] Notation for NA-PINN architecture and training loss could be expanded with explicit equations or a diagram to clarify how the momentum residual is parameterized.
  2. [Verification] The six-tank scenario description would benefit from a schematic or table of geometric parameters to aid reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major point below and have revised the manuscript accordingly to strengthen the verification, coupling analysis, and quantitative support.

read point-by-point responses

  1. Referee: [Verification / Results] Verification section: the central claim of generalization without retraining for the target MELCOR application rests on a single gravity-driven draining scenario at fixed geometry; the NA-PINN inputs exclude friction and form-loss parameters, so stability of the explicit FDM mass update under altered driving forces is unverified and load-bearing for the no-retraining assertion.

    Authors: We agree that the verification is limited to a single gravity-driven draining scenario at fixed geometry without friction or form losses. The current NA-PINN parameterization (water-level difference, initial velocity, time) is chosen to demonstrate the core coupling for this physics. Friction and form-loss coefficients can be incorporated as additional inputs to extend the manifold without retraining, preserving the no-retraining property. In the revised manuscript we add explicit discussion of this extensibility in §4 and note that the presented case is a proof-of-concept; we also include a short stability check of the explicit FDM update under the tested driving forces. revision: partial

  2. Referee: [Method / Coupling] Coupling description (abstract and §3): no analysis or numerical tests address accumulation of small velocity errors in the FDM mass update over many steps or when additional modules modify effective forces, leaving the claimed physical consistency and stability unproven beyond the simple case.

    Authors: We acknowledge the absence of long-term error accumulation analysis. The revised manuscript adds numerical tests in §4 that track velocity and water-level errors over an extended number of steps (beyond the original simulation horizon) for the six-tank case, confirming that errors remain bounded. We also clarify that exact discrete mass conservation is enforced at every FDM step regardless of velocity surrogate error, and we discuss how the modular coupling would propagate errors when additional modules alter effective forces, with this left for future validation. revision: yes

  3. Referee: [Abstract / Results] Abstract and results: reported MAEs lack error bars, network-size convergence studies, or direct comparison against the original MELCOR iterative solver, weakening quantitative support for the accuracy claims across time steps and initial conditions.

    Authors: We will update the abstract and §4 to report MAEs with error bars computed from five independent training runs with different random seeds. A network-size convergence study is added showing that the chosen architecture reaches the reported accuracy plateau. We also include a direct side-by-side comparison of both accuracy and wall-clock time per step against the original MELCOR iterative momentum solver on the same six-tank scenario, confirming comparable accuracy at lower per-step cost. revision: yes

Circularity Check

0 steps flagged

No circularity: verification uses external reference comparisons

full rationale

The P2F method trains a parameterized NA-PINN on the momentum equation and couples it to an explicit FDM mass update. Reported water-level and velocity MAEs (7.85e-5 m and 3.21e-3 m/s) are obtained by direct comparison against an independent reference solution on the six-tank gravity-draining test case, not by algebraic reduction to the network's fitted parameters or inputs. No self-definitional equations, fitted quantities renamed as predictions, or load-bearing self-citations appear in the derivation. Generalization across time steps and initial conditions is demonstrated numerically rather than enforced by construction. The central claim therefore remains independent of its own outputs.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

Only the abstract is available, so the ledger is necessarily incomplete; the central claim rests on the unstated assumption that the momentum residual can be minimized to sufficient accuracy by a single network across the parameter range, plus standard neural-network training assumptions.

free parameters (1)
  • NA-PINN architecture and training hyperparameters
    Network depth, width, activation, loss weights, and optimizer settings are chosen to make the momentum residual small; these are free parameters fitted during training.
axioms (1)
  • domain assumption The momentum conservation equation can be expressed as a residual that a neural network can minimize without boundary data for the specific flow-path geometry.
    Invoked when the authors state the PINN learns the solution manifold for the momentum equation.
invented entities (1)
  • Node-Assigned Parameterized PINN (NA-PINN) no independent evidence
    purpose: Surrogate for momentum conservation that accepts parameters as inputs so one network covers multiple flow paths and conditions.
    New named construct introduced to enable the hybrid solver; no independent evidence outside the paper is provided.

pith-pipeline@v0.9.0 · 5654 in / 1497 out tokens · 37076 ms · 2026-05-13T20:39:58.292180+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

39 extracted references · 39 canonical work pages · 1 internal anchor

  1. [1]

    I. A. E. Agency, Status and evaluation of severe accident simulation codes for water cooled reactors, Tech. rep. (2019)

    work page 2019

  2. [2]

    R. O. Gauntt, et al., Melcor computer code manuals, volume 2: Reference manual, version 2.2.9496, Tech. rep., Albuquerque, NM, sAND2021-xxxx, available from NRC or Sandia National Labs (2021)

    work page 2021

  3. [3]

    Wagner, B

    K. Wagner, B. A. Beeny, D. L. Luxat, F. Gelbard, D. L. Louie, L. I. Albright, L. L. Humphries, Melcor integrated severe accident code ap- plication to safety assessment of high-temperature gas-cooled reactors, Nuclear Engineering and Design 402 (2023) 112083

    work page 2023

  4. [4]

    J. Jeon, J. Lee, S. J. Kim, Finite volume method network for the ac- celeration of unsteady computational fluid dynamics: Non-reacting and reacting flows, International Journal of Energy Research 46 (8) (2022) 10770–10795

    work page 2022

  5. [5]

    Khanal, S

    S. Khanal, S. Baral, J. Jeon, Comparison of cnn-based deep learning architectures for unsteady cfd acceleration on small datasets, Nuclear Engineering and Technology 57 (10) (2025) 103703

    work page 2025

  6. [6]

    Y. H. Chae, H. Kim, J. Bang, P. H. Seong, Development of a data-driven simulation framework using physics-informed neural network, Annals of Nuclear Energy 189 (2023) 109840

    work page 2023

  7. [7]

    W. Wang, W. Ma, Coupling of melcor with surrogate model for quench estimation of conical debris beds, Annals of Nuclear Energy 211 (2025) 110933

    work page 2025

  8. [8]

    W. Wang, W. Ma, Bootstrapped artificial neural network model for uncertainty analysis in melcor simulation of severe accident, Progress in Nuclear Energy 157 (2023) 104556

    work page 2023

  9. [9]

    J. Song, K. Ha, A simulation and machine learning informed diagnosis of the severe accidents, Nuclear Engineering and Design 395 (2022) 111881

    work page 2022

  10. [10]

    Y. Lee, S. H. Song, J. Y. Bae, K. Song, M. R. Seo, S. J. Kim, J. I. Lee, Surrogate model for predicting severe accident progression in nuclear power plant using deep learning methods and rolling-window forecast, Annals of Nuclear Energy 208 (2024) 110816. 33

    work page 2024

  11. [11]

    C. H. Song, W. Choi, S. J. Kim, Performance evaluation of a surrogate model to predict effective dose under hypothesized severe accidents of pressurized water reactor based on neural network, Annals of Nuclear Energy 224 (2025) 111701

    work page 2025

  12. [12]

    Baraldi, C

    P. Baraldi, C. Battistini, F. Antonello, J. Buongiorno, E. Zio, Physics- informed neural networks for the safety analysis of nuclear reactors, Progress in Nuclear Energy 185 (2025) 105745

    work page 2025

  13. [13]

    Raissi, P

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

    work page 2019

  14. [14]

    G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, L. Yang, Physics-informed machine learning, Nature Reviews Physics 3 (6) (2021) 422–440

    work page 2021

  15. [15]

    S. Cai, Z. Wang, S. Wang, P. Perdikaris, G. E. Karniadakis, Physics- informed neural networks for heat transfer problems, Journal of Heat Transfer 143 (6) (2021) 060801

    work page 2021

  16. [16]

    Z. Zhao, Y. Wang, W. Zhang, Z. Ba, L. Sun, Physics-informed neural networks in heat transfer-dominated multiphysics systems: A compre- hensive review, Engineering Applications of Artificial Intelligence 157 (2025) 111098

    work page 2025

  17. [17]

    S. Cai, Z. Mao, Z. Wang, M. Yin, G. E. Karniadakis, Physics-informed neural networks (pinns) for fluid mechanics: A review, Acta Mechanica Sinica 37 (12) (2021) 1727–1738

    work page 2021

  18. [18]

    M. H. Elhareef, Z. Wu, Physics-informed neural network method and ap- plication to nuclear reactor calculations: A pilot study, Nuclear Science and Engineering 197 (4) (2023) 601–622

    work page 2023

  19. [19]

    Coppo Leite, E

    V. Coppo Leite, E. Merzari, A. Novak, R. Ponciroli, L. Ibarra, Applica- tion of a physics-informed convolutional neural network for monitoring the temperature fields in high-temperature gas reactors, Nuclear Science and Engineering 199 (10) (2025) 1712–1732. 34

    work page 2025

  20. [20]

    S. Wang, S. Sankaran, P. Perdikaris, Respecting causality for training physics-informed neural networks, Computer Methods in Applied Me- chanics and Engineering 421 (2024) 116813

    work page 2024

  21. [21]

    Chen, S.-K

    Z. Chen, S.-K. Lai, Z. Yang, At-pinn: Advanced time-marching physics- informed neural network for structural vibration analysis, Thin-Walled Structures 196 (2024) 111423

    work page 2024

  22. [22]

    J. Jeon, J. Lee, R. Vinuesa, S. J. Kim, Residual-based physics-informed transfer learning: A hybrid method for accelerating long-term cfd simu- lations via deep learning, International Journal of Heat and Mass Trans- fer 220 (2024) 124900

    work page 2024

  23. [23]

    XRePIT: A deep learning-computational fluid dynamics hybrid framework implemented in OpenFOAM for fast, robust, and scalable unsteady simulations

    S. Baral, Y. Lee, S. Khanal, J. Jeon, Residual-guided ai-cfd hybrid method enables stable and scalable simulations: from 2d benchmarks to 3d applications, arXiv preprint arXiv:2510.21804 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  24. [24]

    J. Shin, C. Kim, S. Yang, M. Lee, S. J. Kim, J. Jeon, Node assigned physics-informed neural networks for thermal-hydraulic system simula- tion: Cvh/fl module, arXiv preprint arXiv:2504.16447 (2025)

    work page arXiv 2025

  25. [25]

    W. Cho, M. Jo, H. Lim, K. Lee, D. Lee, S. Hong, N. Park, Parameterized physics-informed neural networks for parameterized pdes, 2024

    work page 2024

  26. [26]

    Y. XIE, Y. MA, Y. WANG, Solving steady-state mono-energy neutron diffusion solution set with parameterized physics-informed neural net- work, Atomic Energy Science and Technology 58 (6) (2024) 1242–1249

    work page 2024

  27. [27]

    C. J. Arthurs, A. P. King, Active training of physics-informed neural networkstoaggregateandinterpolateparametricsolutionstothenavier- stokes equations, Journal of Computational Physics 438 (2021) 110364

    work page 2021

  28. [28]

    Jeong, J

    H. Jeong, J. Bai, C. P. Batuwatta-Gamage, C. Rathnayaka, Y. Zhou, Y. Gu, A physics-informed neural network-based topology optimization (pinnto) framework for structural optimization, Engineering Structures 278 (2023) 115484

    work page 2023

  29. [29]

    J. Wang, Y. Mo, B. Izzuddin, C.-W. Kim, Exact dirichlet boundary physics-informed neural network epinn for solid mechanics, Computer Methods in Applied Mechanics and Engineering 414 (2023) 116184. 35

    work page 2023

  30. [30]

    S. Jo, S. Park, J. Shin, J. Park, H. Kim, S. Ko, S. Lee, J. Jeon, Task- aware evolution in physics-informed neural networks: Application to saint-venant torsion problems, Engineering Applications of Artificial In- telligence 168 (2026) 113988

    work page 2026

  31. [31]

    S. Wang, Y. Teng, P. Perdikaris, Understanding and mitigating gradient flow pathologies in physics-informed neural networks, SIAM Journal on Scientific Computing 43 (5) (2021) A3055–A3081

    work page 2021

  32. [32]

    Shukla, A

    K. Shukla, A. D. Jagtap, G. E. Karniadakis, Parallel physics-informed neural networks via domain decomposition, Journal of Computational Physics 447 (2021) 110683

    work page 2021

  33. [33]

    A. D. Jagtap, G. E. Karniadakis, Extended physics-informed neural net- works (xpinns): A generalized space-time domain decomposition based deep learning framework for nonlinear partial differential equations, Communications in Computational Physics 28 (5) (2020)

    work page 2020

  34. [34]

    Khadijeh, V

    M. Khadijeh, V. Cerqueglini, C. Kasbergen, S. Erkens, A. Varveri, Mul- tistage physics informed neural network for solving coupled multiphysics problems in material degradation and fluid dynamics, Engineering with Computers (2025) 1–31

    work page 2025

  35. [35]

    Donnelly, A

    J. Donnelly, A. Daneshkhah, S. Abolfathi, Physics-informed neural net- works as surrogate models of hydrodynamic simulators, Science of the Total Environment 912 (2024) 168814

    work page 2024

  36. [36]

    Poubeau, Y

    B. Poubeau, Y. Richet, L. Chailan, F. Mascari, M. Massone, S. Gianfe- lici, L. E. Herranz, J. Fontanet, T. Lind, C. D’alessandro, et al., Horizon euratom assas project: can machine-learning make fast and accurate se- vere accident simulators a reality?, in: The 11 th European Review Meeting on Severe Accident Research (ERMSAR2024), 2024, pp. 1–12

    work page 2024

  37. [37]

    Xiang, W

    Z. Xiang, W. Peng, W. Yao, W. Zhou, Hybrid finite-difference physics- informed neural networks partial differential equation solver for complex geometries, Journal of Thermophysics and Heat Transfer 39 (4) (2025) 686–703

    work page 2025

  38. [38]

    A. O. Beitalmal, A hybrid physics-informed neural network (pinn) and finite element method (fem) framework for multiscale fluid-structure in- teraction problems, RA Journal of Applied Research 11 (04) (2025). 36

    work page 2025

  39. [39]

    W. Wang, M. Hakimzadeh, H. Ruan, S. Goswami, Time-marching neu- ral operator–fe coupling: Ai-accelerated physics modeling, Computer Methods in Applied Mechanics and Engineering 446 (2025) 118319. 37

    work page 2025