pith. the verified trust layer for science. sign in

arxiv: 2604.27298 · v1 · submitted 2026-04-30 · ⚛️ physics.plasm-ph · physics.comp-ph

DeepPropNet: an operator learning-based predictor for thermal plasma properties

Zuo Wang , Linlin Zhong This is my paper

Pith reviewed 2026-05-07 08:58 UTC · model grok-4.3

classification ⚛️ physics.plasm-ph physics.comp-ph
keywords thermal plasma propertiesoperator learningDeepPropNetthermodynamic propertiestransport propertiesSF6-N2C4F7N-CO2-O2neural network prediction
0
0 comments X p. Extension

The pith

DeepPropNet uses operator learning to map plasma conditions to thermodynamic and transport properties with 0.1-1% error.

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

The paper introduces DeepPropNet to solve the slow evaluation of thermal plasma properties, which depend in a strongly nonlinear way on temperature, pressure, and gas composition. It builds two neural architectures, a single-property network and a mixture-of-experts multi-property network, trained on high-fidelity data for SF6-N2 and C4F7N-CO2-O2 mixtures. Both versions reach relative L2 errors of order 10^{-3} to 10^{-2} and continue to work on operating points withheld from training. The models are then coupled to finite-volume solvers and physics-informed networks to run full plasma simulations. A reader cares because the approach replaces repeated expensive property calculations with fast forward passes inside existing simulation codes.

Core claim

DeepPropNet learns the nonlinear operator that takes plasma operating conditions and returns thermodynamic and transport properties. For binary SF6-N2 and ternary C4F7N-CO2-O2 mixtures, the single-property and MoE-based versions produce predictions whose relative L2 errors lie between 10^{-3} and 10^{-2}. The learned operator generalizes to unseen conditions and can be inserted directly into finite-volume and physics-informed neural-network solvers.

What carries the argument

DeepPropNet, an operator-learning neural network (with S-DeepPropNet and MoE-DeepPropNet variants) that learns the direct mapping from temperature, pressure, and composition to physical properties.

If this is right

  • Property evaluation inside plasma codes becomes orders of magnitude faster than traditional table lookup or direct calculation.
  • The same trained model works for both single properties and all properties together through the MoE architecture.
  • Coupling to finite-volume methods produces accurate plasma-flow solutions without recomputing properties at every step.
  • Coupling to physics-informed networks allows data-driven solution of plasma equations while using the learned properties.
  • The framework scales to additional mixtures once new high-fidelity data are supplied.

Where Pith is reading between the lines

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

  • The same operator-learning strategy could be applied to other nonlinear property sets in high-temperature gases or materials.
  • Embedding the model inside real-time control loops for plasma devices becomes feasible because each evaluation is a cheap neural forward pass.
  • Adding uncertainty quantification on top of DeepPropNet would let engineers know when a prediction is reliable for safety-critical equipment.
  • Training on data from multiple simulation codes could produce a single model usable across different plasma modeling communities.

Load-bearing premise

The high-fidelity datasets used for training cover the relevant ranges of temperature, pressure, and composition so that the learned mapping generalizes without large extrapolation errors.

What would settle it

Run the trained model on a gas mixture or extreme temperature-pressure point never present in the training data and check whether the relative L2 error stays inside 10^{-3} to 10^{-2}; a jump above that range would disprove reliable generalization.

Figures

Figures reproduced from arXiv: 2604.27298 by Linlin Zhong, Zuo Wang.

Figure 1
Figure 1. Figure 1: Diagram of S-DeepPropNet for single-property prediction. view at source ↗
Figure 2
Figure 2. Figure 2: Diagram of MoE-DeepPropNet for multi-property prediction. view at source ↗
Figure 3
Figure 3. Figure 3: Prediction of thermodynamic and transport properties of SF view at source ↗
Figure 4
Figure 4. Figure 4: Prediction of thermodynamic and transport properties of SF view at source ↗
Figure 5
Figure 5. Figure 5: Prediction of thermodynamic and transport properties of C view at source ↗
Figure 6
Figure 6. Figure 6: The integration of DeepPropNet with FVM and PINN frameworks, including the offline forecast and online view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of temperature distributions obtained using plasma properties from traditional calculations and view at source ↗
Figure 8
Figure 8. Figure 8: (a) Computational domain and monitoring locations. (b) Temperature evolution at Position A (arc center), view at source ↗
Figure 9
Figure 9. Figure 9: Prediction of radial temperature distribution of a one-dimensional stationary arc for SF view at source ↗
Figure 10
Figure 10. Figure 10: presents the transient arc simulation for an SF6-N2 (60%-40%) plasma mixture at 1 bar. The MoE-PINN employs a feed-forward neural network with six hidden layers and 200 neurons per layer. The spatiotemporal evolution of temperature shows a gradual cooling process, with the temperature in the arc core decreasing over time while maintaining a smooth spatial distribution. This behavior is physically consiste… view at source ↗
read the original abstract

Thermal plasma properties play a critical role in plasma simulations and plasma-related applications. However, their strong nonlinear dependence on temperature, pressure, and gas composition makes accurate and efficient evaluation challenging. In this work, an operator learning-based model, termed DeepPropNet, is proposed for fast prediction of thermodynamic and transport properties of thermal plasmas. Two architectures are developed, including a single-property model (S-DeepPropNet) and a Mixture of Experts (MoE)-based multi-property model (MoE-DeepPropNet). The proposed models learn the nonlinear mapping from plasma operating conditions to physical properties based on high-fidelity datasets. The MoE architecture enables efficient multi-property prediction within a unified framework. Predictions are performed for binary SF6-N2 and ternary C4F7N-CO2-O2 mixtures. The results show that the proposed models achieve high accuracy, with relative L2 errors on the order of 10-3 to 10-2, while maintaining strong generalization capability under unseen conditions. The applicability of DeepPropNet is further demonstrated by coupling with finite volume method (FVM) and physics-informed neural networks (PINNs). The results indicate that DeepPropNet provides an efficient and scalable approach for plasma property prediction and plasma simulations.

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 DeepPropNet, an operator-learning framework with two variants (S-DeepPropNet for single properties and MoE-DeepPropNet for multi-property prediction) that maps plasma operating conditions (temperature, pressure, composition) to thermodynamic and transport properties for binary SF6-N2 and ternary C4F7N-CO2-O2 mixtures. It reports relative L2 errors of order 10^{-3}–10^{-2} on high-fidelity data, claims strong generalization to unseen conditions, and demonstrates coupling to finite-volume and physics-informed neural network solvers.

Significance. If the reported accuracy and generalization hold under rigorous validation, the work would supply a fast, scalable surrogate for expensive equilibrium property calculations that currently limit plasma CFD and PINN simulations. The explicit demonstration of coupling to FVM and PINNs is a concrete strength that directly addresses practical deployment.

major comments (3)
  1. [Abstract and §4] Abstract and §4 (Results): the headline claim of 'strong generalization capability under unseen conditions' with relative L2 errors 10^{-3}–10^{-2} is not supported by any quantitative description of the train/test split, the sampling grid in (T, p, x) space, or extrapolation metrics (e.g., error versus distance from training points). Without these, the generalization statement cannot be evaluated and remains the load-bearing assumption for the central accuracy claim.
  2. [§3] §3 (Methodology): the high-fidelity dataset generation procedure is described only at a high level; no details are given on the underlying solver, the number of points per mixture, the ranges and discretization of temperature/pressure/composition, or any cross-validation strategy. These omissions make it impossible to judge whether the reported errors reflect interpolation within a densely sampled domain or true extrapolation.
  3. [§4.3] §4.3 (Coupling demonstrations): while the FVM and PINN integrations are shown, no quantitative comparison of wall-clock time or accuracy against the original high-fidelity property routine is provided, so the practical speedup and error propagation into the coupled solver remain unquantified.
minor comments (2)
  1. [§2] Notation for mixture compositions (e.g., x_SF6, x_C4F7N) is introduced without a consistent table of symbols or explicit definition of the mole-fraction normalization.
  2. [Figures 5–7] Figure captions for the error heat-maps do not state the exact number of test points or the precise definition of 'unseen conditions' used in each panel.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed review. The comments identify important omissions that limit the evaluability of our claims on generalization and practical utility. We address each major comment below and will revise the manuscript to incorporate the requested quantitative information and clarifications.

read point-by-point responses

  1. Referee: [Abstract and §4] Abstract and §4 (Results): the headline claim of 'strong generalization capability under unseen conditions' with relative L2 errors 10^{-3}–10^{-2} is not supported by any quantitative description of the train/test split, the sampling grid in (T, p, x) space, or extrapolation metrics (e.g., error versus distance from training points). Without these, the generalization statement cannot be evaluated and remains the load-bearing assumption for the central accuracy claim.

    Authors: We agree that the current text does not supply the quantitative details needed to substantiate the generalization claim. In the revised manuscript we will add, in a new subsection of §4, an explicit description of the train/test split, the sampling strategy used to generate the (T, p, x) grid, and an extrapolation analysis (including error plotted against normalized distance to the nearest training point). These additions will allow readers to distinguish interpolation from extrapolation performance and will directly support the reported accuracy figures. revision: yes

  2. Referee: [§3] §3 (Methodology): the high-fidelity dataset generation procedure is described only at a high level; no details are given on the underlying solver, the number of points per mixture, the ranges and discretization of temperature/pressure/composition, or any cross-validation strategy. These omissions make it impossible to judge whether the reported errors reflect interpolation within a densely sampled domain or true extrapolation.

    Authors: We acknowledge that §3 currently provides only a high-level overview. The revised version will expand this section to name the equilibrium solver employed, state the total number of points generated for each mixture, specify the exact ranges and discretization steps for temperature, pressure and composition, and describe the cross-validation procedure used during training. These details will enable an independent assessment of whether the reported errors correspond to interpolation or extrapolation. revision: yes

  3. Referee: [§4.3] §4.3 (Coupling demonstrations): while the FVM and PINN integrations are shown, no quantitative comparison of wall-clock time or accuracy against the original high-fidelity property routine is provided, so the practical speedup and error propagation into the coupled solver remain unquantified.

    Authors: We agree that the coupling demonstrations lack the quantitative benchmarks required to evaluate practical speedup and error propagation. In the revised §4.3 we will add a table (or inline text) reporting wall-clock times for DeepPropNet versus the original high-fidelity routine and will quantify the effect of the approximation on the coupled solver outputs (e.g., differences in predicted temperature or flow fields). This will provide concrete measures of both computational gain and solution fidelity. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected in claimed derivation chain

full rationale

The paper explicitly frames DeepPropNet as a supervised operator-learning model trained on high-fidelity datasets to approximate the mapping from (T, p, composition) to thermodynamic/transport properties. No first-principles derivation, uniqueness theorem, or ansatz is invoked that would reduce outputs to inputs by construction. Accuracy metrics are reported against held-out data drawn from the same high-fidelity source, which is standard supervised-learning practice and does not meet the criteria for fitted-input-called-prediction or self-definitional circularity. No load-bearing self-citations or imported uniqueness results appear in the abstract or described claims. The work is therefore self-contained as an empirical predictor without circular reduction.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The accuracy and generalization claims rest on the representativeness of the high-fidelity training data and the capacity of the neural network to learn the underlying nonlinear operator without explicit physical constraints.

free parameters (2)
  • Neural network parameters (weights, biases, expert routing)
    Learned by minimizing loss on the high-fidelity dataset; no closed-form values given.
  • Mixture composition and condition ranges for training
    Chosen to generate the datasets on which accuracy is reported.
axioms (1)
  • domain assumption Thermal plasma properties admit an accurate operator representation as a nonlinear function of temperature, pressure, and composition
    Invoked by the choice of operator learning framework.

pith-pipeline@v0.9.0 · 5518 in / 1391 out tokens · 46873 ms · 2026-05-07T08:58:24.704157+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

31 extracted references · 5 canonical work pages · 2 internal anchors

  1. [1]

    A. B. Murphy. A perspective on arc welding research: the importance of the arc, unresolved questions and future directions.Plasma Chem. Plasma Process., 35: 471-489, 2015

    2015

  2. [2]

    Sakuyama, H

    T. Sakuyama, H. Kobayashi, M. Ishikawa, H. Kotsuji, K. Shiraishi, H. Urai, and Y . Tanaka. Comparison of measurement and numerical calculation on spectral profile of ablation arc plasma around current zero in high-voltage gas circuit breakers.J. Phys. D: Appl. Phys., 56: 405503, 2023

    2023

  3. [3]

    Kreuznacht, M

    S. Kreuznacht, M. Purcel, S. Böddeker, P. Awakowicz, W. Xia, M. Muhler, M. Böke, and A. v. Keudell. Comparison of the performance of a microwave plasma torch and a gliding arc plasma for hydrogen production via methane pyrolysis.Plasma Process. Polym., 20: 2200132, 2023

    2023

  4. [4]

    J.-R. Riba, M. Moreno-Eguilaz, and J. A. Ortega. Arc fault protections for aeronautic applications: A review identifying the effects, detection methods, current progress, limitations, future challenges, and research needs. IEEE Trans. Instrum. Meas., 71: 3504914, 2022

    2022

  5. [5]

    Zhong, Y

    L. Zhong, Y . Cressault, and P. Teulet. Thermophysical and radiation properties of high-temperature C4F8-CO2 mixtures to replace SF6 in high-voltage circuit breakers.Phys. Plasmas, 25: 033502, 2018

    2018

  6. [6]

    Ren and L

    H. Ren and L. Zhong. Evaluation of arc quenching ability for SF6 replacements based on time-dependent Elenbaas–Heller and Boltzmann equations.J. Appl. Phys., 137: 023301, 2025

    2025

  7. [7]

    Li and Y

    W. Li and Y . Zhang. Numerical simulations on DC-driven low-temperature plasmas by physics-informed neural networks.J. Appl. Phys., 137: 203304, 2025

    2025

  8. [8]

    H. Li, H. Hua, Z. Yu, and Z. Zuo. Physics-informed neural networks for steady power-law fluid flows around a square cylinder in a two-dimensional channel.Phys. Fluids, 38: 043103, 2026

    2026

  9. [9]

    Boullé and A

    N. Boullé and A. Townsend. A mathematical guide to operator learning.Handbook of Numerical Analysis, V ol. 25 (Elsevier, 2024)

    2024

  10. [10]

    Zampla, A

    E. Zampla, A. H. D. O. Fonseca, J. O. Caro, A. H. Moberly, M. J. Higley, J. Cardin, and D. V . Dijk. Learning integral operators via neural integral equations.Nat. Mach. Intell., 6: 1046-1062, 2024

    2024

  11. [11]

    N. B. Kovachki, S. Lanthaler, and A. M. Stuart. Operator learning: Algorithms and analysis.Handbook of Numerical Analysis, V ol. 25 (Elsevier, 2024)

    2024

  12. [12]

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

    work page internal anchor Pith review arXiv 2010

  13. [13]

    L. Lu, P. Jin, G. Pang, Z. Zhang, and G. Karniadakis. Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators.Nat. Mach. Intell., 3: 218-229, 2021

    2021

  14. [14]

    H. Wang, Y . Cao, Z. Huang, Y . Liu, P. Hu, X. Luo, Z. Song, W. Zhao, J. Liu, and J. Sun. Recent Advances on Machine Learning for Computational Fluid Dynamics: A Survey.arXiv:2408.12171, 2024

    work page arXiv 2024

  15. [15]

    Goswami, A

    S. Goswami, A. D. Jagtap, H. Babaee, B. T. Susi, and G. E. Karniadakis. Learning stiff chemical kinetics using extended deep neural operators.Comput. Methods Appl. Mech. Eng., 419: 116674, 2024

    2024

  16. [16]

    Eivazi, J.-A

    H. Eivazi, J.-A. Tröger, S. Wittek, S. Hartmann, and A. Rausch. EquiNO: A Physics-Informed Neural Operator for Multiscale Simulations.arXiv:2504.07976, 2025

    work page arXiv 2025

  17. [17]

    Gopakumar, S

    V . Gopakumar, S. Pamela, L. Zanisi, Z. Li, A. Anandkumar, and M. Team. Fourier Neural Operator for Plasma Modelling.arXiv:2302.06542, 2023. 15 DeepPropNet: an operator learning-based predictor for thermal plasma properties

    work page arXiv 2023

  18. [18]

    Bonotto, D

    M. Bonotto, D. Abate, and L. Pigatto. Reconstruction of plasma equilibrium and separatrix using convolutional physics-informed neural operator.Fusion Eng. Des., 200: 114193, 2024

    2024

  19. [19]

    Jacobs, M

    R. Jacobs, M. Jordan, S. Nowlan, and G. Hinton. Adaptive Mixtures of Local Experts.Neural Comput., 3: 79-87, 1991

    1991

  20. [20]

    Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer

    N. Shazeer, A. Mirhoseini, K. Maziarz, A. Davis, and J. Dean. Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer.arXiv:1701.06538, 2017

    work page internal anchor Pith review arXiv 2017

  21. [21]

    M. Rong, L. Zhong, Y . Cressault, A. Gleizes, X. Wang, F. Chen, and H. Zheng. Thermophysical properties of SF6–Cu mixtures at temperatures of 300–30,000 K and pressures of 0.01–1.0 MPa: part 1. Equilibrium compositions and thermodynamic properties considering condensed phases.J. Phys. D: Appl. Phys., 47: 495202, 2014

    2014

  22. [22]

    X. Wang, L. Zhong, Y . Cressault, A. Gleizes, and M. Rong. Thermophysical properties of SF6–Cu mixtures at temperatures of 300–30,000 K and pressures of 0.01–1.0 MPa: part 2. Collision integrals and transport coefficients. J. Phys. D: Appl. Phys., 47: 495201, 2014

    2014

  23. [23]

    Zhong, M

    L. Zhong, M. Rong, X. Wang, J. Wu, G. Han, G. Han, Y . Lu, A. Yang, and Y . Wu. Compositions, thermodynamic properties, and transport coefficients of high-temperature C5F10O mixed with CO2 and O2 as substitutes for SF6 to reduce global warming potential.AIP Adv., 7: 075003, 2017

    2017

  24. [24]

    Zhong, A

    L. Zhong, A. B. Murphy, X. Wang, and M. Rong. Calculation of two-temperature plasma composition: part 1. Mass action law methods and extremum searching methods.J. Phys. D: Appl. Phys., 53: 065202, 2020

    2020

  25. [25]

    Cressault, V

    Y . Cressault, V . Connord, H. Hingana, P. Teulet, and A. Gleizes. Transport properties of CF3I thermal plasmas mixed with CO2, air or N2 as an alternative to SF6 plasmas in high-voltage circuit breakers.J. Phys. D: Appl. Phys., 44: 495202, 2011

    2011

  26. [26]

    Wang and L

    Z. Wang and L. Zhong. DeepONet-based fast calculation of net emission coefficients of arc plasmas. In:IEEE 20th Conference on Industrial Electronics and Applications (ICIEA’25), 2025, Art. no. 11148744

    2025

  27. [27]

    Paszke, S

    A. Paszke, S. Gross, F. Massa, A. Lerer, and S. Chintala. PyTorch: An Imperative Style, High-Performance Deep Learning Library. In:Advances in Neural Information Processing Systems 32 (NeurIPS’19), 2019, pp. 8026-8037

    2019

  28. [28]

    Zhong, Q

    L. Zhong, Q. Gu, and B. Wu. Graphite production in two-temperature non-LTE plasmas of C4F7N and C5F10O mixed with CO2, N2, and O2 as eco-friendly SF6 replacements: A numerical study.Plasma Process. Polym., 18: 2100036, 2021

    2021

  29. [29]

    Z. Wang, B. Baheti, L. Zhong, Y . Cressault, and P. Teulet. Fluid simulation of two-temperature non-LTE plasmas of C4F7N and C5F10O mixed with CO2, N2 and O2 as eco-friendly SF6 replacements.J. Phys. D: Appl. Phys., (under review)

  30. [30]

    Zhong, Q

    L. Zhong, Q. Gu, and B. Wu. Deep learning for thermal plasma simulation: Solving 1-D arc model as an example. Comput. Phys. Commun., 257: 107496, 2020

    2020

  31. [31]

    Zhong, B

    L. Zhong, B. Wu, and Y . Wang. Low-temperature plasma simulation based on physics-informed neural networks: Frameworks and preliminary applications.Phys. Fluids, 34: 087116, 2022. 16

    2022