Pith. sign in

REVIEW 2 major objections 4 minor 66 references

A deep-learning surrogate compresses ASTEC vessel physics by over 300 imes and rolls out tens of thousands of time steps in under a minute.

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

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

T0 review · grok-4.5

2026-07-11 19:03 UTC pith:2LPW3CGE

load-bearing objection First real multi-variable ASTEC vessel surrogate with honest metrics and open code; open-loop evaluation is the main soft spot, not a hidden fatal flaw. the 2 major comments →

arxiv 2607.04450 v1 pith:2LPW3CGE submitted 2026-07-05 cs.LG cs.AI

A Deep Learning-based surrogate model for Severe Accidents in nuclear reactors using ASTEC

Alessandro Longhi , Danny Lathouwers , Zolt\'an Perk\'o This is my paper
classification cs.LG cs.AI
keywords surrogate modelingdeep learningsevere accidentsneural ODEsdimensionality reductionmulti-physics simulationsASTECautoencoder
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Severe-accident codes such as ASTEC solve the multi-physics of a nuclear-reactor vessel under core-melt conditions, but each run can take days, so they cannot be used for real-time operator training. This paper shows that an auto-encoder paired with a neural ordinary differential equation can learn the vessel's dynamics from a few hundred ASTEC trajectories of station-blackout and loss-of-coolant accidents. The resulting surrogate simultaneously forecasts roughly eighty scalar and field variables (thermal-hydraulics, core degradation, fission-product release) for up to forty hours of accident time, stays stable for fifty thousand autoregressive steps, and finishes in under a minute on either CPU or GPU. The work is presented as the first systematic test of how far purely data-driven latent dynamics can go on the highly non-linear, discontinuous physics that ASTEC produces.

Core claim

An auto-encoder that maps the vessel state (1913 degrees of freedom) into a six-dimensional latent vector, coupled to a neural ODE that advances that latent state under exogenous primary-circuit boundary conditions, yields a stable, multi-variable surrogate of ASTEC's vessel physics whose wall-clock speed is hundreds of times higher than the original code.

What carries the argument

AE-NODE: a multi-branch auto-encoder that separately encodes global, plenum, core, vessel and face fields, followed by a neural ODE whose right-hand side is conditioned on the first upper-plenum and cold-leg volumes; the whole pipeline is trained with an adaptive-window teacher-forcing / autoregressive schedule.

Load-bearing premise

The vessel can be cleanly decoupled from the rest of the plant by treating the neighbouring primary-circuit volumes as fixed external inputs, without closed-loop feedback that would drive later predictions off the training manifold.

What would settle it

Couple the trained surrogate back to a live primary-circuit model (or to ASTEC itself) and check whether the closed-loop vessel predictions remain within the reported error bounds for the full duration of a severe-accident trajectory; a rapid divergence would falsify the open-loop decoupling claim.

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

If this is right

  • Real-time operator-training simulators become feasible because a full forty-hour vessel evolution finishes in under a minute.
  • The same AE-NODE architecture can be reused for other multi-physics modules once similar open-boundary data sets are generated.
  • Dimensionality reduction by more than two orders of magnitude makes ensemble uncertainty quantification of severe-accident scenarios computationally practical.
  • Stable multi-variable roll-outs of tens of thousands of steps demonstrate that latent neural ODEs can handle the discontinuous, stiff dynamics typical of core-melt codes.

Where Pith is reading between the lines

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

  • If the open-loop assumption fails under closed-loop feedback, hybrid schemes that periodically re-anchor the latent state to a cheap physics residual may be required.
  • The same latent-space construction could serve as a common interface for coupling independently trained surrogates of primary, secondary and containment domains.
  • Because the auto-encoder already isolates the major source of error, future gains will come more from better latent reconstruction than from more sophisticated time steppers.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper constructs the first (to the authors’ knowledge) purely data-driven surrogate of the ASTEC vessel domain for LOCA and SBO scenarios. An AutoEncoder compresses ~1913 degrees of freedom into a 6-dimensional latent space; a Neural ODE then advances the latent state under exogenous boundary conditions taken from the adjacent primary-circuit volumes. Separate models are trained on 700 LOCA and 286 SBO trajectories obtained by Sobol sampling of operator actions. On held-out trajectories the surrogate produces stable open-loop rollouts of 10k–50k macro-steps (4–40 h of accident time) in under a minute on CPU or GPU, simultaneously predicting ~80 scalar and field variables, with a reported wall-clock speed-up of roughly 500–900 imes relative to the ICARE module alone.

Significance. If the open-loop results hold under closed-loop coupling, the work would supply a practical building block for real-time severe-accident training simulators and would demonstrate that AE-NODE architectures can handle the multi-physics, multi-scale, discontinuous signals characteristic of integral SA codes. Strengths that should be credited include the public source code, the explicit multi-topology AE design, the adaptive-window autoregressive training schedule, the quantitative RMSE decomposition that isolates AE versus NODE error, and the frank discussion of failure modes on near-zero and highly oscillatory variables. These elements make the study a useful first benchmark even if the operational claim remains provisional.

major comments (2)
  1. The headline claim (Abstract; §4; Table 2) of a usable multi-hour, multi-variable surrogate is measured exclusively under open-loop forcing: at every macro-step the true ASTEC states of the first upper-plenum and cold-leg volumes are supplied as exogenous inputs (Eq. 2, §2.4, Fig. 4). Section 5.1 itself notes that even modest errors in the predicted boundary fluxes s_B1/s_B2 will alter the primary-circuit response and can drive subsequent vessel predictions off the training manifold. Because no closed-loop experiment (or even a simple one-step feedback test) is reported, the evidence does not yet support the operational claim that AE-NODE can replace the ICARE–CESAR coupling inside a real-time simulator. At minimum the authors should quantify error growth under a synthetic closed loop or clearly re-scope the claim to open-loop vessel approximation.
  2. Figures 9 and 15 show that several safety-relevant scalars (x_alpha, P_H2, m_gas, Q_liq_vap, porosity, FP mass-flow rates, and the boundary fluxes themselves) exhibit large RMSE_mean and RMSE_std while remaining low only on RMSE_max. The discussion attributes this to sharp late-time jumps and near-zero plateaus, yet no quantitative criterion is given for when a variable is considered “acceptable” for operator-training use. Without such a criterion, or without an ablation that isolates the contribution of these poorly predicted variables to the overall safety metrics, it is difficult to judge whether the surrogate is already useful or still requires substantial data filtering / capacity increases.
minor comments (4)
  1. Table 2 reports ASTEC times for ICARE only; the caption correctly notes this is a lower bound, but the abstract and conclusions still quote the speed-up relative to “ASTEC”. A single clarifying sentence would avoid over-statement.
  2. The adaptive-window schedule (Appendix E) and the precise values of the loss weights α–ω are given only in tables; a short paragraph in the main text would help reproducibility.
  3. Figures 26–29 illustrate the most difficult variables; adding the corresponding RMSE time series (as already done for s_B1/s_B2 in Figs. 30–31) would make the failure-mode analysis more quantitative.
  4. Minor notation: the same symbol S is used both for the solution manifold and for the solution space of the PDE; a brief distinction would improve readability.

Circularity Check

0 steps flagged

No circularity: purely empirical data-driven surrogate; self-citation of AE-NODE is methodological reuse, not a load-bearing derivation that forces the reported metrics.

full rationale

The paper constructs and evaluates a non-intrusive AE-NODE surrogate (AutoEncoder + Neural ODE) on ASTEC-generated LOCA/SBO trajectories. All headline claims (simultaneous prediction of ~80 vessel variables, stable 10k–50k-step autoregressive rollouts, >300 imes reduction from 1913 to 6 latent dimensions, wall-clock <1 min for ~40 h of accident time) are measured empirical outcomes on held-out test trajectories under open-loop boundary forcing (Eqs. 1–2, §2.4, Figs. 9–21, Table 2). Latent dimension λ=6, network widths, loss weights α…ω and adaptive-window schedule are free hyperparameters chosen by the authors; they do not algebraically force the reported RMSE, stability or speed-up figures. The sole self-citation ([43], same authors) supplies the base AE-NODE architecture that is then adapted (multi-domain sub-encoders, adaptive AR windows, extra loss terms). That citation is ordinary methodological reuse; the present paper’s results do not reduce to any uniqueness theorem, fitted constant or definitional identity imported from [43]. No step of the form “X derives Y but X is defined via Y”, no parameter fitted on a subset and then “predicted”, and no renaming of a known empirical pattern. The acknowledged open-loop limitation (§5.1) is a correctness/scope issue, not circularity. The derivation chain is therefore self-contained and non-circular.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central empirical claim rests on the manifold hypothesis, the adequacy of the sampled operator-action space, the decoupling of vessel from primary circuit, and a collection of architectural and training hyperparameters. No new physical entities are postulated; free parameters are standard ML knobs.

free parameters (4)
  • latent dimension λ = 6
    Fixed at 6 for both LOCA and SBO models (Table 14); chosen by the authors to achieve the reported compression factor.
  • loss weights α,β,γ,δ,ν,μ,ω = 1.0 each
    All set to 1.0 (Table 9); control the relative importance of reconstruction, latent symmetry, teacher-forcing and autoregressive terms.
  • adaptive window schedule L, Nl = L=[10,1000], Nl=[25,25]
    Sequence of window lengths and epoch counts used to ramp autoregressive training (Table 8); hand-chosen.
  • network widths and filter counts = e.g. fθ layers [200,200,200,200]
    Fully-connected and convolutional layer sizes listed in Tables 9–13; architectural free parameters.
axioms (4)
  • domain assumption High-dimensional vessel state lives on a low-dimensional manifold that an autoencoder can learn (manifold hypothesis).
    Invoked in §3.1 to justify the 1913→6 compression; standard in modern ROM/DL but unproved for ASTEC physics.
  • domain assumption Operator-action vectors sampled by low-discrepancy Sobol sequences with the listed temporal constraints adequately cover the relevant accident manifold.
    §2.1 and Appendix C; the training distribution is defined by this sampling.
  • ad hoc to paper Vessel dynamics can be written as a single autonomous PDE driven only by the adjacent primary-circuit volumes as exogenous inputs.
    Equation (2) and §2.4; enables the surrogate to be trained independently of the rest of ASTEC.
  • domain assumption Neural ODEs with explicit Runge–Kutta integration can stably advance the latent state for tens of thousands of steps when trained with adaptive-window teacher-forcing + autoregressive losses.
    §3.2 and Appendix E; standard NODE practice but not guaranteed a priori for discontinuous SA data.

pith-pipeline@v1.1.0-grok45 · 38621 in / 2896 out tokens · 32625 ms · 2026-07-11T19:03:48.374067+00:00 · methodology

0 comments
read the original abstract

Integral codes like the Accident Source Term Evaluation Code (ASTEC) are powerful tools to study the physics of Severe Accidents (SAs) in nuclear reactors. Real time SA simulators can also be helpful in training operators of nuclear plants to react correctly to malfunctions. However, SA simulators can take up to several days per simulation, making their use infeasible for real time applications. In this work we show how to speed up a SA simulator with a fast, Deep Learning based (DL), surrogate model (SM). The SM is built as a combination of a dimensionality reduction stage, via an AutoEncoder, and a time-stepping stage, via a Neural Ordinary Differential Equation. The data on which the SM is trained are obtained from the ASTEC simulator, by sampling a set of operator actions for station blackout (SBO) and loss-of-coolant accidents (LOCA). The objective of the developed SM is to approximate multiple spatio-temporal fields for the thermal-hydraulic physics, core degradation, and fission product release modules in ASTEC's vessel domain. The SM predicts simultaneously around $80$ different physical variables (both scalar and fields), maintaining a stable autoregressive rollout up to $50$ thousand time steps. In addition, the AutoEncoder achieves a dimensionality reduction by a factor of over $300$, which allows the SM to predict up to $40$ hours of simulation in under a minute, both on CPU and GPU. This work is the first study of the capabilities and limits of DL based surrogate modeling in approximating the challenging, highly non-linear physics of ASTEC.

Figures

Figures reproduced from arXiv: 2607.04450 by Alessandro Longhi, Danny Lathouwers, Zolt\'an Perk\'o.

Figure 1
Figure 1. Figure 1: Left: the twelve modules of ASTEC linked to the reactor modules they model. Right: the modules of ASTEC communicate with each other through the ASTEC’s dynamic database. δti is the micro time step of each module, while ∆t A is the macro time-step of ASTEC. Image adapted from the ASNR version in [53]. step of each module, while ∆t A is the macro time-step of ASTEC. Each module is advanced by δti up to ∆t A … view at source ↗
Figure 2
Figure 2. Figure 2: Coupling of the ICARE (core-degradation) and CESAR (thermal-hydraulics) modules. ∆ [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: (a): blue and orange cells are the volumes considered for the construction of the SM. The [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Coupling of the surrogate model (AE-NODE) with the primary circuit. At time [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: On the left the data space W ⊂ R N to which the PDE solution s potentially belongs is displayed together with the actual PDE solution manifold S, to which the solution s is constrained by the PDE. Since the dimensionality of S is λ << N, we can describe the solution s(x, ti |kj ) with a vector ε(ti |kj ) ∈ E ⊂ R λ of only λ dimensions displayed on the right. The Decoder ψ : R λ → S maps the low-dimensional… view at source ↗
Figure 6
Figure 6. Figure 6: On the left, the Encoder φθ = {φ sc θ , φ p θ , φv θ , φcr θ , φ f θ , φcnc θ } is used to encode the solution s(x, ti |kj ) into the latent representation ε(ti |kj ) ∈ R λ of dimension λ. On the right, the latent vector ε(ti |kj ) is mapped by the Decoder ψθ = {ψ sc θ , ψsc θ , ψv θ , ψcr θ , ψf θ , ψcnc θ } into the reconstructed solution s˜(x, ti |kj ). The loss function L reg AE acts as a regularizer o… view at source ↗
Figure 7
Figure 7. Figure 7: The training procedure to optimize fθ during training is split between a Teacher Forcing (TF) and an Autoregressive (AR) approach. At the top the solution fields of a data-point Db,l,j , arranged from the left to the right, are encoded into their latent representations going from ε j b to ε j b+l . In the TF approach, the Processor πθ is only applied once to ε j b+i to give a latent prediction ε j b+i+1,b+… view at source ↗
Figure 8
Figure 8. Figure 8: The application of AE-NODE at inference time. From the left: the initial condition [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Comparison of RMSEmean, RMSEmax and RMSEstd metrics when doing a simple AutoEncoding (AE) and when running the actual inference (AE-NODE) on the LOCA testing set. The uncertainty bars are the standard deviations of the metrics computed across testing trajectories. The green horizontal line is placed at the value 0.5, which is a common baseline for RMSEstd. of the porosity and of the void fraction in the ve… view at source ↗
Figure 10
Figure 10. Figure 10: AE-NODE predictions vs. ground truth for selected (testing) LOCA trajectories of the total [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: AE-NODE predictions vs. ground truth for selected (testing) LOCA trajectories of the void [PITH_FULL_IMAGE:figures/full_fig_p017_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: AE-NODE predictions vs. ground truth for selected (testing) LOCA trajectories of the liquid [PITH_FULL_IMAGE:figures/full_fig_p017_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: AE-NODE predictions vs. ground truth at 4 different time steps of volume and the mass of [PITH_FULL_IMAGE:figures/full_fig_p018_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: AE-NODE predictions vs. ground truth at 4 different time steps of the temperature of the fuel [PITH_FULL_IMAGE:figures/full_fig_p018_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Comparison of RMSEmean, RMSEmax and RMSEstd metrics when doing a simple AutoEn￾coding (AE) and when running the actual inference (AE-NODE) on the SBO testing set. The uncertainty bars are the standard deviations of the metrics computed across testing trajectories. The green horizontal line is placed at the value 0.5, which is a common baseline for RMSEstd. prediction, we plot in Figures 32 and 33 some lat… view at source ↗
Figure 16
Figure 16. Figure 16: AE-NODE predictions vs. ground truth for selected (testing) SBO trajectories of the H2 [PITH_FULL_IMAGE:figures/full_fig_p020_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: AE-NODE predictions vs. ground truth for selected (testing) SBO trajectories of the liquid [PITH_FULL_IMAGE:figures/full_fig_p020_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: AE-NODE predictions vs. ground truth for selected (testing) SBO trajectories of the tempera [PITH_FULL_IMAGE:figures/full_fig_p020_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: AE-NODE predictions vs. ground truth at 4 different time steps of porosity and void fraction [PITH_FULL_IMAGE:figures/full_fig_p021_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: AE-NODE predictions vs. ground truth at 4 different time steps of temperature of the fuel in [PITH_FULL_IMAGE:figures/full_fig_p021_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: CPU and GPU comparison of time (in seconds) required by AE-NODE to simulate the full [PITH_FULL_IMAGE:figures/full_fig_p022_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: The variables belonging to the faces of the vessel, given by ASTEC as 1D vectors of dimension [PITH_FULL_IMAGE:figures/full_fig_p033_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Application of the Savitzky–Golay filter on the void fraction in the plenum over time. [PITH_FULL_IMAGE:figures/full_fig_p037_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: AE-NODE predictions vs. ground truth at 4 different (during the degradation phase) time [PITH_FULL_IMAGE:figures/full_fig_p037_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: AE-NODE predictions vs. ground truth at 4 different (during the degradation phase) time [PITH_FULL_IMAGE:figures/full_fig_p038_25.png] view at source ↗
Figure 26
Figure 26. Figure 26: AE-NODE predictions vs. ground truth for scalar variables with large [PITH_FULL_IMAGE:figures/full_fig_p038_26.png] view at source ↗
Figure 27
Figure 27. Figure 27: AE-NODE predictions vs. ground truth for scalar variables with large [PITH_FULL_IMAGE:figures/full_fig_p039_27.png] view at source ↗
Figure 28
Figure 28. Figure 28: AE-NODE predictions vs. ground truth at 4 different time steps of debris 1 in the vessel and [PITH_FULL_IMAGE:figures/full_fig_p039_28.png] view at source ↗
Figure 29
Figure 29. Figure 29: Challenging trajectory examples showing AE-NODE predictions vs. ground truth for trajectories [PITH_FULL_IMAGE:figures/full_fig_p039_29.png] view at source ↗
Figure 30
Figure 30. Figure 30: RMSEmean per time step for all the LOCA testing trajectories for the boundary variables sB1 and sB2 [PITH_FULL_IMAGE:figures/full_fig_p040_30.png] view at source ↗
Figure 31
Figure 31. Figure 31: RMSEmean per time step for all the SBO testing trajectories for the boundary variables sB1 and sB2 . (a) Trajectory 804 (b) Trajectory 820 (c) Trajectory 838 (d) Trajectory 844 [PITH_FULL_IMAGE:figures/full_fig_p040_31.png] view at source ↗
Figure 32
Figure 32. Figure 32: Auto-regressive latent vector prediction (6 dimensions) over time for some LOCA selected [PITH_FULL_IMAGE:figures/full_fig_p040_32.png] view at source ↗
Figure 33
Figure 33. Figure 33: Auto-regressive latent vector prediction (6 dimensions) over time for some SBO selected [PITH_FULL_IMAGE:figures/full_fig_p041_33.png] view at source ↗
Figure 34
Figure 34. Figure 34: Auto-regressive latent vector prediction (6 dimensions) over time for some LOCA selected [PITH_FULL_IMAGE:figures/full_fig_p041_34.png] view at source ↗
Figure 35
Figure 35. Figure 35: Auto-regressive latent vector prediction (6 dimensions) over time for some SBO selected [PITH_FULL_IMAGE:figures/full_fig_p041_35.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

66 extracted references · 13 canonical work pages

  1. [1]

    Springer International Publishing, 2014

    Reduced Order Methods for Modeling and Computational Reduction. Springer International Publishing, 2014. URL: http://dx.doi.org/10.1007/978-3-319-02090-7 , doi:10.1007/ 978-3-319-02090-7

    doi:10.1007/978-3-319-02090-7 2014

  2. [2]

    Openturns

    Airbus, EDF, IMACS, ONERA, and Phimeca. Openturns. URL: https://openturns.github.io/ www/,doi:10.1007/978-3-319-11259-6_64-1

    doi:10.1007/978-3-319-11259-6_64-1

  3. [3]

    Sarnet, a success story

    Thierry Albiol, Jean-Pierre Van Dorsselaere, and Nils Reinke. Sarnet, a success story. survey of major achievements on severe accidents and of knowledge capitalization within the astec code. 01 2008

    2008

  4. [4]

    Universal physics transformers: A framework for efficiently scaling neural operators

    Benedikt Alkin, Johannes Brandstetter, Andreas F¨ urst, Lukas Gruber, Markus Holzleitner, and Simon Schmid. Universal physics transformers: A framework for efficiently scaling neural operators. InAdvances in Neural Information Processing Systems 37, NeurIPS 2024, page 25152–25194. Neural Information Processing Systems Foundation, Inc. (NeurIPS), 2024. URL...

    work page doi:10.52202/079017-0793 2024

  5. [5]

    Ascher and Linda R

    Uri M. Ascher and Linda R. Petzold. Computer methods for ordinary differential equations and differential-algebraic equations. 1998. URL: https://api.semanticscholar.org/CorpusID: 32366732

    1998

  6. [6]

    The astec software system

    ASNR. The astec software system. https://research-assessment.asnr.fr/ astec-software-system, 2025

    2025

  7. [7]

    ASSAS – Artificial intelligence for simulation of severe accidents

    ASSAS Consortium. ASSAS – Artificial intelligence for simulation of severe accidents. Horizon Europe Project, coordinated by ASNR, 2023-2026. 14 partners from the European Union, Switzerland and Ukraine. URL:https://assas-horizon-euratom.eu

    2023

  8. [8]

    Layer normalization.arXiv preprint arXiv:1607.06450, 2016

    Jimmy Lei Ba. Layer normalization.arXiv preprint arXiv:1607.06450, 2016

    Pith/arXiv arXiv 2016

  9. [9]

    Joon Young Bae, Chang Hyun Song, JinHo Song, Jeong Ik Lee, Miro Seo, and Sung Joong Kim. Prediction of severe accident progression using machine learning with data-driven surrogate modeling as operator support tool.International Journal of Energy Research, 2026(1), January 2026. URL: http://dx.doi.org/10.1155/er/1416259,doi:10.1155/er/1416259

    work page doi:10.1155/er/1416259 2026

  10. [10]

    Representation equivalent neural operators: a framework for alias-free operator learning

    Francesca Bartolucci, Emmanuel de Bezenac, Bogdan Raonic, Roberto Molinaro, Siddhartha Mishra, and Rima Alaifari. Representation equivalent neural operators: a framework for alias-free operator learning. InThirty-seventh Conference on Neural Information Processing Systems, 2023. URL: https://openreview.net/forum?id=7LSEkvEGCM

    2023

  11. [11]

    Bruinsma, Ana Lucic, Megan Stanley, Anna Allen, Johannes Brandstetter, Patrick Garvan, Maik Riechert, Jonathan A

    Cristian Bodnar, Wessel P. Bruinsma, Ana Lucic, Megan Stanley, Anna Allen, Johannes Brandstetter, Patrick Garvan, Maik Riechert, Jonathan A. Weyn, Haiyu Dong, Jayesh K. Gupta, Kit Thambiratnam, Alexander T. Archibald, Chun-Chieh Wu, Elizabeth Heider, Max Welling, Richard E. Turner, and Paris Perdikaris. A foundation model for the earth system.Nature, 641(...

    doi:10.1038/s41586-025-09005-y 2025

  12. [12]

    Worrall, and Max Welling

    Johannes Brandstetter, Daniel E. Worrall, and Max Welling. Message passing neural PDE solvers. InInternational Conference on Learning Representations, 2022. URL: https://openreview.net/ forum?id=vSix3HPYKSU

    2022

  13. [13]

    Report on the modelling strategy

    Jure Brence, Saso Dzeroski, and Bastien Poubeau. Report on the modelling strategy. Technical report, ASSAS, May 2024. URL: https://assas-horizon-euratom.eu/wp-content/uploads/2024/11/ Report-on-modelling-strategySigned.pdf

    2024

  14. [14]

    Brunton and J

    Steven L. Brunton and J. Nathan Kutz. Promising directions of machine learning for partial differential equations.Nature Computational Science, 4(7):483–494, June 2024. URL: http://dx. doi.org/10.1038/s43588-024-00643-2,doi:10.1038/s43588-024-00643-2

    doi:10.1038/s43588-024-00643-2 2024

  15. [15]

    Brunton, Joshua L

    Steven L. Brunton, Joshua L. Proctor, and J. Nathan Kutz. Discovering governing equations from data by sparse identification of nonlinear dynamical systems.Proceedings of the National Academy of Sciences, 113(15):3932–3937, March 2016. URL: http://dx.doi.org/10.1073/pnas.1517384113, doi:10.1073/pnas.1517384113. 25

    doi:10.1073/pnas.1517384113 2016

  16. [16]

    Data-driven modeling and prediction of non-linearizable dynamics via spectral submanifolds.Nature Com- munications, 13(1), February 2022

    Mattia Cenedese, Joar Ax˚ as, Bastian B¨ auerlein, Kerstin Avila, and George Haller. Data-driven modeling and prediction of non-linearizable dynamics via spectral submanifolds.Nature Com- munications, 13(1), February 2022. URL: http://dx.doi.org/10.1038/s41467-022-28518-y , doi:10.1038/s41467-022-28518-y

    doi:10.1038/s41467-022-28518-y 2022

  17. [17]

    Overview of ASTEC integral code status and perspectives

    Lionel Chailan, Loic Bosland, Laure Car´ enini, Julien Chambarel, Frederic Cousin, et al. Overview of ASTEC integral code status and perspectives. In9th European Review Meeting on Severe Accident Research (ERMSAR2019), Prague, Czech Republic, March 2019.doi:irsn-04106726

    2019

  18. [18]

    Chatelard, Joelle Fleurot, Olivier Marchand, and Patrick Drai

    P. Chatelard, Joelle Fleurot, Olivier Marchand, and Patrick Drai. Assessment of icare/cathare v1 severe accident code. 07 2006.doi:10.1115/ICONE14-89307

    work page doi:10.1115/icone14-89307 2006

  19. [19]

    Chatelard, N

    P. Chatelard, N. Reinke, S. Arndt, S. Belon, L. Cantrel, L. Carenini, K. Chevalier-Jabet, F. Cousin, J. Eckel, F. Jacq, C. Marchetto, C. Mun, and L. Piar. Astec v2 severe accident integral code main features, current v2.0 modelling status, perspectives.Nuclear Engineering and De- sign, 272:119–135, June 2014. URL: http://dx.doi.org/10.1016/j.nucengdes.201...

    work page doi:10.1016/j.nucengdes.2013.06.040 2014

  20. [20]

    NUCLEA - Thermodynamic properties and phase equilibria in nuclear systems.Journal de Physique IV Proceedings, 113:61 – 64, January 2004

    Bertrand Cheynet, Patricia Chaud, Pierre-Yves Chevalier, Evelyne Fischer, Paul Mason, and Mike Mignanelli. NUCLEA - Thermodynamic properties and phase equilibria in nuclear systems.Journal de Physique IV Proceedings, 113:61 – 64, January 2004. URL: https://hal.univ-grenoble-alpes. fr/hal-01895657,doi:10.1051/jp4:20040014

    work page doi:10.1051/jp4:20040014 2004

  21. [21]

    Autoencoder diagrams

    Keenan Crane. Autoencoder diagrams. https://cs.cmu.edu/~kmcrane/AutoencoderDiagrams. zip

  22. [22]

    P. Drai, N. Girault, L. Car´ enini, L. Laborde, J.A. Zambaux, L. Cloarec, E. Delaume, F. Fichot, C. Marchetto, S. M. O. Souvi, O. Coindreau, F. Virot, M. Jobelin, G. Astier, F. Kremer, K. Chevalier- Jabet, L. Bosland, H. Bloch, L. Chailan, R. Monod, M. Johnson, L. Foucher, A. Bleyer, A. Commande, J.J. Ingremeau, and M. Szogradi. ASTEC V3: A Comprehensive ...

    2025

  23. [23]

    Koeberg switchboard, switchboard components and plant ca- bling evaluation for long-term operation

    Eskom Holdings SOC Ltd. Koeberg switchboard, switchboard components and plant ca- bling evaluation for long-term operation. Technical Report 32-T-IPDK-008, Eskom, 2021. URL: https://www.eskom.co.za/wp-content/uploads/2024/04/32-T-IPDK-008_Koeberg_ Switchboard_Switchboard_Components_and_Plant_Cabling_Evaluation_for_LTO.pdf

    2021

  24. [24]

    Testing the manifold hypothesis

    Charles Fefferman, Sanjoy Mitter, and Hariharan Narayanan. Testing the manifold hypothesis. Journal of the American Mathematical Society, 29(4):983–1049, February 2016. URL: http://dx. doi.org/10.1090/jams/852,doi:10.1090/jams/852

    doi:10.1090/jams/852 2016

  25. [25]

    Nicola Rares Franco, Stefania Fresca, Filippo Tombari, and Andrea Manzoni. Deep learning-based surrogate models for parametrized pdes: Handling geometric variability through graph neural networks.Chaos: An Interdisciplinary Journal of Nonlinear Science, 33(12), December 2023. URL: http://dx.doi.org/10.1063/5.0170101,doi:10.1063/5.0170101

    work page doi:10.1063/5.0170101 2023

  26. [26]

    A comprehensive deep learning-based approach to reduced order modeling of nonlinear time-dependent parametrized pdes.Journal of Scientific Computing, 87:1–36, 2021

    Stefania Fresca, Luca Dede’, and Andrea Manzoni. A comprehensive deep learning-based approach to reduced order modeling of nonlinear time-dependent parametrized pdes.Journal of Scientific Computing, 87:1–36, 2021

    2021

  27. [27]

    Learning physics-based reduced-order models from data using nonlinear manifolds.Chaos: An Interdisciplinary Jour- nal of Nonlinear Science, 34(3), March 2024

    Rudy Geelen, Laura Balzano, Stephen Wright, and Karen Willcox. Learning physics-based reduced-order models from data using nonlinear manifolds.Chaos: An Interdisciplinary Jour- nal of Nonlinear Science, 34(3), March 2024. URL: http://dx.doi.org/10.1063/5.0170105, doi:10.1063/5.0170105

    doi:10.1063/5.0170105 2024

  28. [28]

    MIT Press, 2016

    Ian Goodfellow, Yoshua Bengio, and Aaron Courville.Deep Learning. MIT Press, 2016. http: //www.deeplearningbook.org

    2016

  29. [29]

    Vectorized Conditional Neural Fields: A framework for solving time-dependent parametric partial differential equations

    Jan Hagnberger, Marimuthu Kalimuthu, Daniel Musekamp, and Mathias Niepert. Vectorized Conditional Neural Fields: A framework for solving time-dependent parametric partial differential equations. In Ruslan Salakhutdinov, Zico Kolter, Katherine Heller, Adrian Weller, Nuria Oliver, 26 Jonathan Scarlett, and Felix Berkenkamp, editors,Proceedings of the 41st I...

    2024

  30. [30]

    Towards a definition of disentangled representations.arXiv preprint arXiv:1812.02230, 2018

    Irina Higgins, David Amos, David Pfau, Sebastien Racaniere, Loic Matthey, Danilo Rezende, and Alexander Lerchner. Towards a definition of disentangled representations.arXiv preprint arXiv:1812.02230, 2018

    Pith/arXiv arXiv 2018

  31. [31]

    Current State of Knowledge

    Didier Jacquemain, Gerard Cenerino, Francois Corenwinder, Emmanuel IRSN Raimond, Ahmed Bentaib, Herve Bonneville, Bernard Clement, Michel Cranga, Florian Fichot, Vincent Koundy, and et al.Nuclear power reactor core melt accidents. Current State of Knowledge. EDP Sciences, Nov 2015

    2015

  32. [32]

    Solving inverse problems in physics by optimizing a discrete loss: Fast and accurate learning without neural networks.PNAS Nexus, 3(1), December 2023

    Petr Karnakov, Sergey Litvinov, and Petros Koumoutsakos. Solving inverse problems in physics by optimizing a discrete loss: Fast and accurate learning without neural networks.PNAS Nexus, 3(1), December 2023. URL: http://dx.doi.org/10.1093/pnasnexus/pgae005, doi:10.1093/ pnasnexus/pgae005

    doi:10.1093/pnasnexus/pgae005 2023

  33. [33]

    On large-batch training for deep learning: Generalization gap and sharp minima.ArXiv, abs/1609.04836, 2016

    Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima.ArXiv, abs/1609.04836, 2016. URL:https://api.semanticscholar.org/CorpusID:5834589

    Pith/arXiv arXiv 2016

  34. [34]

    Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014

    Pith/arXiv arXiv 2014

  35. [35]

    Space-time continuous PDE forecasting using equivariant neural fields

    David M Knigge, David Wessels, Riccardo Valperga, Samuele Papa, Jan-Jakob Sonke, Erik J Bekkers, and Stratis Gavves. Space-time continuous PDE forecasting using equivariant neural fields. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL: https://openreview.net/forum?id=wN5AgP0DJ0

    2024

  36. [36]

    Neural operator: Learning maps between function spaces with applications to pdes.Journal of Machine Learning Research, 24(89):1–97, 2023

    Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Learning maps between function spaces with applications to pdes.Journal of Machine Learning Research, 24(89):1–97, 2023. URL: http: //jmlr.org/papers/v24/21-1524.html

    2023

  37. [37]

    Jonas Kusch and Pia Stammer. A robust collision source method for rank adaptive dynamical low-rank approximation in radiation therapy.ESAIM: Mathematical Modelling and Numerical Analysis, 57(2):865–891, March 2023. URL: http://dx.doi.org/10.1051/m2an/2022090, doi: 10.1051/m2an/2022090

    work page doi:10.1051/m2an/2022090 2023

  38. [38]

    Carlberg

    Kookjin Lee and Kevin T. Carlberg. Model reduction of dynamical systems on nonlinear manifolds using deep convolutional autoencoders.Journal of Computational Physics, 404:108973, March 2020. URL:http://dx.doi.org/10.1016/j.jcp.2019.108973,doi:10.1016/j.jcp.2019.108973

    doi:10.1016/j.jcp.2019.108973 2020

  39. [39]

    Yeonha Lee, Seok Ho Song, Joon Young Bae, Kyusang Song, Mi Ro Seo, Sung Joong Kim, and Jeong Ik Lee. Surrogate model for predicting severe accident progression in nuclear power plant using deep learning methods and rolling-window forecast.Annals of Nuclear Energy, 208:110816, December 2024. URL: http://dx.doi.org/10.1016/j.anucene.2024.110816, doi:10.1016...

    work page doi:10.1016/j.anucene.2024.110816 2024

  40. [40]

    Maap5 simulation of the pwr severe accident induced by pressurizer safety valve stuck-open accident.Progress in Nuclear Energy, 77:141–151, November 2014

    Longze Li, Yapei Zhang, Wenxi Tian, Guanghui Su, and Suizheng Qiu. Maap5 simulation of the pwr severe accident induced by pressurizer safety valve stuck-open accident.Progress in Nuclear Energy, 77:141–151, November 2014. URL: http://dx.doi.org/10.1016/j.pnucene.2014.06.014, doi:10.1016/j.pnucene.2014.06.014

    work page doi:10.1016/j.pnucene.2014.06.014 2014

  41. [41]

    Fourier neural operator with learned deformations for pdes on general geometries.Journal of Machine Learning Research, 24(388):1–26, 2023

    Zongyi Li, Daniel Zhengyu Huang, Burigede Liu, and Anima Anandkumar. Fourier neural operator with learned deformations for pdes on general geometries.Journal of Machine Learning Research, 24(388):1–26, 2023. URL:http://jmlr.org/papers/v24/23-0064.html

    2023

  42. [42]

    Fourier neural operator for parametric partial differential equations

    Zongyi Li, Nikola Borislavov Kovachki, Kamyar Azizzadenesheli, Burigede liu, Kaushik Bhat- tacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. InInternational Conference on Learning Representations, 2021. URL: https://openreview.net/forum?id=c8P9NQVtmnO. 27

    2021

  43. [43]

    Latent space modeling of parametric and time-dependent pdes using neural odes.Computer Methods in Applied Mechanics and Engineering, 448:118394, January 2026

    Alessandro Longhi, Danny Lathouwers, and Zolt´ an Perk´ o. Latent space modeling of parametric and time-dependent pdes using neural odes.Computer Methods in Applied Mechanics and Engineering, 448:118394, January 2026. URL: http://dx.doi.org/10.1016/j.cma.2025.118394, doi:10.1016/ j.cma.2025.118394

    work page doi:10.1016/j.cma.2025.118394 2026

  44. [44]

    J. L. Lumley. The structure of inhomogeneous turbulence. In A. M. Yaglom and V. I. Tatarski, editors,Atmospheric Turbulence and Wave Propagation. Nauka, Moscow, 1967

    1967

  45. [45]

    Andrea Malizia, Andrea Chierici, Sergio Biancotto, Marco D’Arienzo, Gian Marco Ludovici, Francesco d’Errico, Guglielmo Manenti, and Fabio Marturano. The hotspot code as a tool to improve risk analysis during emergencies: Predicting i-131 and cs-137 dispersion in the fukushima nuclear accident.International Journal of Safety and Security Engineering, 11(4)...

    work page doi:10.18280/ijsse.110421 2021

  46. [46]

    A brief review of reduced order models using intrusive and non-intrusive techniques.PAMM, 24(4), November 2024

    Guglielmo Padula, Michele Girfoglio, and Gianlugi Rozza. A brief review of reduced order models using intrusive and non-intrusive techniques.PAMM, 24(4), November 2024. URL: http://dx.doi. org/10.1002/pamm.202400210,doi:10.1002/pamm.202400210

    doi:10.1002/pamm.202400210 2024

  47. [47]

    Users’ expectations – simulator specifications

    Isabel Parrado-Rodriguez. Users’ expectations – simulator specifications. WP6 Report Version 1, ASSAS, February 2024

    2024

  48. [48]

    Film: Visual reasoning with a general conditioning layer

    Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. InProceedings of the AAAI conference on artificial intelligence, volume 32, 2018

    2018

  49. [49]

    Radaideh, Connor Pigg, Tomasz Kozlowski, Yujia Deng, and Annie Qu

    Majdi I. Radaideh, Connor Pigg, Tomasz Kozlowski, Yujia Deng, and Annie Qu. Neural-based time series forecasting of loss of coolant accidents in nuclear power plants.Expert Systems with Applications, 160:113699, December 2020. URL: http://dx.doi.org/10.1016/j.eswa.2020.113699, doi:10. 1016/j.eswa.2020.113699

    work page doi:10.1016/j.eswa.2020.113699 2020

  50. [50]

    Raissi, P

    M. Raissi, P. Perdikaris, and 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, 378:686–707, February 2019. URL: http://dx.doi.org/10. 1016/j.jcp.2018.10.045,doi:10.1016/j.jcp.2018.10.045

    doi:10.1016/j.jcp.2018.10.045 2019

  51. [51]

    Convolutional neural operators for robust and accurate learning of pdes.Advances in Neural Information Processing Systems, 36:77187–77200, 2023

    Bogdan Raonic, Roberto Molinaro, Tim De Ryck, Tobias Rohner, Francesca Bartolucci, Rima Alaifari, Siddhartha Mishra, and Emmanuel de B´ ezenac. Convolutional neural operators for robust and accurate learning of pdes.Advances in Neural Information Processing Systems, 36:77187–77200, 2023

    2023

  52. [52]

    Tarabelli, G

    D. Tarabelli, G. Ratel, R. P´ elisson, G. Guillard, M. Barnak, and P. Matejovic. Astec application to in-vessel corium retention.Nuclear Engineering and Design, 239(7):1345–1353, 2009. URL: http: //dx.doi.org/10.1016/j.nucengdes.2009.02.021,doi:10.1016/j.nucengdes.2009.02.021

    work page doi:10.1016/j.nucengdes.2009.02.021 2009

  53. [53]

    Organized by ASNR (Autorit´ e de Sˆ uret´ e Nucl´ eaire et de Radioprotection)

    Training Course, Aix-en-Provence, France.ASTEC (Accident Source Term Evaluation Code) Software: Models, Numerical Structure, Data Management, January 23–28 2023. Organized by ASNR (Autorit´ e de Sˆ uret´ e Nucl´ eaire et de Radioprotection)

    2023

  54. [54]

    U.S. NRC. TRACE v5.840 Theory Manual: Field Equations, Solution Methods, and Physical Models. Theory manual, US Nuclear Regulatory Commission, Washington, D.C., United States, 2013

    2013

  55. [55]

    Oliphant, Matt Haberland, Tyler Reddy, David Cournapeau, Evgeni Burovski, Pearu Peterson, Warren Weckesser, Jonathan Bright, St´ efan J

    Pauli Virtanen, Ralf Gommers, Travis E. Oliphant, Matt Haberland, Tyler Reddy, David Cournapeau, Evgeni Burovski, Pearu Peterson, Warren Weckesser, Jonathan Bright, St´ efan J. van der Walt, Matthew Brett, Joshua Wilson, K. Jarrod Millman, Nikolay Mayorov, Andrew R. J. Nelson, Eric Jones, Robert Kern, Eric Larson, C J Carey, ˙Ilhan Polat, Yu Feng, Eric W....

    doi:10.1038/s41592-019-0686-2 2020

  56. [56]

    Wiewel, M

    S. Wiewel, M. Becher, and N. Thuerey. Latent space physics: Towards learning the temporal evolution of fluid flow.Computer Graphics Forum, 38(2):71–82, May 2019. URL: http://dx.doi. org/10.1111/cgf.13620,doi:10.1111/cgf.13620. 28

    doi:10.1111/cgf.13620 2019

  57. [57]

    Group normalization

    Yuxin Wu and Kaiming He. Group normalization. InComputer Vision – ECCV 2018: 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part XIII, page 3–19, Berlin, Heidelberg, 2018. Springer-Verlag.doi:10.1007/978-3-030-01261-8_1. 29 Table 3: Variables belonging tos g,s p,s B1 ands B2 State space Variables sg H2 cumulated mass in the ...

    doi:10.1007/978-3-030-01261-8_1 2018

  58. [58]

    The idea behind this is that the latent spaceEneeds to be shaped before the dynamics can be approximated

    we choose a low value for l = l0, typically l0 = 25, and we only train the AutoEncoder for a number NAE of epochs on Ml0 tr, i.e., we set δ = ν = µ = ω = 0.0 from Ltot. The idea behind this is that the latent spaceEneeds to be shaped before the dynamics can be approximated

  59. [59]

    This is done for NT F epochs

    we keep Ml0 tr and we switch on the LT F and Ltm losses, while keeping ν = ω = 0, i.e., we do not train yet autoregressively. This is done for NT F epochs. The reason why we do so is to start building upf θ in the easiest possible setting; 33 Table 8: Hyper-parameters of AR training NAE NT F L N l LOCA 25 10 [10,1000] [25,25] SBO 25 10 [10,1000] [25,25]

  60. [60]

    We wait Nl0 epochs and after that we reshape the dataset into Ml1 tr, where l1 > l 0 and we train on Ml1 tr for Nl1 epochs

    we start with Ml0 tr and we switch on also the AR losses LAR and Lf ull AR . We wait Nl0 epochs and after that we reshape the dataset into Ml1 tr, where l1 > l 0 and we train on Ml1 tr for Nl1 epochs. We repeat this process up to a certain li with the waiting times and each li pre-defined before the training starts. Thus, prior to the training, we have to...

  61. [61]

    The latent vector is concatenated with the flattened vector p(ˆx, tj i |kj), resulting in a concatenated vector of dimensionλ+d h1 +d c1

    simple concatenation. The latent vector is concatenated with the flattened vector p(ˆx, tj i |kj), resulting in a concatenated vector of dimensionλ+d h1 +d c1

  62. [62]

    FiLM [48] application. The latent vector is transformed as ε→β f ilm θ (p) ⊙ε + δf ilm θ (p), where βf ilm θ : Sp →R λ and δf ilm θ : Sp →R λ are two linear transformations parametrized by weights found during training.⊙is the point-wise (Hadamard product). We did not find noticeable difference in the application of the two, so for simplicity we use conca...

  63. [63]

    RMSE j m = 1 Fj −1 PFj −1 i=1 ||˜s(xm,ti+1|kj)−sm(xm,ti+1|kj)||2 2 |Ωm| 1 2 ∈R dm

  64. [64]

    RMSE j m,mean = RMSEj m Mean{sm(xm,ti|kj)} Fj i=1 ∈R dm

  65. [65]

    RMSE j m,max = RMSEj m Max{sm(xm,ti|kj)} Fj i=1 ∈R dm

  66. [66]

    So each metric is a vector containing dm errors, each associated to a variable belonging to sm

    RMSE j m,std = RMSEj m Std{sm(xm,ti|kj)} Fj i=1 ∈R dm where the index m∈ {g, p, cr, v, B 1, B2} as defined in Subsection 2.2 and ˜s(xm, ti+1|kj) can be either ˜sm i+1,1(xm, ti+1|kj), i.e., the actual prediction of AE-NODE, or ψ◦φ (sm(xm, ti+1|kj)), i.e., a simple autoencoding, as done in Figures 15 and 9. So each metric is a vector containing dm errors, e...