{"id":"2b9297e3-efc0-4c47-8ca8-73479efa19bd","arxiv_id":"2611.16511","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A recurrent Fourier neural operator trained on coarse hydrodynamic grids can predict fine-grid temperature profiles and flow anisotropies without retraining and can continue predictions beyond the training time interval.","lead":"This paper develops a recurrent Fourier neural operator (rFNO) that emulates (2+1)-dimensional relativistic hydrodynamics, achieving zero-shot super-resolution: models trained on coarse grids can predict fine-grid profiles without retraining. The rFNO could accelerate event-by-event hydrodynamic simulations by 6-8x, which would speed up Bayesian parameter estimation for the quark-gluon plasma.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"The demonstration of zero-shot super-resolution and recurrent continuation is based entirely on test events from the same T_RENTo initial condition model as training; generalization to other initial condition distributions remains untested, which is load-bearing for the practical reach of the claim.","rationale":"The reader identified boost invariance and ideal/conformal EOS as the weakest assumptions, which are acknowledged by the paper. My concern moves one level deeper: even within the (2+1)-d ideal hydro framework with conformal EOS, the rFNO's success may be contingent on the specific distribution of initial conditions used during training. The paper tests only one initial condition family (T_RENTo with a single parameter set), so it does not establish that the learned operator generalizes across different initial condition distributions. This is load-bearing because the practical value of the method for event-by-event simulations in Bayesian analyses depends on its ability to handle varied initial conditions. The concrete test I propose directly probes this generalization. If the test passes (errors remain within about 10%), the claim is strengthened and the concern resolved; the verdict could be upgraded to ACCEPT. If the test fails, the claim must be qualified. Therefore I recommend a CONDITIONAL verdict: accept conditional on passing this generalization test or on an explicit scoping statement limiting the claim to the T_RENTo initial conditions studied. My agreement with the reader is partial because we both identify scope limitations, but I focus on a different aspect (initial condition distribution) rather than the underlying physics assumptions (boost invariance, EOS).","tokens_in":14594,"tokens_out":7698,"duration_ms":92297,"concrete_test":"Train or select 100 initial conditions from a qualitatively different model (e.g., MC-Glauber with slightly deformed nuclei, or T_RENTo with a different normalization width p parameter) for Au+Au at sqrt(s)=200 GeV, 0-10% centrality. Run VISH2+1 on these with the same settings to produce ground-truth profiles on the 128x128 grid. Evaluate the pre-trained 32x32 and 64x64 rFNOs (without retuning) on these 128x128 initial conditions. Compute the event-by-event epsilon_2, epsilon_3, epsilon_4, and epsilon_p at tau-tau0 = 4.8 fm/c and 6.4 fm/c. If the median relative error for any of these integrated quantities exceeds 20%, that would indicate the super-resolution capability does not generalize across initial condition models, constraining the scope of the claim.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that the rFNO 'realizes zero-shot super-resolution predictions for relativistic hydrodynamics' is supported by quantitative comparisons on a test set drawn from the same T_RENTo model used for training. The paper does not evaluate whether the rFNO can maintain its accuracy when applied to initial conditions with different statistical properties—for example, from MC-Glauber, or from T_RENTo with different parameters, or from collisions at a different energy or centrality. Because the rFNO learns mappings conditioned on the training distribution, it may exploit specific patterns in the initial conditions (e.g., smoothness scale, fluctuation spectrum) that are not universal. If the super-resolution and recurrent continuation fail for a distinct initial condition model, the claim that the method works 'for relativistic hydrodynamics' would be weakened, and its utility for accelerating Bayesian analyses that vary initial condition parameters would be reduced. The paper states that the T_RENTo parameters are 'tuned close to the ones extracted from the Bayesian analyses for Pb+Pb at the LHC,' but applies them to Au+Au at a different energy, so the training data themselves depart from the exact posterior. This makes the distributional specificity a load-bearing concern: the reported accuracy may be an artifact of the particular statistical structure of the training set rather than a property of the learned operator.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper develops a recurrent Fourier neural operator (rFNO) that emulates (2+1)-dimensional relativistic ideal hydrodynamics event by event. Three rFNOs are trained on 32×32, 64×64, and 128×128 grids using profiles from the VISH2+1 code with T_RENTo initial conditions for Au+Au at √sNN=200 GeV. The two coarser models are then evaluated on a 128×128 grid without retraining (zero-shot super-resolution), and all models are recurrently continued beyond the supervised interval (τ−τ0=4.8 fm/c) to τ−τ0=6.4 fm/c. Comparisons with VISH2+1 show that the rFNO predictions accurately reproduce temperature profiles, event-by-event eccentricities ε_n, and momentum anisotropy ε_p within the supervised interval, with relative errors for integrated quantities remaining within about 10% after recurrent continuation. The paper concludes that the rFNO achieves zero-shot super-resolution and can be recurrently continued, demonstrating potential to accelerate realistic event-by-event hydrodynamic simulations.","tokens_in":14993,"tokens_out":2281,"duration_ms":21982,"significance":"If the results hold, the paper provides a practical neural operator emulator that can accelerate the computationally expensive step of relativistic hydrodynamic evolution in heavy-ion collisions. The zero-shot super-resolution ability means that a model trained on a coarse grid can directly produce accurate predictions on a finer grid, which is valuable for multi-scale simulations. The recurrent continuation beyond the trained time window further extends applicability. The paper is clearly written, the experiments are well designed, and the quantitative comparisons with VISH2+1 ground truth (on an independent test set) give confidence in the approach. The work builds on prior FNO applications to hydrodynamics and advances the state of the art by demonstrating both super-resolution and recurrence together.","major_comments":[],"minor_comments":[{"comment":"The claim that the rFNO 'realizes zero-shot super-resolution predictions for relativistic hydrodynamics' should be qualified to reflect that the demonstration is for one specific initial condition model (T_RENTo) and collision system (Au+Au at √sNN=200 GeV, 0–10% centrality). The paper does not test generalization to other initial condition models (e.g., MC-Glauber) or collision energies, which limits the scope of the claim as stated. Please add an explicit discussion of this limitation in the conclusion.","section":"Abstract and §IV","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The paper uses ideal hydrodynamics (P=e/3) and the assumption of longitudinal boost invariance. These are clearly stated but are important for interpreting the results. I recommend adding a brief remark in the conclusion that generalizing to viscous hydrodynamics and (3+1)-D evolution is left for future work.","section":"§II.A","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The caption of Table II and the surrounding text report δ_E and δ_P deviations. The smaller δ_E for the 128×128 rFNO at τ−τ0=6.4 fm/c is explained as a cancellation of errors. This is acceptable, but it would be helpful to provide the standard deviation or range of δ_E across events to indicate whether this cancellation is robust or accidental.","section":"Table II and §III","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The temporal reflection construction (Eq. 6) is used to reduce temporal boundary mismatch. I suggest briefly commenting on how much the temporal reflection improves the prediction accuracy, perhaps with a one-sentence reference to an ablation test if available.","section":"§II.B","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Minor typographical: In §II.A, the definition of ε_n in Eq. (3) uses 'r dr dϕ' which should be 'r dr dϕ' (it is fine). Also, in Eq. (4), the transverse momentum anisotropy definition uses T^{xx} and T^{yy} but the text in the following line uses 'T_{xx}, T_{yy}' – these are the same, but please ensure consistent notation.","section":"General","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"minor_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Colleague,\n\nThis is a clean, well-executed paper that builds a recurrent Fourier neural operator (rFNO) to emulate (2+1)-d ideal hydrodynamics. The main new things are: (i) a recurrent architecture with a gated predictor that propagates profiles interval by interval, (ii) a demonstration of zero-shot super-resolution—models trained on 32×32 or 64×64 grids evaluated on 128×128 without retraining—and (iii) recurrent continuation beyond the supervised time window with relative errors on ε_n and ε_p staying around 10%.\n\nThe experiments are carefully done. The comparisons to VISH2+1 cover temperature profiles, eccentricities, momentum anisotropy, and energy conservation. The speedup estimate (6–8× over GPU-accelerated VISH2+1) is plausible and relevant. The paper is clearly written and the limitations are stated: ideal hydro, conformal EOS, boost invariance, a single initial-condition model (T_RENTo).\n\nNow the soft spots, in proportion. The stress-test concern about distributional generality is real and load-bearing for the practical claim. The zero-shot super-resolution and recurrent continuation are demonstrated only on test events drawn from the same T_RENTo model as training. The rFNO may exploit specific patterns in the initial-condition fluctuations (smoothness scale, spectrum) that are not universal. If the method fails on MC-Glauber or on T_RENTo with different parameters, the claim that it works \"for relativistic hydrodynamics\" is weakened, and its utility for Bayesian parameter estimation—which varies initial-condition parameters—is reduced. The paper acknowledges using T_RENTo parameters tuned to LHC Pb+Pb but applied to Au+Au at RHIC, which is a mild distribution shift, but that is not the same as testing on a genuinely different initial-condition model. This is a missing experiment, not a flaw in what they did.\n\nA second, minor point: the paper uses ideal hydro with a conformal EOS. The authors say this is for simplicity, and they are upfront about it. But the practical relevance for heavy-ion phenomenology is limited until the method is extended to viscous hydro with a realistic EOS. That is future work, not a flaw.\n\nThe paper does not release code or data, but the description is detailed enough for reimplementation.\n\nWho is this for? Researchers working on surrogate models for heavy-ion simulations, especially those interested in operator learning for PDEs. The paper deserves a serious referee—the claims are specific, the experiments are reproducible in principle, and the method is a genuine step forward within the FNO line. I would accept a peer-review invitation for this.\n\nVerdict: accept with minor revision. The main revision should be to add a discussion (and ideally a test) of generalization to other initial-condition models, or at least to state the distributional limitation more prominently.","headline":"Solid operator-learning paper for (2+1)-d hydro; the zero-shot super-resolution claim is real but the distributional generality is untested, which matters for Bayesian workflows.","tokens_in":15591,"tokens_out":698,"duration_ms":7708},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"A recurrent Fourier neural operator learns relativistic hydrodynamics with zero-shot super-resolution and temporal extrapolation.","keywords":["recurrent Fourier neural operator","relativistic hydrodynamics","quark-gluon plasma","zero-shot super-resolution","event-by-event simulation","operator learning","heavy-ion collisions","neural emulator"],"falsifier":"Extend the training data to include viscous hydrodynamics with a lattice-QCD equation of state and test the rFNO on events with large transverse momentum gradients or at forward rapidity; if relative errors exceed 20% for integrated quantities, then the current ideal-boost-invariant assumption is essential and the method's range is narrower than claimed.","tokens_in":16072,"feed_emoji":"⚛","tokens_out":4041,"duration_ms":34926,"temperature":0.7,"pith_summary":"The paper develops a recurrent Fourier neural operator (rFNO) that emulates (2+1)-dimensional relativistic hydrodynamics for quark-gluon plasma evolution. Trained on coarse grids (32x32, 64x64), the rFNO directly predicts accurate temperature profiles on a fine 128x128 grid without retraining, achieving zero-shot super-resolution. It can also be recurrently continued beyond the trained time window, maintaining relative errors within about 10% for event-by-event eccentricities and momentum anisotropy. This capability could substantially accelerate Bayesian parameter estimation and event-by-event simulations in heavy-ion physics.","feed_headline":"Neural network emulates quark-gluon fluid with zero-shot super-resolution","feed_subtitle":"Trained on coarse grids, it predicts fine-grid temperature and flow without retraining, and extrapolates in time within 10% error.","key_machinery":"The key machinery is the recurrent Fourier neural operator (rFNO), which maps the hydrodynamic state at one proper time to a sequence of profiles over a finite interval via a three-dimensional FNO acting on (τ, x, y). The architecture includes a learned soft mask that suppresses predictions in the dilute exterior region, and a progressive recurrent training curriculum that gradually increases the number of rollout steps. The Fourier layers retain transverse modes up to a fixed cutoff; during cross-resolution evaluation the same Fourier weights operate on a finer grid, enabling zero-shot super-resolution.","core_discovery":"The central discovery is that a recurrent Fourier neural operator trained on coarse spatial grids learns a discretization-invariant representation of relativistic hydrodynamic evolution. When evaluated on a finer grid than seen during training, the network produces temperature profiles, spatial eccentricities ε_n (n=2,3,4), and momentum anisotropy ε_p that agree closely with the VISH2+1 ground truth, without any fine-tuning. Furthermore, the recurrent architecture allows the model to propagate predictions beyond the supervised proper-time interval (τ-τ0=4.8 fm/c) to 6.4 fm/c, with relative errors staying below about 10% for the integrated quantities. This demonstrates that operator learning can capture both spatial super-resolution and temporal extrapolation for a nonlinear hyperbolic PDE system in a physically meaningful regime.","pith_inferences":["The demonstrated discretization invariance suggests that similar rFNO architectures could be applied to other nonlinear PDEs in physics, such as magnetohydrodynamics or cosmological structure formation, where multi-resolution modeling is beneficial.","The use of an ideal equation of state (P=e/3) and longitudinal boost invariance limits the direct applicability to realistic LHC or RHIC collisions; extending to viscous hydrodynamics with a lattice-QCD equation of state would test whether the rFNO retains its accuracy in more complex regimes.","The progressive training curriculum may be essential for the recurrent stability; it would be informative to ablate this curriculum to see whether direct multi-step training leads to divergence."],"forward_implications":["The rFNO can accelerate event-by-event hydrodynamic simulations by a factor of 6–8 compared to a GPU-accelerated VISH2+1, and by a factor of 400 compared to a single-threaded CPU implementation.","Zero-shot super-resolution means that expensive fine-grid simulations can be emulated using models trained on cheap coarse grids, reducing the computational cost of generating training data.","The recurrent continuation capability allows modeling longer time evolutions without retraining, which is useful for studying late-time observables such as Hanbury Brown–Twiss correlations or freeze-out dynamics.","The accuracy of integrated quantities (ε_n, ε_p) suggests that rFNO can be used as a fast surrogate in Bayesian parameter estimation pipelines, where many forward evaluations are needed.","The approach is complementary to generative models that produce final-state hadron spectra; combining them could yield end-to-end fast emulators for heavy-ion collisions."],"supporting_citations":[{"why":"Introduces the Fourier neural operator that the rFNO extends with recurrence and gating.","marker":"[34]"},{"why":"Provides the theoretical foundation for learning maps between function spaces with discretization invariance.","marker":"[35]"},{"why":"The VISH2+1 code generates the ground-truth hydrodynamic profiles used for training and evaluation.","marker":"[39]"},{"why":"The T_RENTo model generates the initial conditions for the hydrodynamics events.","marker":"[41]"},{"why":"Inspires the progressive recurrent training curriculum that stabilizes multi-step rollouts.","marker":"[43]"},{"why":"First application of FNO to QGP hydrodynamics, but without recurrence or super-resolution; the present work improves upon it.","marker":"[38]"},{"why":"Motivates the need for fast emulators by describing Bayesian parameter estimation for heavy-ion collisions.","marker":"[16]"}],"fun_headline_variants":["Zero-shot super-resolution for quark-gluon plasma via recurrent Fourier net","Coarse-trained neural operator predicts fine-grid hydro without retraining","Recurrent Fourier operator extrapolates QGP flow within 10% error","Grid-invariant learning achieves super-res for relativistic hydrodynamics"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The evolution is assumed to be longitudinally boost-invariant and uses an ideal conformal equation of state (P=e/3); both are approximations that limit applicability to midrapidity and to the regime where viscous effects are negligible.","fun_headline_variants_meta":{"raw":{"variants":["Zero-shot super-resolution for quark-gluon plasma via recurrent Fourier net","Coarse-trained neural operator predicts fine-grid hydro without retraining","Recurrent Fourier operator extrapolates QGP flow within 10% error","Grid-invariant learning achieves super-res for relativistic hydrodynamics"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.00027,"raw_usage":{"cost":0.000270368,"is_byok":false,"cost_details":{"upstream_inference_cost":0.000270368,"upstream_inference_prompt_cost":0.000111888,"upstream_inference_completions_cost":0.00015848},"total_tokens":1570,"prompt_tokens":1004,"completion_tokens":566,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":490}},"tokens_in":1004,"tokens_out":566,"cost_source":"openrouter_usage","duration_ms":5632,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":490,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-11-13T14:43:55.612446+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Extend the training data to include viscous hydrodynamics with a lattice-QCD equation of state and test the rFNO on events with large transverse momentum gradients or at forward rapidity; if relative errors exceed 20% for integrated quantities, then the current ideal-boost-invariant assumption is essential and the method's range is narrower than claimed.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the Fourier neural operator that the rFNO extends with recurrence and gating."},{"cited_title":"Neural operator: Learning maps between function spaces with applications to PDEs, Journal of Machine Learning Re- search24, 1","cited_arxiv_id":null,"evidence_quote":"Provides the theoretical foundation for learning maps between function spaces with discretization invariance."},{"cited_title":"Causal viscous hydrodynamics in 2+1 dimensions for relativistic heavy-ion collisions","cited_arxiv_id":"0712.3715","evidence_quote":"The VISH2+1 code generates the ground-truth hydrodynamic profiles used for training and evaluation."},{"cited_title":"Alternative ansatz to wounded nucleon and binary collision scaling in high-energy nuclear collisions","cited_arxiv_id":"1412.4708","evidence_quote":"The T_RENTo model generates the initial conditions for the hydrodynamics events."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Inspires the progressive recurrent training curriculum that stabilizes multi-step rollouts."},{"cited_title":"Fast prediction of the hydrodynamic QGP evolution in ultra-relativistic heavy-ion collisions using Fourier Neural Operators","cited_arxiv_id":null,"evidence_quote":"First application of FNO to QGP hydrodynamics, but without recurrence or super-resolution; the present work improves upon it."},{"cited_title":"Applying Bayesian parameter estimation to relativistic heavy-ion collisions: simultaneous characterization of the initial state and quark-gluon plasma medium","cited_arxiv_id":"1605.03954","evidence_quote":"Motivates the need for fast emulators by describing Bayesian parameter estimation for heavy-ion collisions."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}