Recurrent Fourier Neural Operator for Event-by-Event Relativistic Hydrodynamics
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Nuclear Theory. The day in nuclear theory: structure, reactions and the models behind them.
Kai: We're looking at the paper: "Recurrent Fourier Neural Operator for Event-by-Event Relativistic Hydrodynamics".
Mira: Here's the gist of the paper. Relativistic hydrodynamics simulations are computationally demanding when used to constrain model parameters through Bayesian analyses of experimental data The recurrent Fourier neural operator (rFNO) realizes zero-shot super-resolution predictions for relativistic hydrodynamics and can be recurrently continued beyond the supervised interval Model Development and Training Setup The rFNO is developed to emulate the evolution of (2 + 1)-dimensional relativistic hydrodynamics Three networks, referred to as the 32 × 32, 64 × 64, and 128 × 128 rFNOs, are trained on the corresponding lattices using the same event-by-event hydrodynamic profiles The networks trained on the coarser lattices are then directly evaluated on the 128×32 lattice without retraining, while the 128×32 rFNO provides the native-resolution reference.
Kai: Recurrent Propagation and Resolution Transfer The network maps the hydrodynamic state at one proper time to a sequence of transverse profiles over a finite interval The input state is replicated along the temporal direction and processed by a threedimensional FNO acting on (τ, x, y) with four Fourier layers and hidden width 64.
Mira: A learned output gate suppresses small-amplitude predictions in the dilute exterior region During crossresolution evaluation, both the Fourier weights and the learned model parameters are kept unchanged.
Kai: Training Curriculum and Loss Function The training curriculum is structured in three stages to account for errors accumulated during recurrent evolution Stage I involves τ0→ I1, τ1→ I2, and τ2→ I3.
Paper summary: Kai: So we're looking at this paper today which is titled "Recurrent Fourier Neural Operator for Event-by-Event Relativistic Hydrodynamics." It sounds like they are tackling the computational headache of running full relativistic hydrodynamics simulations when you need to do lots of Bayesian analyses with experimental data.
Mira: Exactly, it seems the core idea here is developing a recurrent Fourier neural operator, or rFNO, that can handle this evolution without needing to re-run the whole expensive simulation every single time. The paper claims this rFNO can perform zero-shot super-resolution predictions for relativistic hydrodynamics and even continue its prediction beyond the supervised time interval.
Lev: That sounds promising because running those full event-by-event simulations is just too slow for what we need in many experimental contexts, especially when you're trying to constrain model parameters. The training setup involves three networks, the thirty-two by thirty-two sixty-four by sixty-four and one hundred twenty-eight by one hundred twenty-eight rFNOs on different lattices using event-by-event hydrodynamic profiles from VISH2+one up to a proper time of four point eight femtoseconds <ref:2611.16511#pg1>.
Kai: And what's interesting about the training is that they use the same event-by-event profiles for all those different lattice sizes. The networks trained on the coarser lattices are then evaluated directly on the one hundred twenty-eight by one hundred twenty-eight lattice without any retraining, while that one hundred twenty-eight by thirty-two rFNO is used as a kind of native resolution reference <ref:2611.16511#pg1,The networks trained on the coarser lattices are then>.
Mira: The methodology involves mapping the hydrodynamic state at one proper time to a sequence of transverse profiles over a finite interval. They feed the input state into this three-dimensional Fourier neural operator acting on the proper time, space, and two spatial dimensions with four Fourier layers and a hidden width of sixty-four.
Lev: And they add something called a learned output gate that suppresses small amplitude predictions in the dilute exterior region during crossresolution evaluation. The authors keep both the Fourier weights and the learned model parameters unchanged when doing this crossresolution part.
Kai: The training curriculum is structured in three stages to handle errors accumulated during recurrent evolution, moving from time zero to one point eight femtoseconds, then one point eight femtoseconds to two point six femtoseconds, and finally two point six femtoseconds to three point four femtoseconds. They use the element-wise mean squared error over aligned predictions and targets for the loss function.
Mira: They also use an auxiliary binary cross-entropy loss on a learned gate mask weighted by alpha equals zero point zero one. This suggests they're trying to learn *how* to suppress those small errors in that exterior region too. The performance claims show that the rFNO predictions capture the magnitude and structures of the temperature profiles, as well as event-by-event eccentricities epsilon n, momentum anisotropy epsilon p, and their time evolution within that supervised proper time interval.
Lev: And they achieve zero-shot super-resolution, which means it can predict finer details than what it was directly trained on. They also mentioned that after recurrent continuation to four point eight femtoseconds minus the initial time zero point eight femtoseconds, which is six point four femtoseconds, the relative errors of those epsilon n and epsilon p remain within about ten percent.
Kai: That's quite a bit of detail for a prediction beyond what it was directly trained on. And they also address conservation properties by saying that the rFNO values for delta E and delta P are of the same order as those from VISH2+one where the nonzero delta E mainly reflects different time discretizations used for the longitudinal pressure source and evaluation of WL <ref:2611.16511#pg1>.
Mira: The authors note that because of this, the smaller delta E seen in their one hundred twenty-eight by thirty-two rFNO at six point four femtoseconds results from a partial cancellation between its prediction error and that discretization error, rather than from better energy conservation <ref:2611.16511#pg1>.
Lev: From an engineering standpoint, the computational efficiency is where it really shines. The rFNO predictions are much faster than the VISH2+one simulations <ref:2611.16511#pg1>. For instance, the prediction times per event for the thirty-two by thirty-two rFNO are about fifty milliseconds, while for the sixty-four by sixty-four and one hundred twenty-eight by thirty-two they are about fifty five and sixty four milliseconds respectively <ref:2611.16511#pg1>.
Kai: If we look at scaling up to the six forty by sixty four rFNO on that twelve eighty by thirty lattice, they expect that to further reduce the prediction time by a factor of four. That's significant speedup for event-by-event work.
Mira: The conclusion is that the rFNO could make reliable super-resolution predictions and it can be recurrently continued beyond the supervised interval, which shows its potential to substantially accelerate realistic event-by-event hydrodynamic simulations. The time resolved profiles predicted by this rFNO are complementary to recent generative diffusion and flow matching models that emulate final state hadron spectra of hybrid hydrodynamic simulations end to end.
Lev: It seems like the real implication here is in speeding up those realistic event-by-event hydrodynamic simulations, which is a big deal for the kind of detailed analysis we need when we're trying to constrain model parameters with experimental data. The paper itself, "Recurrent Fourier Neural Operator for Event-by-Event Relativistic Hydrodynamics," shows how this recurrent approach works and its potential for faster simulation.
Conclusion: Kai: So we’re wrapping up our look at this paper on "Recurrent Fourier Neural Operator for Event-by-Event Relativistic Hydrodynamics." Basically, they built this rFNO that can take a simulation state and predict the next part of its evolution without running the whole thing over again.
Mira: Yeah, it seems to be about using these neural operators to solve the problem of making those big relativistic hydrodynamic simulations actually faster when you need them for experimental data analysis.
Lev: From an error correction standpoint, that recurrence is a huge deal because it means you're not just doing one step; you're chaining predictions together, which is where the real overhead usually lands on real hardware.
Kai: It’s interesting how they trained these networks on different grid sizes—thirty-two by thirty-two sixty-four by sixty-four and one hundred twenty-eight by one hundred twenty-eight—and then they can use the smaller ones to predict on a bigger one like twelve eighty by thirty without retraining.
Mira: That suggests the learned mapping is quite universal across those resolutions, which is what makes it powerful for super-resolution predictions.
Lev: But they did mention that for conservation properties, things like energy and momentum are similar to what you get from VISH2 plus one simulations, though there’s a bit of cancellation in the error terms at later times.
Kai: So the main point is that this method can give them these super-resolved predictions for temperature profiles and other things like eccentricities and momentum anisotropy within a specific time window.
Mira: It really changes how we might look at simulating those complex systems because it offers a path toward making event-by-event work much more feasible.
Lev: And the efficiency gains are noticeable; they’re talking about prediction times that are much faster than what you’d get from running those traditional simulations.
Kai: It looks like this could be a big help for experimentalists who need these detailed, time-resolved profiles to constrain their models with real data.
Mira: It opens up a new way to look at how we handle the complexity of relativistic fluid dynamics in practice, especially when you’re dealing with high-energy collisions.
Lev: The real question now is how robust this recurrence is if you try to push it beyond that supervised time interval they tested.
Xiangrun Bao, *, Lingxiao Wang, •, Huichao Song
School of Physics, Peking University · RIKEN Center for Interdisciplinary Theoretical and Mathematical Sciences (iTHEMS) · Institute for Physics of Intelligence, Graduate School of Science, The University of Tokyo · Center for High Energy Physics, Peking University
nucl-th, hep-ph
Submitted: 2026-11-12
Updated: 2026-11-12
Comments: 9 pages, 4 figures
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 98/100
The gist: Relativistic hydrodynamics simulations are computationally demanding when used to constrain model parameters through Bayesian analyses of experimental data The recurrent Fourier neural operator…
Key concepts
- Recurrent Fourier Neural Operator (rFNO)
- An AI model designed to emulate the time evolution of (2+1)-dimensional relativistic hydrodynamics. It maps a state at one proper time to a sequence of transverse profiles over a finite interval. It uses Fourier layers and a learned output gate, allowing it to predict hydrodynamic states efficiently.
- Zero-shot Super-resolution
- The ability of the rFNO to generate high-resolution predictions for hydrodynamic profiles without needing retraining on those specific resolutions. The network is trained on coarser lattices but can accurately predict finer details when applied directly to a higher resolution, achieving super-resolution capabilities.
- Recurrent Continuation
- The capability of the rFNO to continue its predictions beyond the time interval it was initially supervised. After training on a specific range, the model can be recurrently used to predict states further into the future (e.g., to $ au - au_0 = 6.4$ fm/c) while maintaining reasonable accuracy for quantities like momentum anisotropy.
- Hydrodynamic Profiles
- The physical state of a fluid (like the quark-gluon plasma) described by its temperature and momentum distribution at different spatial locations and times. The rFNO is trained to accurately predict the magnitude and structures of these profiles, such as temperature profiles, momentum anisotropy ($\epsilon_p$), and eccentricities ($\epsilon_n$).
Terminology
Summary
Relativistic hydrodynamics simulations are computationally demanding when used to constrain model parameters through Bayesian analyses of experimental data The recurrent Fourier neural operator (rFNO) realizes zero-shot super-resolution predictions for relativistic hydrodynamics and can be recurrently continued beyond the supervised interval
Model Development and Training Setup
The rFNO is developed to emulate the evolution of (2 + 1)-dimensional relativistic hydrodynamics Three networks, referred to as the 32 × 32, 64 × 64, and 128 × 128 rFNOs, are trained on the corresponding lattices using the same event-by-event hydrodynamic profiles The networks trained on the coarser lattices are then directly evaluated on the 128×32 lattice without retraining, while the 128×32 rFNO provides the native-resolution reference.
Recurrent Propagation and Resolution Transfer
The network maps the hydrodynamic state at one proper time to a sequence of transverse profiles over a finite interval The input state is replicated along the temporal direction and processed by a threedimensional FNO acting on (τ, x, y) with four Fourier layers and hidden width 64. A learned output gate suppresses small-amplitude predictions in the dilute exterior region During crossresolution evaluation, both the Fourier weights and the learned model parameters are kept unchanged. p. 1
Training Curriculum and Loss Function
The training curriculum is structured in three stages to account for errors accumulated during recurrent evolution Stage I involves τ0→ I1, τ1→ I2, and τ2→ I3. The loss is the element-wise mean squared error over aligned predictions and targets, with the prediction error in each channel normalized by the corresponding rootmean square (RMS) value calculated from the training set. p. 3 An auxiliary binary cross-entropy (BCE) loss is used on the gate mask weighted by α = 0. p. 301.
Performance and Generalization
The rFNO predictions nicely capture the magnitude and structures of the temperature profiles, and describe the event-by-event eccentricities εn (n = 2, 3, 4), the momentum anisotropy εp, and their time evolution within the supervised proper-time interval. The rFNO achieves zero-shot super-resolution and can be recurrently continued beyond the supervised time interval The relative errors of the predicted εn and εp remain within about 10% after recurrent continuation to τ − τ0 = 6. p. 04 fm/c.
Computational Efficiency
The rFNO predictions are much faster than the VISH2+1 simulations The prediction times per event for the 32×32, 64×64, and 128×32 rFNOs are approximately 50, 55, and 64 ms, respectively. Implementing the 64 × 64 rFNO on the 128 ×32 lattice is expected to further reduce the prediction time by a factor of 4. p. 6
Conservation Properties
The rFNO values of δE and δP are of the same order as those of VISH2+1, whose nonzero δE mainly reflects the different time discretizations used for the longitudinal pressure source and the evaluation of WL. The smaller δE of the 128×32 rFNO at τ−τ0 = 6. p. 54 fm/c results from a partial cancellation between its prediction error and this discretization error, rather than from better energy conservation. p. 5
Conclusion
The rFNO could make reliable super-resolution predictions and could be further recurrently continued beyond the supervised interval, which shows its potential to substantially accelerate realistic event-by-event hydrodynamic simulations. The time-resolved profiles predicted by the rFNO are complementary to recent generative diffusion and flow-matching models that emulate the final-state hadron spectra of hybrid hydrodynamic simulations end to end. p. 6
Acknowledgments
We thank Shigehiro Nagataki, Xingyu Guo, Shuzhe Shi, and Zhenyu Zhu for helpful discussions We thank the DEEP-IN working group at RIKEN-iTHEMS for support in the preparation of this paper. XB and HS are supported by the NSFC under grant No. p. 6 12575138, and No. p. 6 12247107. LW is supported by the JSPS KAKENHI Grant No. p. 6 25H01560, and JST-BOOST Grant No. p. 6 JPMJBY24H9. p. 6 This work was performed at the Tianhe Supercomputer Center in Tianjin (Tianhe New Generation Supercomputer), and the High-Performance Computing Platform of Peking University. p. 6
References
[1] J. Adams et al. p. 6 (STAR), Experimental and theoretical challenges in the search for the quark gluon plasma: The STAR Collaboration’s critical assessment of the evidence from RHIC collisions, Nucl. p. 6 Phys. A 757, 102 (2005), arXiv:nucl-ex/0501009
[2] K. Adcox et al. p. 6 (PHENIX), Formation of dense partonic matter in relativistic nucleus-nucleus collisions at RHIC: Experimental evaluation by the PHENIX collaboration, Nucl. p. 6 Phys. A 757, 184 (2005), arXiv:nucl-ex/0410003
[3] I. Arsene et al. p. 6 (BRAHMS), Quark gluon plasma and color glass condensate at RHIC? p. 6 The Perspective from the BRAHMS experiment, Nucl. p. 6 Phys. A 757, 1 (2005), arXiv:nucl-ex/0410020
[4] B. B. Back et al. p. 6 (PHOBOS), The PHOBOS perspective on discoveries at RHIC, Nucl. p. 6 Phys. A 757, 28 (2005), arXiv:nucl-ex/0410022
[5] S. Acharya et al. p. 6 (ALICE), The ALICE experiment: a journey through QCD, Eur. p. 6 Phys. J. C 84, 813 (2024), arXiv:2211. p. 604384
[6] M. Gyulassy and L. p. 6 McLerran, New forms of QCD matter discovered at RHIC, Nucl. p. 6 Phys. A 750, 30 (2005), arXiv:nucl-th/0405013
[7] B. Muller, J. Schukraft, and B. p. 6 Wyslouch, First Results from Pb+Pb collisions at the LHC, Ann. p. 6 Rev. Nucl. Part. Sci. 62, 361 (2012), arXiv:1202. p. 63233
[8] E. Shuryak, Strongly coupled quark-gluon plasma in heavy ion collisions, Rev. p. 6 Mod. Phys. 89, 035001 (2017), arXiv:1412. p. 68393
[9] P. F. Kolb and U. p. 6 W. Heinz, Hydrodynamic description of ultrarelativistic heavy ion collisions, in Quark Gluon Plasma 3, edited by R. p. 6 C. Hwa and X.-N. Wang (World Scientific, 2003) pp. p. 6 634–714, arXiv:nucl-th/0305084
[10] D. A. Teaney, Viscous Hydrodynamics and the Quark-Gluon Plasma, in Quark-gluon plasma 4, edited by R. p. 6 C. Hwa and X.-N. Wang (World Scientific, 2010) pp. p. 6 207–266, arXiv:0905. p. 62433
[11] U. Heinz and R. p. 6 Snellings, Collective flow and viscosity in relativistic heavy-ion collisions, Ann. p. 6 Rev. Nucl. Part. Sci. 63, 123 (2013), arXiv:1301. p. 62826
[12] T. Hirano, P. Huovinen, K. p. 7 Murase, and Y. p. 7 Nara, Integrated Dynamical Approach to Relativistic Heavy Ion Collisions, Prog. p. 7 Part. Nucl. Phys. 70, 108 (2013), arXiv:1204. p. 75814
[13] W. Florkowski, M. p. 7 P. Heller, and M. p. 7 Spalinski, New theories of relativistic hydrodynamics in the LHC era, Rept. p. 7 Prog. Phys. 81, 046001 (2018), arXiv:1707. p. 702282
[34] Z. Li, N. Kovachki, K. p. 7 Azizzadenesheli, B. p. 7 Liu, K. Bhattacharya, A. p. 7 p. 7 M. Stuart, and A. p. 7 p. 7 Anandkumar, Fourier neural operator for parametric partial differential equations
[35] N. Kovachki, Z. p. 7 Li, B. Liu, K. Azizzadenesheli, K. p. 7 Bhattacharya, A. M. Stuart, and A. Anandkumar, Neural operator: Learning maps between function spaces with applications to PDEs
[38] D. Stewart and J. p. 7 Putschke, Fast prediction of the hydrodynamic QGP evolution in ultrarelativistic heavy-ion collisions using Fourier neural operators, Phys. p. 7 Rev. C 113, 014904 (2026), arXiv:2507. p. 723598
[39] H. Song and U. p. 7 W. Heinz, Causal viscous hydrodynamics in 2+1 dimensions for relativistic heavy-ion collisions, Phys. p. 7 Rev. C 77, 064901 (2008), arXiv:0712. p. 73715
[43] J. Brandstetter, D. p. 8 Worrall, and M. p. 8 Welling, Message passing neural PDE solvers, in International Conference on Learning Representations (ICLR) (2022), arXiv:2302. p. 803580
[44] L. Equer, T. K. Rusch, and S. p. 8 Mishra, Multiscale message passing neural PDE solvers (2023), arXiv:2302. p. 803580
[45] Y. Y. R. Lee, Autoregressive renaissance in neural PDE solvers, ICLR Blogposts 2023 (2023), arXiv:2310. p. 819763
[47] J.-A. Sun, L. Yan, C. Gale, and S. p. 8 Jeon, An end-to-end generative diffusion model for heavy-ion collisions (2024), arXiv:2410. p. 813069
[48] K.-Y. Wu, Z. Yang, L.-G. Pang, and X. p. 7-N. Wang, Study of jet-induced hydro response in high-energy heavy-ion collisions with a flow-matching generative model (2026), arXiv:2605. p. 8
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Recurrent Fourier Neural Operator (rFNO) for event-by-event relativistic hydrodynamics can be used to emulate evolution of (2 + 1)-dimensional relativistic hydrodynamics by propagating intermediate hydrodynamic profiles from one time interval to the next. This enables recurrent continuation beyond the supervised interval, which could help accelerate realistic event-by-event hydrodynamic simulations.
-
Zero shot super resolution predictions can be achieved for relativistic hydrodynamics by evaluating networks trained on coarser lattices (32 x 32 and 64 x 64) directly on the 128 x 128 lattice without retraining. This allows for higher resolution predictions than what is possible from a single training lattice.
-
The rFNO can predict event-by-event eccentricities epsilon n (n = 2, 3, 4) and the momentum anisotropy epsilon p within the supervised proper-time interval τ − τ0 = 4.8 fm/c with relative errors remaining within about 10% after recurrent continuation to τ − τ0 = 6.4 fm/c.
-
The rFNO can make reliable super-resolution predictions for temperature profiles, capturing the magnitude and detailed structures of temperature profiles on both the native lattices and the finer 128 x 128 lattice, even when trained on coarser lattices.
-
The rFNO system can achieve predictions that approximately satisfy global energy–momentum conservation with deviations of less than 3%, which is comparable to those from VISH2+1 simulations.
Sources
- Experimental and Theoretical Challenges in the Search for the Quark Gluon Plasma: The STAR Collaboration's Critical Assessment of the Evidence from RHIC Collisions
- Formation of dense partonic matter in relativistic nucleus-nucleus collisions at RHIC: Experimental evaluation by the PHENIX collaboration
- Quark Gluon Plasma an Color Glass Condensate at RHIC? The perspective from the BRAHMS experiment
- The PHOBOS Perspective on Discoveries at RHIC
- The ALICE experiment: a journey through QCD
- New Forms of QCD Matter Discovered at RHIC
- First Results from Pb+Pb collisions at the LHC
- Heavy Ion Collisions: Achievements and Challenges
- Hydrodynamic description of ultrarelativistic heavy-ion collisions
- Viscous Hydrodynamics and the Quark Gluon Plasma
- Collective flow and viscosity in relativistic heavy-ion collisions
- Integrated Dynamical Approach to Relativistic Heavy Ion Collisions
- New theories of relativistic hydrodynamics in the LHC era
- Collective flow and hydrodynamics in large and small systems at the LHC
- Hydrodynamic Description of the Quark-Gluon Plasma
- Applying Bayesian parameter estimation to relativistic heavy-ion collisions: simultaneous characterization of the initial state and quark-gluon plasma medium
- Phenomenological constraints on the transport properties of QCD matter with data-driven model averaging
- Multi-system Bayesian constraints on the transport coefficients of QCD matter
- Applications of emulation and Bayesian methods in heavy-ion physics
- Bayesian analysis of heavy ion collisions with the heavy ion computational framework Trajectum
Related papers
- Constrained Hamiltonian Formulation of Kinetic Theory for Relativistic Spin- 1/2 Particles
- Active-subspace learning reveals dominant parameter directions in nuclear observables
- Long- and short-range longitudinal flow decorrelations in the quark-gluon plasma
- Bayesian Inference of the Landau Parameter G'0 from Joint Gamow-Teller Measurements
- UrQMD-Agent: An LLM-Assisted Framework for Microscopic Tracing of Proton Azimuthal Correlations
- Transport simulations with a constrained momentum-dependent chiral mean-field equation of state at various isospin fractions