acceptodds
Under review as a conference paper at ICLR 2027

Learn the Dynamics, Enforce the Physics: Generative Latent Operators for High-Dimensional Beam Evolution

Abstract

Particle-in-cell (PIC) simulation is indispensable for modeling beam dynamics in particle accelerators, but its computational cost limits many-query workflows, such as configuration exploration, virtual diagnostics, and controller design. We introduce *GLOBE*, a physics-constrained **G**enerative **L**atent **O**perator for **B**eam **E**volution, modeling the nonlinear, control-dependent, multiscale evolution of 3D beam-field dynamics covering spatial cells. GLOBE compresses beam-field states into a compact latent space and factors their evolution into two components. A control-conditioned neural operator learns the dominant dynamics, while an amplitude-normalized generative model captures the residual distribution, recovering fine-scale structures and conditional variations. GLOBE enforces field constraints by regime-specific, parameter-free operators: a Poisson solve reconstructs the electric field in the electrostatic regime, and a spectral projection imposes Gauss's law and a divergence-free magnetic field in the electromagnetic regime. On two high-resolution 3D particle-accelerator benchmarks, bunched-beam transport and plasma-wakefield acceleration, GLOBE lowers relative- error by 9% and 13%, low-frequency spectral error by 69% and 51%, and mid-frequency spectral error by 54% and 20% relative to the strongest baseline (DPOT), satisfies field constraints, and reduces hours of PIC simulation to a minute per rollout.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Particle Accelerator, Operator Learning, Generative Modeling, Vlasov-Maxwell System, PDE Learning, Particle-in-Cell Simulation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_ZQWHOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents GLOBE, a surrogate model for high-resolution 3D beam–field PIC simulations. It supports 10^7 PIC cells, achieves lower relative-L2 error, and captures finer spatial structure by reducing the error in low-frequency and medium-frequency domains.

Critical strengths
  1. GLOBE predicts full 3D beam–field evolution on grids containing 10 million cells, achieving over 100× faster inference than PIC simulation.
  2. The framework enforces physical constraints with little loss of prediction accuracy.
  3. Low-frequency and medium-frequency errors are significantly suppressed.
Critical weaknesses
  1. The 2nd contribution bullet point in introduction does not point to a research gap. In the 2nd paragraph of introduction, if I understand correctly, scalability of these hybrid designs is the sole declared research gap.
  2. It is unclear if the comparison with baselines is fair.
  3. Although GLOBE runs on large 3D grids, the scalability advantage remains unclear.
Questions for authors
  1. The 6600s and 4130s in Section 5 Experiments, do they refer to GPU seconds? I presume they do not mean the number of GPUs.
  2. Are the baselines adequately trained, and exposed to the same information as GLOBE and variances?
  3. Could you test the model with data outside training scope, to evaluate the model's transferability?
  4. Could you systematically increase resolution and measure how runtime and memory grow compared to baselines?
6 comments
AuthorsOpenReview ↗

Response to Reviewer ZQWH (1/2)

We thank the reviewer for the detailed and constructive feedback. We address W1, Q1, and W2 / Q2 here, and Q3 and W3 / Q4 in the next comment.

W1. The second contribution does not point to a research gap.

We acknowledge that the current introduction emphasizes the computational challenge of scaling hybrid deterministic–generative models but does not sufficiently motivate the separate challenge addressed by our physics-constrained operators. A learned surrogate predicts the charge density and fields at each time step without any mechanism that keeps them consistent with each other. They should satisfy the instantaneous constraints in Eqs. (2)–(3): Gauss's law ∇⋅E=ρ/ε0\nabla\cdot\mathbf E=\rho/\varepsilon_0 and ∇⋅B=0\nabla\cdot\mathbf B=0, and, in the electrostatic regime, the Poisson equation linking E\mathbf E to ρ\rho. Decoding from a compressed latent space and adding a sampled residual make violations more likely. Table 3 quantifies this. Without the constraint operator, the decoded PWFA fields violate Gauss's law and ∇⋅B=0\nabla\cdot\mathbf B=0 at the 10−210^{-2} to 10−310^{-3} level. Penalty terms in the loss can reduce such violations but cannot guarantee that they vanish. Our second contribution closes this gap with parameter-free operators built from the simulation's own discretization, which reduce both violations to the 10−710^{-7} level with nearly unchanged accuracy (Rel-L2L_2 0.191 without, 0.194 with). We will revise the introduction to state three gaps: (1) long-horizon, mesh-resolved prediction of 3D beam–field evolution is largely unexplored, (2) existing hybrid deterministic–generative designs scale poorly to such grids and to heterogeneous field amplitudes, and (3) learned surrogates give no guarantee that predicted fields satisfy the field constraints.

Q1. Do 6600 s and 4130 s refer to GPU seconds?

Yes. Each is the wall-clock time of one WarpX simulation on one GPU, i.e., GPU seconds per trajectory (about 1.8 h and 1.1 h), not a number of GPUs. This matches the total compute reported for generating the two datasets in Appendix A (about 900 and 600 GPU hours for 474 and 504 runs). All surrogate inference times are also measured on one GPU, one trajectory at a time.

W2 / Q2. Are the baselines adequately trained and given the same information?

  • Same information. GLOBE, its variants, and all baselines use the same trajectory-level train / validation / test splits, the same observed history (8 / 4 frames), the same machine controls, and the same free-rollout evaluation over all frames.
  • Same capacity. Every baseline is parameter-matched to GLOBE's full pipeline (90.0M to 92.6M for bunched-beam transport and 107.7M to 111.0M for PWFA, against 91.4M / 111.8M for GLOBE-C, Table 1). Each is trained with its published implementation.
  • Convergence. The table gives each baseline's validation rollout error (Rel-L2L_2) at the checkpoint at 90% of training and at the checkpoint after training finished (final checkpoint).
DatasetModelRel-L2L_2 at 90%Rel-L2L_2 finalChange
Bunched-beam transportDPOT0.08940.0887−0.8%
MPP0.08820.0874−0.9%
UPT0.12910.1281−0.8%
P3D0.10660.0987−7.4%
PWFADPOT0.08820.0863−2.2%
MPP0.07760.0768−1.0%
UPT0.42600.4215−1.1%
P3D0.12320.1095−11.1%

DPOT, MPP, and UPT change by about 1% or less over the last 10% of training (2.2% for DPOT on PWFA), so their training has converged. P3D's official recipe uses a constant learning rate, so its training is oscillatory. On PWFA, its error rises at 91% of training (0.123 to 0.128), drops sharply at 97% (0.105), and rises again at the final checkpoint (0.110), which produces the −11.1%. Bunched-beam transport shows a similar late step: the error stays at about 0.106 from 67% to 85% of training and then drops to 0.099 at 97% (−7.4%).

Revision. We will state the three research gaps in the introduction and add the training setup and convergence curves to the appendix.

AuthorsOpenReview ↗

Response to Reviewer ZQWH (2/2)

Q3. Testing outside the training scope.

We generated new out-of-distribution (OOD) test sets of 19 bunched-beam transport and 20 PWFA trajectories. For each simulation, we sample multiple design parameters outside their training ranges, for example sampling phase advance between 50–60° or 85–89° (training 60–85°) for bunched-beam transport and driver charge Λ\Lambda 1.2–1.6 (training up to 1.2) for PWFA. None of these configurations is seen during training, and all models are evaluated with the same protocol as the main test sets (reported in Rel-L2L_2 error):

ModelBunched-beam transport (19)PWFA (20)
GLOBE-D0.1930.311
GLOBE-C, mean of 8 samples0.1940.312
GLOBE-C, single sample0.2070.319
DPOT0.3080.319
MPP0.4360.526
UPT0.8650.747
P3D-D0.9040.739
P3D-G0.6551.829
HFS (in our latent space)0.3591.569
HFS+DM (in our latent space)0.3712.536

On bunched-beam transport, GLOBE-D's error is 37% lower than that of the best baseline (DPOT, 0.308). On PWFA, GLOBE-D (0.311) and DPOT (0.319) are comparable, and both are far more accurate than the other baselines. Every model's error is higher than the error on in-distribution test sets (Table 1), as expected under a distribution shift. These results assess robustness to unseen configurations outside the training ranges. We do not claim transfer to other machines, lattices, or geometries, which remains future work.

W3 / Q4. Computational scalability and resolution dependence.

We clarify that GLOBE's advantage is computationally tractable prediction on high-resolution 3D grids that preserves fine-scale structure, not resolution invariance property. Unlike discretization-invariant neural operators such as FNO (Li et al., 2021), GLOBE is trained at a fixed spatial resolution, and we do not claim zero-shot transfer to other discretizations.

  • Computational advantage. GLOBE compresses 28.9M / 15.7M grid cells into 28.2K / 30.6K latent tokens, which makes global attention and generative refinement feasible at this resolution. GLOBE-C is 102× / 189× faster than WarpX and 23× / 14× faster than P3D-G, which generates the full state, and the hybrid HFS baselines did not scale to our full grids (Section 5.3). GLOBE is not the fastest surrogate, since DPOT has has faster inference runtime, but DPOT is less accurate (Rel-L2L_2 0.177 / 0.224 vs. 0.144 / 0.174 for GLOBE-D) and recovers much less fine-scale structure (mid-band spectral error 0.986 / 0.377 vs. 0.457 / 0.300 for GLOBE-C).
  • Cost structure. The convolutional encoder and decoder scale linearly with the number of grid cells, and the dynamics models see only the latent tokens. Full attention remains quadratic in the number of latent tokens, so its cost is set by the latent grid rather than the mesh, and compression reduces the token count by 1024× / 512×.
  • Resolution-scaling study (in progress). Increasing the resolution further is beyond our resources: the current benchmarks took about 1,500 GPU hours to simulate and occupy about 7 TB, and doubling the resolution per axis would multiply both by at least 8. We are therefore conducting a controlled resolution-scaling study on progressively downsampled bunched-beam transport trajectories, retraining GLOBE-D, GLOBE-C, and baselines at each resolution and measuring inference time, training cost, peak GPU memory, and prediction error. This experiment characterizes computational scaling rather than zero-shot resolution transfer. We will report the results in a separate comment during the discussion period once it is completed.
  • Limitation. The study uses coarsened versions of our PIC data, so it does not establish accuracy at resolutions finer than our benchmarks. GLOBE's learned latent representation is also resolution-dependent, and extending one trained model to arbitrary discretizations remains future work.

Revision. We will clarify the distinction between computational scalability and resolution invariance, add the resolution study and the OOD results to the appendix, and state these limitations.

We hope these clarifications address the reviewer's concerns.

References

  • Z. Li et al. Fourier neural operator for parametric partial differential equations. ICLR (2021).

Dear Reviewer ZQWH,

The authors have submitted their rebuttal and we are closing in on the end of the discussion period. Please take some time to carefully read over the rebuttal and respond whether your concerns have been addressed.

Best, AC

Reviewer_Edp4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GLOBE, a physics-constrained generative latent operator for high-dimensional beam–field evolution in particle accelerators. The method combines a compressed latent representation, a deterministic neural operator for dominant dynamics, a generative residual model for fine-scale corrections, and parameter-free operators enforcing field constraints. Experiments on two high-resolution 3D particle-accelerator benchmarks demonstrate improvements in prediction accuracy and computational efficiency over several baselines.

Critical strengths

Practical significance. The paper addresses the high computational cost of particle-in-cell simulations, a meaningful bottleneck in accelerator modeling. The reported speedups and predictive accuracy improvements suggest practical value.

Strong empirical validation. Experiments on two high-resolution 3D benchmarks demonstrate improvements over several competitive baselines. The physics-constraint ablation is particularly convincing, showing substantially reduced constraint violations with negligible loss in predictive accuracy.

Critical weaknesses

Missing comparison to relevant prior work. The paper overlooks Learning Efficient Hybrid Particle-Continuum Representations of Non-Equilibrium N-Body Systems (Wu, Sun et al., 2022). Both works address efficient learning-based simulation of multiscale, nonequilibrium particle dynamics, including plasma systems with electromagnetic interactions. Although LHPC focuses on learned particle–continuum coupling rather than latent beam–field evolution, the overlap warrants discussion to better contextualize GLOBE's novelty.

Insufficient justification of architectural choices. The framework introduces numerous physics equations alongside ML design choices, without clearly distinguishing physically necessary constraints from empirical modeling decisions. While Tables 2–4 provide useful ablations for the Haar transform, multiscale shortcuts, backbone, and physics projection, other components, particularly the three-level residual normalization and flow-matching formulation, lack comparable justification. This makes it difficult to assess which design choices are essential to the reported improvements.

Questions for authors
  1. How does GLOBE relate to LHPC (Wu, Sun et al., 2022), particularly regarding multiscale plasma dynamics and computational efficiency?
  2. Can the authors more explicitly distinguish physics-mandated components from empirical architectural choices? In particular, what evidence supports the three-level residual normalization and the chosen generative formulation over simpler alternatives?
5 comments
AuthorsOpenReview ↗

Response to Reviewer Edp4 (1/2)

We thank the reviewer for the constructive feedback and for pointing us to LHPC. We address Q1 / W1 here, and Q2 / W2 in the next comment.

Q1 / W1. How does GLOBE relate to LHPC (Wu et al., 2022), particularly regarding multiscale plasma dynamics and computational efficiency?

We agree that LHPC is highly relevant to learning-based acceleration of plasma simulation, and we will include it in our literature review. Both methods aim to make multiscale plasma simulation more efficient, but they differ fundamentally in their physical representation, their training objective, and their reliance on a numerical solver at inference.

  • Representation and training objective. LHPC is a hybrid solver. It splits the local velocity distribution into a thermal part, advanced as a continuum by a neural network, and a non-thermal part, advanced as particles by a conventional particle solver. A third network learns the exchange of mass, momentum, and energy between the two. GLOBE is a surrogate for the full simulation. It learns the evolution of mesh-resolved charge densities and electromagnetic fields in a compressed latent space, conditioned on machine controls, without tracking any particles at inference.
  • Multiscale modeling. LHPC handles multiscale, non-equilibrium physics in velocity space and keeps kinetic detail for the non-thermal population. GLOBE handles it across spatial scales, separated by spectral energy. The deterministic operator captures the energetic large-scale (low-frequency) evolution, and the amplitude-normalized generative residual recovers the weak fine-scale (higher-frequency) structure.
  • Computational efficiency. LHPC reports an 8× speedup over a classical particle solver on an intense laser-plasma interaction problem, while still advancing the non-thermal particles with that solver. GLOBE-C requires no particle push and is 102× / 189× faster than WarpX on our two 3D benchmarks with 28.9M / 15.7M grid cells. These figures come from different problems, hardware, and predicted quantities, so they are not directly comparable.

Why LHPC is not included as a baseline. LHPC is trained on particle-level data and keeps a particle solver in the loop, while our benchmarks store mesh-resolved fields. It therefore cannot be trained on our benchmarks as is, so we add it as related work rather than as a baseline.

Revision. We will cite and discuss LHPC in Section 2, including this distinction and the complementary strengths of the two approaches.

References

  • T. Wu et al. Learning efficient hybrid particle-continuum representations of non-equilibrium N-body systems. NeurIPS 2022 Workshop on AI for Science: Progress and Promises (2022).
AuthorsOpenReview ↗

Response to Reviewer Edp4 (2/2)

Q2 / W2. Which components are physics-mandated and which are empirical? What supports the three-level residual normalization and the generative formulation over simpler alternatives?

We agree that the paper should draw this line explicitly.

(a) Physics-mandated versus empirical components, and the motivation for each.

  • Physics-mandated: the field constraints and the variables they determine. In the electrostatic regime, the field is fully determined by the charge density, so we predict only the density and recover the field by a Poisson solve with the simulation's boundary conditions. In the electromagnetic regime, densities and fields carry independent information and are predicted jointly, and a spectral projection matched to the simulation's discretization enforces Gauss's law and ∇⋅B=0\nabla\cdot\mathbf B=0. These operators have no learned parameters.
  • Physics-guided: operator learning plus a generative residual. The beam–field evolution follows a PDE (Vlasov–Poisson or Vlasov–Maxwell, Eqs. 1–3), whose solution operator maps the current state and machine controls to future states. We therefore learn this map with a deterministic neural operator. The PIC simulation also contains realization-dependent fine-scale structure from finite-particle sampling. Deterministic neural operators, however, exhibit spectral bias: they primarily predict the high-energy content and over-smooth the weak fine-scale structure (Khodakarami et al., 2025). We therefore add a generative component that learns this residual as a distribution, so that each prediction recovers realistic fine-scale structure (see our response to Reviewer F2iH, W1).
  • Empirical: the network architecture and training details. We chose a Transformer backbone because global attention captures long-range spatial dependencies, and attention is a discretized kernel integral operator (Kovachki et al., 2023, and Cao, 2021). Similarly, flow matching is one standard way to learn the sampling density of the residual. The autoencoder design (Table 2), the backbone (Table 4), and the residual normalization (Table 5) are ablated. Input noise and pushforward training follow prior work.

(b) Why a three-level normalization.

The generative model learns the residual r\mathbf r between the true latent state and the deterministic forecast, which carries the fine-scale structure. Its amplitude, however, varies by orders of magnitude across channels, forecast windows, and spatial locations. Training flow matching directly on these raw residuals biases the objective toward large-amplitude components, making weaker structures difficult to learn consistently. To quantify this, an ANOVA of the log residual energy log⁡r2\log\mathbf r^2 on the test sets measures how much of its variance is explained by channels, forecast windows, and spatial tokens.

Source of variationBunched-beam transportPWFA
Channel4.1%0.6%
Forecast window10.1%54.2%
Spatial token73.7%22.8%
Window-dependent spatial variation5.2%10.5%
Covered by our three-level scale93.1%88.1%
Other6.9%11.9%

Most of the variance lies along spatial tokens, forecast windows, and their interaction. Capturing it therefore requires scales at multiple levels, and the factors represented by our three-level scale account for 93.1% / 88.1% of the variance. Our learned normalizer reduces the window RMS ratio (90th/10th percentile) from 4.46 / 47.63 to 1.49 / 1.62 and the top-10% token energy share from 97.9% / 78.0% to 25.2% / 31.4%. The ablation in Table 5 (Appendix D.1) is consistent with this. With the channel scale alone, the mid-band spectral error on bunched-beam transport rises from 0.457 to 0.894 (GLOBE-D: 0.994). On PWFA, the mid-band spectral error rises from 0.300 to 4.32 (GLOBE-D: 0.350) and Rel-L2L_2 from 0.194 to 0.266.

(c) Evidence for the generative formulation.

In Table 1, compared with the deterministic operator alone (GLOBE-D), the combined GLOBE-C reduces the mid-band spectral error from 0.994 / 0.350 to 0.457 / 0.300. Compared with a single generative model forecasting the latent state directly (GLOBE-G), GLOBE-C reduces Rel-L2L_2 from 0.251 / 1.954 to 0.161 / 0.194.

Revision. In Section 4, we will separate physics-mandated from empirical components and point to Appendix D.1, where we add the variance analysis.

References

  • S. Khodakarami et al. Mitigating spectral bias in neural operators via high-frequency scaling for physical systems. Neural Networks, 108027 (2025).
  • N. Kovachki et al. Neural operator: Learning maps between function spaces with applications to PDEs. J. Mach. Learn. Res. 24 (2023).
  • S. Cao. Choose a Transformer: Fourier or Galerkin. Advances in Neural Information Processing Systems 34 (2021).

Dear Reviewer Edp4,

The authors have submitted their rebuttal and we are closing in on the end of the discussion period. Please take some time to carefully read over the rebuttal and respond whether your concerns have been addressed.

Best, AC

Reviewer_F2iHOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The proposed model includes 3 components: a convolutional autoencoder first compresses the high-dimensional simulation states into a  latent representation, a deterministic Transformer then predicts how this representation evolves, using previous states and accelerator control parameters, because deterministic forecasts tend to miss fine spatial details, a second, generative transformer trained with flow matching models the residual errors, and the final prediction combines the two.

Critical strengths

They demonstrate strong performance on actual 3d simulation tasks. The paper grounds the predictions using actual physics. The paper provides good ablations to motivate their components of the models, especially the convolutional autoencoder, especially they showed why using Haar transform helps. The paper evaluates the method on two large-scale 3D accelerator simulation datasets; their numbers seem to good competitive results.

Critical weaknesses

One thing that sort of confused me is that their generative component meant to capture the intricate details essentially degrades a lot of metrics compared to if they just used their transformer, their argument is that this generative component improves some spectral predictions but on the actual simulation metrics it always degrades the performance, makes me think why is this finer capturing even physically important, they don't elaborate on this. The physical constraints are enforced at individual time steps, there doesn't need to be a guarantee that full complete simulation follows these laws, I guess they could point out the scope of these a bit more

Questions for authors

Could you clarify the runtime and efficiency of the proposed approach compared to standard ML approaches? Does algorithm complexity factor into the design?

4 comments
AuthorsOpenReview ↗

Response to Reviewer F2iH (1/2)

Response to Reviewer F2iH (1/2)

We thank the reviewer for the constructive feedback. We address W1 here, and W2 and Q1 in the next comment.

W1. The generative component increases Rel-L2L_2. Is fine-scale structure physically important?

Why fine-scale structure matters.

  • Bunched-beam transport. A key goal in tuning high-intensity beam transport is to limit the beam halo, a small fraction of particles that drift away from the dense core, hit the beam pipe, and are lost. The halo is driven by the beam's own repulsive forces, especially when the beam is not well matched to the focusing channel (Fedotov, 2006). Because the dense core dominates pointwise metrics, a smoothed prediction can attain low Rel-L2L_2 error while missing the low-density region and fine density structure where the halo develops.
  • PWFA. A key goal in PWFA is to accelerate the witness bunch with a small energy spread. Because the accelerating field is produced by the beams and plasma themselves rather than by external devices, it is shaped by small-scale structures such as the thin plasma-electron sheath and the profile of the witness bunch, and small errors in these structures change the energy spread (Tzoufras et al., 2008). These are local, fine-scale details, so a smoothed prediction can attain low Rel-L2L_2 error while blurring them.

Why Rel-L2L_2 error rises.

In operator learning, let cc be the conditional information (observed history and controls). Because finite-particle sampling introduces realization-dependent variations that are not fully determined by cc, the ground-truth reference trajectory YY in the data is one realization drawn from p(Y∣c)p(Y\mid c). A model trained with squared error tends to predict the conditional mean μ(c)=E[Y∣c]\mu(c)=\mathbb E[Y\mid c]. From the energy-spectrum perspective, however, the mean is not a realistic state. For each realization, decompose it as Y=μ(c)+δY=\mu(c)+\delta, where δ\delta is the run-dependent residual with E[δ∣c]=0\mathbb E[\delta\mid c]=0. With ⋅^(k)\hat{\cdot}(k) the Fourier coefficient at wavenumber kk, E[∣Y^(k)∣2∣c]=∣μ^(k)∣2+E[∣δ^(k)∣2∣c]\mathbb E[|\hat Y(k)|^2\mid c]=|\hat\mu(k)|^2+\mathbb E[|\hat\delta(k)|^2\mid c]. The spectrum of the mean thus lacks the energy of δ\delta, mostly at small scales where realizations differ. A generative residual restores this energy, but its details cannot align pointwise with the reference run. Even an ideal sample Y~∼p(Y∣c)\tilde Y\sim p(Y\mid c), drawn independently of YY, has E∥Y~−Y∥2=2 E∥δ∥2\mathbb E\|\tilde Y-Y\|^2=2\,\mathbb E\|\delta\|^2, twice the squared error of μ(c)\mu(c). Other fields show the same trade-off. In weather downscaling, a diffusion residual restores fine-scale variance missing from a regression model while slightly increasing MAE (Mardani et al., 2025). In turbulence super-resolution, diffusion refinement reduces the energy-spectrum error from 0.160 to 0.014 while raising the field error from 0.054 to 0.069 (Oommen et al., 2026).

Ensemble averaging largely recovers deterministic accuracy.

By design, GLOBE-C adds a residual r~\tilde{\mathbf r} sampled by Rξ\mathcal R_\xi to the GLOBE-D forecast z~det\tilde{\mathbf z}^{\rm det} (the estimated conditional mean), and r~\tilde{\mathbf r} aims to restore the zero-mean variation δ\delta introduced above. Averaging MM independent samples therefore cancels most of r~\tilde{\mathbf r}, so the sample mean should nearly match GLOBE-D, while each sample keeps r~\tilde{\mathbf r} and should have a higher error. We test this with M=8M=8 on the same test sets as Table 1:

Bunched-beam transport (52)PWFA (56)
GLOBE-D0.14370.1747
GLOBE-C, mean of 8 samples0.1460 (+1.6%)0.1802 (+3.1%)
GLOBE-C, single sample0.1609 (+12.0%)0.1941 (+11.1%)

The 8-sample mean stays close to GLOBE-D, while a single sample is 12% / 11% higher. Most of the increase thus comes from sampled variation, not the mean forecast. This variation carries the fine-scale content, reducing mid-band spectral error from 0.994 to 0.457 (bunched-beam transport) and from 0.350 to 0.300 (PWFA, Table 1). Despite the increase relative to GLOBE-D, GLOBE-C still outperforms the strongest baseline, DPOT, in Rel-L2L_2 (0.161 / 0.194 vs. 0.177 / 0.224).

Revision. In Section 5.1, we will present GLOBE-D and GLOBE-C as complementary and add the ensemble-mean result, with the motivation and analysis above in the appendix.

References

  • M. Tzoufras et al. Beam loading in the nonlinear regime of plasma-based acceleration. Phys. Rev. Lett. 101, 145002 (2008).
  • A. V. Fedotov. Beam halo formation in high-intensity beams. Nucl. Instrum. Methods Phys. Res. A 557, 216–219 (2006).
  • M. Mardani et al. Residual corrective diffusion modeling for km-scale atmospheric downscaling. Commun. Earth Environ. 6, 124 (2025).
  • V. Oommen et al. Learning turbulent flows with generative models for super resolution and sparse flow reconstruction. Nat. Commun. 17, 3707 (2026).
AuthorsOpenReview ↗

Response to Reviewer F2iH (2/2)

W2. The physical constraints are enforced at individual time steps. The scope should be stated.

We agree and will state the scope explicitly. GLOBE enforces the instantaneous physical relations exactly and learns the temporal evolution from data.

  • Enforced: instantaneous field constraints. These relations contain no time derivative, so they can be imposed on every predicted frame. In the electrostatic (bunched-beam transport) regime, the field is fully determined by the charge density (Eq. 3), so GLOBE computes E\mathbf E from the predicted density with a Poisson solve. The Poisson equation, Gauss's law, and the curl-free condition then hold to solver tolerance. In the electromagnetic (PWFA) regime, a spectral projection enforces the two constraints in Eq. (2), Gauss's law (in the interior, Appendix B.4) and ∇⋅B=0\nabla\cdot\mathbf B=0. Table 3 shows that it reduces the normalized violations from 2.72×10−22.72\times10^{-2} to 2.77×10−72.77\times10^{-7} and from 6.85×10−36.85\times10^{-3} to 2.78×10−72.78\times10^{-7}, while Rel-L2L_2 changes only from 0.191 to 0.194.
  • Learned: temporal evolution. The time evolution governed by the Vlasov equation (Eq. 1) and by Faraday's and Ampère's laws (Eq. 2) is the learning objective of GLOBE. We use the attention-based latent operator to learn this evolution accurately from the simulation data, conditioned on the observed history and machine controls. The learned evolution tracks beam-level quantities over the full rollout (Figures 3 and 4).

We will state this scope, as "instantaneous field constraints", in the abstract, contributions, and method section.

Q1. Runtime and efficiency compared with standard ML approaches. Does algorithmic complexity factor into the design?

Yes, computational efficiency is a central design consideration, and each component of GLOBE keeps the cost low.

  • Compact latent space for surrogate learning. The fields are nonlocal, so the model needs global spatial coupling, but full attention over 2.9×1072.9\times10^7 / 1.6×1071.6\times10^7 grid cells is infeasible. GLOBE compresses each state to 28.2K / 30.6K latent tokens (about 106×10^6\times / 2.6×105×2.6\times10^5\times fewer attention pairs), where both operator learning and generative refinement take place. Generation never touches the full 3D state, and only the final latent forecast is decoded.
  • Autoencoder. A fixed, invertible 3D Haar transform maps each 2×2×22\times2\times2 block to eight subbands before any learned layer, so the convolutions operate on an 8× smaller grid without information loss. This cuts autoencoder training time by 6.2× (Table 2). The convolutional encoder and decoder scale linearly with the number of cells.
  • Operator learning. Full self-attention over the latent tokens captures global coupling at a tractable cost. Each operator call predicts a window of h=8h=8 / 44 frames jointly, so a 393-frame rollout needs about 50 sequential calls instead of 393, and a 56-frame rollout needs 14. The 16 Euler steps of the residual sampler are likewise paid once per window.
  • Parameter-free physics operators. The constraints are imposed by a closed-form projection at each Fourier mode or a Krylov Poisson solve. Neither has learnable parameters, so they add no training cost and no penalty weights to tune.

The measured costs support these choices. We compare rollout time and training cost, all measured on the same GPU type (s = seconds, h = hours, best in bold):

MethodBunched-beam transport: rolloutBunched-beam transport: trainingPWFA: rolloutPWFA: training
WarpX (PIC)6600 sn/a4130 sn/a
DPOT7.55 s68.1 h3.04 s93.3 h
P3D-D91.3 s178.6 h17.9 s115.4 h
MPP943 s576.5 h73.5 s398.4 h
UPT1384 s62.4 h114.5 s89.0 h
P3D-G1490 s173.4 h309 s131.5 h
GLOBE-D (ours)16.3 s42.8 h5.4 s91.1 h
GLOBE-C (ours)64.7 s78.4 h21.8 s140.0 h

Inference. Among deterministic models, GLOBE-D is faster than every baseline except DPOT. DPOT uses lightweight layers on patchified states, whereas GLOBE decodes its latent forecasts with 3D convolutions, but DPOT is less accurate (Rel-L2L_2 0.177 / 0.224 vs. 0.144 / 0.174). Among generative models, GLOBE-C is 23× / 14× faster than P3D-G. Compared with WarpX, GLOBE-C is 102× / 189× faster.

Training. Among deterministic models, GLOBE-D has the lowest training time on bunched-beam transport, and on PWFA only UPT trains faster (89.0 h vs. 91.1 h), because it evaluates its loss on a uniform subset of about 1% of the grid points to keep GPU memory tractable. Among generative models, GLOBE-C trains in less than half the time of P3D-G on bunched-beam transport and comparable time on PWFA.

We will add the training costs to the appendix. We hope these clarifications address the reviewer's concerns.

Dear Reviewer F2iH,

The authors have submitted their rebuttal and we are closing in on the end of the discussion period. Please take some time to carefully read over the rebuttal and respond whether your concerns have been addressed.

Best, AC

AuthorsOpenReview ↗

Revised manuscript uploaded

We thank the reviewer F2iH for the thoughtful feedback. We have uploaded an updated manuscript incorporating the revisions proposed by reviewers during the rebuttal, with the changes highlighted in blue for ease of reference.

As the discussion period comes to an end, we sincerely appreciate the reviewer's time, careful evaluation, and valuable suggestions, which have helped us improve the clarity and quality of our work.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.