Three-dimensional beam evolution skips frame-by-frame PIC: dynamics learned in latent space, physics enforced by construction
Synopsis
GLOBE compresses three-dimensional beam–field states into a compact latent space, learns the dominant evolution with a control-conditioned neural operator, and captures fine-scale residuals with a generative model, then applies parameter-free constraint operators after decoding. On two high-resolution 3D benchmarks, bunched-beam transport and plasma-wakefield acceleration, relative error is 9% and 13% lower than the strongest baseline DPOT, mid-frequency spectral error is 54% and 20% lower, and hours of particle simulation become about a minute per rollout.
Interpretation
Three-dimensional beam–field evolution can now be predicted in a compressed latent space rather than advanced frame by frame on the raw mesh. Earlier data-driven surrogates mostly mapped machine settings to low-dimensional beam quantities, reconstructed phase-space projections, or forecast evolution across successive linac modules; long-horizon prediction of mesh-resolved, self-consistent 3D beam–field states remained largely unexplored. On two high-resolution 3D benchmarks, bunched-beam transport and plasma-wakefield acceleration, relative error is 9% and 13% lower than the strongest baseline DPOT, low-frequency spectral error is 69% and 51% lower, and mid-frequency spectral error is 54% and 20% lower.
The latent forecast is split into a deterministic dominant term and a generative residual that restores fine-scale structure. Deterministic neural operators smooth away weaker fine-scale structure, while evolving the whole dynamics generatively forces one model to learn temporal evolution and conditional variation together. Adding the generative residual lowers mid-frequency spectral error from 0.994 to 0.457 on bunched-beam transport and from 0.350 to 0.300 on plasma-wakefield acceleration, while overall relative error is largely preserved.
Field constraints are imposed directly by parameter-free operators chosen per physical regime instead of being learned by the network. Separating constraint satisfaction from dynamics learning keeps the prediction from violating the instantaneous field equations. On the plasma-wakefield acceleration data, applying the constraint operator reduces violations of Gauss's law and the divergence-free magnetic-field constraint from the 10⁻²–10⁻¹ range to the 10⁻⁶ range with essentially unchanged relative error.
Perspective
The report covers two WarpX-generated 3D particle-accelerator benchmarks: bunched-beam transport in the electrostatic regime and plasma-wakefield acceleration in the electromagnetic regime, evaluated by relative error, banded spectral error, constraint violation, and inference time. The intended readers are researchers and engineers working on accelerator modeling, surrogate models, or scientific machine learning. The stated next step is to extend the approach toward a foundation model for accelerator simulation across machines, control settings, and initial or boundary conditions, and to use the learned dynamics for beam-state estimation and control.
On bunched-beam transport the generative residual raises overall relative error from 0.144 to 0.161, so spectral improvement trades against pointwise error. The latent-backbone comparison shows a U-Net-based operator is competitive on single-channel bunched-beam transport but degrades markedly on multi-channel plasma-wakefield acceleration, so the advantage of global attention is tied to that strongly coupled setting. The constraint operator satisfies Gauss's law exactly in the interior while the boundary shell remains unconstrained.