Learning a Latent Pulse Shape Interface for Photoinjector Laser Systems

Alexander Klemps; Denis Ilia; Pradeep Kr. Banerjee; Ye Chen; Henrik Tünnermann; Nihat Ay

cs.LGFeb 19, 2026

Learning a Latent Pulse Shape Interface for Photoinjector Laser Systems

Authors: Alexander Klemps, Denis Ilia, Pradeep Kr. Banerjee, Ye Chen, Henrik Tünnermann, Nihat Ay

Organizations: Institute for Data Science Foundations, Hamburg University of Technology, Germany · Deutsches Elektron-Synchrotron (DESY), Hamburg, Germany

Abstract

Controlling the longitudinal laser pulse shape in photoinjectors of Free-Electron Lasers is a powerful lever for optimizing electron beam quality, but systematic exploration of the vast design space is limited by the cost of brute-force pulse propagation simulations. We present a generative modeling framework based on Wasserstein Autoencoders to learn a differentiable latent interface between pulse shaping and downstream beam dynamics. Our empirical findings show that the learned latent space is continuous and interpretable while maintaining high-fidelity reconstructions. Pulse families such as higher-order Gaussians trace coherent trajectories, while standardizing the temporal pulse lengths shows a latent organization correlated with pulse energy. Analysis via principal components and Gaussian Mixture Models reveals a well behaved latent geometry, enabling smooth transitions between distinct pulse types via linear interpolation. The model generalizes from simulated data to real experimental pulse measurements, accurately reconstructing pulses and embedding them consistently into the learned manifold. Overall, the approach reduces reliance on expensive pulse-propagation simulations and facilitates downstream beam dynamics simulation and analysis.

Figures & tables

Figure 1: Layout of the modeled photoinjector laser system. Initially in the front-end created pulses undergo nonlinear transformations during propagation to the photoinjector system.

Parameter	Distribution	Unit
$A(\omega)$	$\mathcal{U}(\{S,P,F,T,G\})$	-
$p_{T}$	$\mathcal{U}(\{1,2,4\})$	-
$p_{G}$	$\mathcal{U}(\{1,2,3,4,5,10\})$	-
$\sigma_{t}$	$\mathcal{U}([2,\,40])$	ps
$\varphi_{2}$	$\mathcal{N}(0,\frac{100}{3\sigma_{t}^{2}})$	$\text{s}^{2}$
$\varphi_{3}$	$\mathcal{N}(0,\frac{100}{3\sigma_{t}^{3}})$	$\text{s}^{3}$

Table 1: Pulse parameters and sample ranges.

Figure 2: Representative simulated pulse pair before and after propagation through the fiber system. The input pulse (blue) and the propagated pulse (red) are shown as normalized intensity profiles after centroid and support alignment. Propagation leads to temporal reshaping due to dispersion and nonlinear effects, highlighting the variety of pulse morphologies within the dataset.

Figure 3: Structure of the convolutional encoder and decoder forming the Wasserstein Autoencoder. Each block consists of one-dimensional convolutions followed by batch normalization and Leaky-ReLU activations. Residual skip connections are included to stabilize training and preserve fine temporal details. Layer output sizes are shown below each block, and kernel sizes are indicated by numbers inside the convolutional layers. The decoder’s final output layer uses a hyperbolic tangent activation to constrain the reconstruction to the normalized signal range.

Figure 4: Comparison between input pulse profiles from the test dataset (blue) and their corresponding reconstructions (orange) obtained with the trained Wasserstein Autoencoder. The reconstructions capture not only the overall pulse envelope but also higher-frequency oscillatory components, demonstrating the model’s ability to preserve fine temporal structure.

Figure 5: Visualization of the learned WAE latent space projected onto the first two principal components. Each point represents an encoded pulse, color-coded by normalized pulse energy. The overlaid trajectories correspond to parameterized pulse families shown on the left, whose latent embeddings evolve smoothly along coherent paths across the manifold.

Figure 6: Latent interpolation between two pulse profiles. Left: source and target pulses with a

W_{2}

distance of

1.080

. Right: sequence of intermediate decoded pulses from latent interpolations via a discrete linear interpolation path

\Gamma_{0}

(blue) and a

W_{2}

-optimized path

\Gamma^{*}

(orange) using

N=10

waypoints. Both yield smooth transitions w.r.t. the

W_{2}

distance, where the optimized interpolation path attains a near-optimal length of

L(\Gamma^{*})=1.084

compared to slightly higher

L(\Gamma_{0})=1.195

Figure 7: Comparison between experimentally measured and preprocessed pulse profiles (colored solid) and corresponding reconstructions (black dotted) produced by the trained Wasserstein Autoencoder. Each row to the right resembles representative samples from recorded sequences of pulses measured at the photoinjector laser system, in which optical filter parameters were varied through a gradient-based optimization to morph the initial profile (triangle/first column) into a desired target shape (square/last column). Pulses in the first two sequences are infrared ones, whereas the third one contains UV pulses. The reconstructions closely match the measured data, demonstrating high accuracy and generalization to real experimental conditions.

Model	Simulated Test Data			Experimental Data
	MSE $\downarrow$	SNR $\uparrow$	COR $\uparrow$	MSE $\downarrow$	SNR $\uparrow$	COR $\uparrow$
$\beta$ -VAE 0.5	6.58e-3	14.74	0.19	8.17e-3	15.29	0.18
$\beta$ -VAE 0.7	8.17e-3	13.89	0.18	1.19e-2	13.77	0.16
$\beta$ -VAE 1.0	9.95e-3	12.86	0.16	1.33e-2	13.03	0.14
WAE	2.40e-4	28.99	0.67	3.70e-4	28.38	0.78

Table 2: Comparison of

\beta

-VAEs for different

\beta

and WAE on simulated test and experimental datasets.

\uparrow

denotes that higher values indicate better performance, while

\downarrow

indicates that lower values are preferable.

Figure 8: Linear correlations between the first 5 latent principal components and

\varepsilon_{xy}

. Oscillations between

2.5\,\mathrm{m}

and

4.5\,\mathrm{m}

arise from an accelerating module within the accelerator lattice, causing fluctuations in transverse momenta.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 9: Left panel: PCA projection of the train and test embeddings colored w.r.t. the fit GMM components and their means. Middle panel: Decoded means of the GMM components. Correspondence to components is indicated by the coloring from the left panel. Right panel: Pairwise normalized 2-Wasserstein distances between GMM components.

Figure 10: Pulse shapes decoded from latent samples drawn via a latent Gaussian mixture model (orange) are interpreted as probability density functions for electron emission time. Inverse transform sampling produces zero-centered emission time histograms (blue) of

200\,000

particles each that faithfully match the decoded pulse profiles.

Figure 11: Left panel: Distribution of reconstruction MSE on the test set for different values of $\beta$ and the WAE. Middle panel: Reconstruction of a representative flattop pulse, illustrating progressive loss of sharp temporal features with increasing $\beta$ , while the WAE preserves the target shape. Right panel: PCA projection of the learned latent spaces with trajectories of selected pulse families, showing degraded geometric organization for $\beta$ -VAEs compared to the smooth and interpretable structure obtained with the WAE.

Representation Learning

Representation learning develops data transformations that expose useful structure for prediction, transfer, generation, and control.

See all 798 papers

Feb 26, 2026

Advancing accelerator virtual beam diagnostics through latent evolution modeling: an integrated solution to forward, inverse, tuning, and UQ problems

Virtual beam diagnostics relies on computationally intensive beam dynamics simulations where high-dimensional charged particle beams evolve through the accelerator. We propose Latent Evolution Model (LEM), a hybrid machine learning framework with an autoencoder that projects high-dimensional phase spaces into lower-dimensional representations, coupled with transformers to learn temporal dynamics in the latent space. This approach provides a common foundational framework addressing multiple interconnected challenges in beam diagnostics. For forward modeling, a Conditional Variational Autoencoder (CVAE) encodes 15 unique projections of the 6D phase space into a latent representation, while a transformer predicts downstream latent states from upstream inputs. For inverse problems, we address two distinct challenges: (a) predicting upstream phase spaces from downstream observations by utilizing the same CVAE architecture with transformers trained on reversed temporal sequences along with aleatoric uncertainty quantification, and (b) estimating RF settings from the latent space of the trained LEM using a dedicated dense neural network that maps latent representations to RF parameters. For tuning problems, we leverage the trained LEM and RF estimator within a Bayesian optimization framework to determine optimal RF settings that minimize beam loss. This paper summarizes our recent efforts and demonstrates how this unified approach effectively addresses these traditionally separate challenges.

Mahindra Rautela, Alexander Scheinker

Los Alamos National Laboratory, Los Alamos, NM, US

Latent Dynamics Modeling Inverse Problems

Ask PDF

Oct 10, 2026

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

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 Generative Latent Operator for Beam Evolution, modeling the nonlinear, control-dependent, multiscale evolution of 3D beam--field dynamics covering

10^7

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-

L_2

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.

Haohan Zou, Alexander Scheinker, Yuanyuan Shi

University of California San Diego · Los Alamos National Laboratory

Latent Dynamics Modeling Physics-Informed Generative Modeling PDE Surrogate Modeling Physics-Informed Neural Operators

Ask PDF

Oct 5, 2026

Generative World Models Enable Predictive Control of Laser Melt Pool Dynamics

World models, which learn how environments respond to actions, are emerging as a powerful paradigm for planning through imagined futures, transforming decision-making across games, robotics and autonomous driving. Bringing this capability to manufacturing could enable process decisions on timescales inaccessible to high-fidelity simulation. Here we introduce a generative world model for localized highly dynamic laser melt pool that predicts evolution from histories of temperature and phase morphology under candidate actions. Its generative latent dynamics capture the effects of unresolved melt flow, enabling more accurate recursive rollouts than deterministic regressors under transient laser inputs. Because the learned dynamics are differentiable, the model can serve directly as a predictive control plant. Gradients through imagined futures optimize laser schedules that regulate melt-pool depth over previously unseen geometry, path, initialization. We further distil this optimization into an amortized policy that produces control actions in a single forward pass, providing a proof of concept for real deployment on machines.

Yiyang Yan, Markus Bambach, Mohamadreza Afrasiabi

Advanced Manufacturing Lab, ETH Zürich, Leonhardstrasse 21, Zürich 8092, Switzerland

World Model Learning Model Predictive Control Additive Manufacturing

Ask PDF

Abstract

Representation Learning

Explore similar work

Advancing accelerator virtual beam diagnostics through latent evolution modeling: an integrated solution to forward, inverse, tuning, and UQ problems

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

Generative World Models Enable Predictive Control of Laser Melt Pool Dynamics