Time-inversion of spatiotemporal beam dynamics using uncertainty-aware latent evolution reversal.

Area of Science:

• Accelerator Physics
• Computational Physics
• Machine Learning

Background:

• Charged particle dynamics in electromagnetic fields present complex spatiotemporal challenges.
• Existing physics-based simulators are computationally intensive, hindering online inverse problem solving.
• Estimating upstream six-dimensional (6D) phase space from downstream measurements is a critical inverse problem in accelerators.

Purpose of the Study:

• To develop a computationally efficient model for temporal inversion of charged particle beam dynamics.
• To enable accurate prediction of upstream 6D phase space from downstream measurements.
• To incorporate and propagate uncertainty in predictions for robust accelerator modeling.

Main Methods:

• A two-step, self-supervised deep learning framework combining a conditional variational autoencoder (CVAE) and a long short-term memory (LSTM) network.
• CVAE projects 6D phase space into a lower-dimensional latent distribution.
• LSTM autoregressively learns inverse temporal dynamics within the latent space.

Main Results:

• The coupled CVAE-LSTM model successfully predicts 6D phase space projections across upstream accelerating sections.
• The model utilizes single or multiple downstream measurements as input for predictions.
• Aleatoric uncertainty from input data is captured and propagated, providing uncertainty bounds for upstream predictions.

Conclusions:

• The proposed reverse latent evolution model offers an efficient solution for inverse problems in charged particle beam dynamics.
• The framework demonstrates robustness to input perturbations by effectively propagating uncertainty.
• This approach enhances the utility of deep learning for online accelerator analysis and control.