Latent evolution models

Spatiotemporal learning of charged-particle beam dynamics (in particle accelerators)

The 6D phase space (x, y, z, p_x, p_y, p_z) of charged-particle bunches evolves under the influence of electromagnetic fields along the accelerator. This can be viewed as a spatiotemporal dynamical system in which parameters such as RF-cavity amplitude and phase and magnet strength modulate the charged-particle beam.

1. In this paper, we propose a two-step unsupervised deep-learning framework called the Conditional Latent Autoregressive Recurrent Model (CLARM) for learning forward spatiotemporal dynamics.
   • The model can generate phase space at various accelerator modules by sampling and decoding the structured latent space representation.
   • The model also forecasts future states (downstream states) of charged particles from past states (upstream states). Learn more on the project page.

1. In this paper, we use a reverse latent evolution model (CLARM is a special case of a broader LEM) to solve the inverse problem of predicting 6D phase-space projections across all upstream sections from downstream phase-space measurements.
   • The proposed model also captures the aleatoric uncertainty of the high-dimensional input data within the latent space.
   • The uncertainty is propagated through the temporal learner in the latent space to estimate bounds for all upstream predictions, demonstrating robustness to in-distribution variations in the input data.
2. In this paper, we use a classifier-pruned Bayesian optimizer for efficient exploration within a temporally structured latent space.
   • CBOL-Tuner adaptively searches and filters the latent space for optimal solutions (i.e., RF settings for minimal beam loss in the accelerator).
   • CBOL-Tuner demonstrates superior performance in identifying multiple optimal settings and outperforms alternative global optimization methods.