Complex dynamical systems, such as particle accelerators, require tuning over high-dimensional, nonlinear state and parameter spaces while experimental measurements and high-fidelity simulations can be expensive. Learned latent representations provide a compact domain for such optimization, but poorly supported regions of the learned manifold may decode into unrealistic physical states and yield deceptively favorable objectives. To address this challenge, we propose the Classifier-pruned Bayesian Optimization-based Latent-space Tuner (CBOL-Tuner), which performs Bayesian optimization over a temporally structured latent representation of 6D beam phase-space dynamics. The CBOL-Tuner integrates a conditional variational autoencoder for latent space representation, a long short-term memory network for temporal dynamics, a lightweight neural network for parameter estimation, and a classifier-pruned Bayesian optimizer to adaptively search and filter the latent space for optimal solutions. This framework enables feasibility-aware exploration of learned scientific representations for accelerator tuning.
Figures & tables
Figure 1: Architecture of the CBOL-Tuner for beam optimization in particle accelerators. A. Modular latent evolution model, comprising encoder and decoder of CVAE, forecaster (LSTM) and estimator (DNN) for learning spatiotemporal dynamical system, B. Classifier-pruning using ResNet-50 to classify trajectories into 48 classes (modules), C. Bayesian optimization over the latent space of CVAE. (*) denotes components trained prior to the optimization stage and held fixed during Bayesian optimization..
Figure 2: 3 out of 15 2d projections of 6d phase space of charged particle beam in the LANSCE linear accelerator. Accelerating modules - 1 to 4 are 201 MHz drift tube linac (DTL) and 5 to 48 are 805 MHz coupled cavity linac (CCL). The beam serves various scientific areas like isotope production facility (IPF), ultra-cold neutrons (UCN), proton radiography (PRAD), weapons neutron research (WNR), proton storage ring (PSR).
Figure 3: Forecasting results: Forecasted projections (shown E−ϕ only) across different modules given first four projections as inputs. The original projection is presented against the forecasted along with the absolute difference between both. The MSE for the entire training and test set are plotted as a shaded plot where the the central line is the mean and the boundaries of the region defines the standard deviation.
Figure 4: Target vs parameter search space for Bayesian optimization and classifier-pruned BO. Classifier-pruning eliminates latent trajectories that do not satisfy the selection criterion.