Speaker
Description
The design of fourth-generation light sources is hindered by high dimensionality and strong nonlinearities, rendering the traditional paradigm of "manual matching followed by stepwise fine-tuning" inefficient and lacking in physical interpretability. This paper proposes a modular optimization framework that fuses physics priors with statistical learning to achieve synergistic optimization of linear optics and nonlinear dynamics. First, utilizing Twiss parameter evolution as an intermediate physical representation, we employ a physics-prior screening mechanism driven by linear transport combined with the Covariance Matrix Adaptation Evolution Strategy (CMA-ES). This approach successfully identifies stable periodic solutions with significantly reduced natural emittance within minutes, demonstrating reproducible efficiency gains over manual initializations. Based on the generated dataset, machine learning classifiers are trained to perform high-confidence pruning of the solution space, effectively retaining high-quality solutions. Finally, constructing a local trust region around these "promising solutions," we introduce the Sequential Model-based Algorithm Configuration (SMAC) strategy based on Random Forests for refined iteration. This method provides an efficient and intelligent pathway for complex, high-dimensional lattice design.
| In which format do you inted to submit your paper? | Word |
|---|