Deep Learning-Optimized Dielectric Laser Accelerators: High-Gradient Performance and Cascaded Photonic Chip

摘要: Dielectric laser accelerators (DLAs) can achieve acceleration gradients exceeding those of conventional radio-frequency accelerators by one to two orders of magnitude. Existing DLA design approaches rely heavily on empirical parameter tuning and single-variable optimization, which fundamentally constrains performance enhancement. Here, a new optimization strategy for DLA structures is proposed based on the Gated Adaptive Network for Deep Automated Learning of Features (GANDALF). This framework integrates key parameters such as geometric configurations, material properties, and optical field characteristics into a comprehensive analysis. By accurately predicting particle energy gain, the structural parameters are optimized, significantly improving DLA performance. The proposed approach outperforms traditional computational methods, particularly for nonperiodic structures, enabling continuous particle acceleration. The GANDALF model demonstrates high accuracy, robustness, and adaptability, yielding an average acceleration gradient of 2.8 GV/m (Y2O3), enabling sustained acceleration in the majority of the acceleration channel, with a beam spot radius of 3.13 µm. Additionally, a cascaded DLA design concept is introduced and validated, paving the way for extended acceleration lengths on photonic chips.