THP043
ポスターセッション② 8月7日 61A 13:00-15:00
リングサイクロトロンにおける多目的RF空洞最適化のための変分オートエンコーダによる次元削減
Dimensionality Reduction with Variational Autoencoders for Multiobjective RF Cavity Optimization in Ring Cyclotron
○Shali Ahsani Hafizhu, 福田 光宏, 依田 哲彦, 神田 浩樹, 松田 洋平, 荘 浚謙, 趙 航, 松井 昇大朗, 井村 友紀, 板倉 菜美, 石畑 翔, 辻坂 匡（大阪大学）
○Shali Ahsani Hafizhu, Mitsuhiro Hukuda, Tetsuhiko Yorita, Hiroki Kanda, Yohei Matsuda, Tsun Him Chong, Hang Zhao, Shotaro Matsui, Tomoki Imura, Nami Itakura, Sho Ishihata, Tasuku Tsujisaka（The University of Osaka）
A variational autoencoder-assisted multi-objective optimization scheme for the design of RF cavities in a ring cyclotron is studied. The cavity geometry is parameterized using Non-Uniform Rational B-Splines (NURBS), with control point positions and weights acting as the design space. A variational autoencoder is trained to learn low-dimensional latent representation of NURBS parameters, allowing optimization processes to run efficiently. Multi-objective optimization is performed in the latent space using differential evolution, with objective functions evaluated using an ensemble of neural network surrogate models trained on eigenmode simulations from Ansys HFSS. Both the surrogate model and the variational autoencoder are periodically retrained during the optimization process using new simulation data, improving the accuracy. In this presentation, we demonstrate that dimensionality reduction of the design space using a variational autoencoder can reduce the number of function evaluations required by differential evolution, potentially improving overall optimization efficiency.