Predicting the Expansion of Supernova Shells Using Deep Learning toward Highly Resolved Galaxy Simulations
Predicting the Expansion of Supernova Shells Using Deep Learning toward Highly Resolved Galaxy Simulations
复制标题
使用深度学习预测超新星壳的膨胀以实现高分辨率的星系模拟
DOI：
10.1017/s1743921322001739
复制
发表时间：
2023
期刊：
影响因子：
--
通讯作者：
Makino Junichiro
中科院分区：
文献类型：
--
作者：
Hirashima Keiya;Moriwaki Kana;Fujii Michiko;Hirai Yutaka;Saitoh Takayuki;Makino Junichiro
并行计算的负载不平衡和通信开销是银河系模拟的关键瓶颈。一个成功的方法来提高天文模拟的可扩展性是一个哈密顿分裂方法，它需要确定这样的区域集成与更小的时间步长比全球时间步长整合整个星系。在星系模拟的情况下，超新星（SN）壳内的区域需要最小的步骤。我们开发了深度学习模型来预测在一个全局步骤中受SN壳扩张影响的区域。此外，我们使用图像处理识别出具有小时间步长的颗粒。与使用Sedov-Taylor解决方案的分析方法相比，我们可以使用我们的方法以更高的识别率（平均88%至98%）和更低的“非目标”与“目标”分数（平均6.4至5.5）识别目标颗粒。我们使用哈密顿分裂和深度学习的方法将提高极高分辨率星系模拟的性能。
The load imbalance and communication overhead of parallel computing are crucial bottlenecks for galaxy simulations. A successful way to improve the scalability of astronomical simulations is a Hamiltonian splitting method, which needs to identify such regions integrated with smaller timesteps than the global timestep for integrating the entire galaxy. In the case of galaxy simulations, the regions inside supernova (SN) shells require the smallest steps. We developed the deep learning model to forecast the region affected by the SN shell’s expansion during one global step. In addition, we identified the particles with small timesteps using image processing. We can identify target particles using our method with a higher identification rate (88 % to 98 % on average) and lower “non-target”-to-“target” fraction (6.4 to 5.5 on average) compared to the analytic approach with the Sedov-Taylor solution. Our method using Hamiltonian splitting and deep learning will improve the performance of extremely high-resolution galaxy simulations.