Collaborative Research: CDS&E: AI-Enhanced Exascale Simulations Of The Earliest Galaxies

最近在使用深度神经网络训练的代理模型加速基于物理的科学模拟方面取得了令人印象深刻的进展，这刺激了这一项目，即使用这种模型来加速对第一批星系形成的宇宙学模拟，包括人口III恒星的化学预浓缩。代理模型将在高分辨率自适应网格细化(AMR)模拟上进行训练，并通过将其合并到较低空间分辨率的AMR模拟中进行验证。一旦得到验证，这种方法将成为Enzo-E的一部分，Enzo-E是广泛使用的Enzo代码的高度可伸缩版本。一个主要目标是确定这种寻路方法的有效性和有用性，通过将数值模拟与机器学习相结合来解决天文学中的一个前沿问题。这种技术方法可以为其他天体物理情况定制，在这些情况下，大范围的尺度阻止了直接的数字攻击。它还将有助于促进在多个学科中采用代理模型，并导致改进的科学知识深度学习神经网络设计。代理模型将被称为Pop3Net，由几个定制设计的卷积深度神经网络组成，这些网络预测POP III恒星形成的时间和地点，以及其他模拟辐射和化学反馈效应的网络。Pop3Net将使用最先进的全物理模拟第一批星系的形成进行培训。混合的Enzo-E+Pop3Net代码将被用来模拟大型高红移星系系综的形成，使用Powderday程序包计算合成观测，以与詹姆斯·韦伯太空望远镜和南希·格雷斯·罗马太空望远镜未来的数据进行比较。这部分工作将检验星系属性对第一批星系的化学富集史有多敏感，包括超音速重子-暗物质流动速度效应。这一奖项反映了NSF的法定任务，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Impressive recent advances in accelerating physics-based scientific simulations using deep neural network-trained surrogate models have stimulated this project, to use such models to accelerate cosmological simulations of the formation of the first galaxies, including chemical pre-enrichment by Population III stars. The surrogate model will be trained on high-resolution adaptive mesh refinement (AMR) simulations and validated by incorporating it into AMR simulations of lower spatial resolution. Once validated, this approach will become part of Enzo-E, the highly scalable version of the widely used Enzo code. A principal goal is to establish the validity and usefulness of this pathfinding approach, addressing a frontier problem in astronomy by combining numerical simulations with machine learning. This technical methodology can be customized for other astrophysical situations where a vast range of scales inhibits a direct numerical assault. It will also help to stimulate adoption of surrogate models in multiple disciplines, and lead to improved science-informed deep learning neural network designs.The surrogate model, to be called Pop3Net, is composed of several convolutional deep neural networks of custom design which predict when and where Pop III stars form, and other networks that emulate radiative and chemical feedback effects. Pop3Net will be trained using state-of-the-art full-physics simulations of the formation of the first galaxies. The hybrid Enzo-E+Pop3Net code will be used to simulate the formation of large ensembles of high redshift galaxies, computing synthetic observations using the Powderday package, to compare with future data from the James Webb Space Telescope and the Nancy Grace Roman Space Telescope. This part of the work will examine how sensitive galaxy properties are to the chemical enrichment history of the first galaxies, including supersonic baryon-dark matter streaming velocity effects.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.