The JAG inertial confinement fusion simulation dataset for multi-modal scientific deep learning. In Lawrence Livermore National Laboratory (LLNL) Open Data Initiative

The goal of this project is to build better surrogate models for Inertial Confinement Fusion (ICF) using neural networks. Particularly, we are interested in replicating the behavior of the JAG 1D Semi-analytic simulator for ICF. The JAG model has been designed to give a rapid description of the observables from ICF experiments, which are all generated very late in the implosion. In this way the very complex and computationally expensive transport models needed to describe the capsule drive can be avoided, allowing a single solution in $ ilde$ seconds. The trade-off is that JAG inputs do not relate to actual experimental observables, rather the state of the implosion once the laser drive has switched off. At that point, an analytic description of the spatial profile inside the hotspot can be found [1,2], leaving only a set of coupled ODEs describing the temporal energy balance inside the entire problem which can be solved easily [3]. The various terms in the energy balance equation relate to different physics processes (radiation, electron conduction, heating by alpha particles, etc), making JAG useful for investigating the role of various potentially uncertain physics models. Combined with a thin-shell model describing the 3D hydrodynamic evolution of the hotspot [4], JAG has a detailed description of the spatial and temporal evolution of all thermodynamic variables which can be post-process to predict a full range of experimental observables. References: 1. Betti et al., Physics of Plasmas 9, 2277 (2002) 2. Springer et al., EPJ Web. Conferences 59:04001 (2013) 3. Betti et al., Physical Review Letters 114:255003 (2015) 4. Ott et al., Physical Review letters 29:1429 (1995)本项目的目标是利用神经网络构建更优质的惯性约束聚变（Inertial Confinement Fusion, ICF）替代模型。我们尤其关注复现用于惯性约束聚变的JAG 1D半解析模拟器的行为。 JAG模型旨在快速给出惯性约束聚变实验的可观测量描述，此类可观测量均产生于内爆的极晚阶段。借此可规避描述胶囊驱动所需的极端复杂且计算成本高昂的输运模型，仅需极短时间即可完成单次求解。其权衡之处在于，JAG的输入并不对应实际的实验可观测量，而是激光驱动关闭后内爆的状态。此时可通过解析方法得到热斑内部的空间分布描述[1,2]，仅余下一组耦合常微分方程（Ordinary Differential Equations, ODE）来描述整个问题内部的时间尺度能量平衡，该方程可被轻松求解[3]。 能量平衡方程中的各项对应不同的物理过程（辐射、电子热传导、α粒子加热等），这使得JAG可用于探究各类存在不确定性的物理模型所发挥的作用。结合用于描述热斑三维流体力学演化的薄壳模型[4]，JAG可对所有热力学变量的时空演化进行详细描述，后续可通过后处理预测全范围的实验可观测量。 参考文献： 1. Betti 等，《等离子体物理学》，第9卷，2277页（2002年） 2. Springer 等，《欧洲物理学会期刊会议论文集》，第59卷：04001（2013年） 3. Betti 等，《物理评论快报》，第114卷：255003（2015年） 4. Ott 等，《物理评论快报》，第29卷：1429（1995年）