ELMO：用于快速台基线性稳定性预测的不确定性感知模拟代理工作流
ELMO: An Uncertainty-Aware Simulation-to-Surrogate Workflow for Fast Pedestal Linear-Stability Prediction
- Lawrence Livermore National Laboratory（劳伦斯利弗莫尔国家实验室）
- General Atomics（通用原子能公司）
- Oak Ridge National Laboratory（橡树岭国家实验室）
- Columbia University（哥伦比亚大学）
机构由 AI 辅助整理，请以论文原文为准。
AI总结：
本研究提出ELMO工作流，集成多技术实现快速台基线性稳定性预测，其GPR代理预测精度高、速度快，大幅降低计算成本。
AI中文摘要：
快速预测台基线性稳定性对探索托卡马克运行空间、不确定性量化及未来模型辅助控制至关重要，但使用BOUT++的模式分辨磁流体动力学稳定性计算计算成本高昂。本文提出了ELMO（Edge Learning and Modeling Orchestrator，边界学习与建模编排器）的聚焦实现，作为一种不确定性感知的模拟代理工作流，集成了平衡生成、场对齐网格构建、大规模BOUT++计算、自动化任务执行与数据缩减以及高斯过程回归（GPR）。针对单个DIII-D等离子体位形，在请求的7992个配置中，有3869个完成了平衡重建、网格生成、稳定性计算和质量控制。每个保留的平衡在16个环向模数（n=5至80，Δn=5）下，采用理想MHD和理想加抗磁模型进行评估，产生了123808个模式分辨计算。利用8个台基特征，GPR代理预测两个16模式增长率谱，并提供潜在后验不确定性估计。在5个独立测试实现中，最大增长率预测对理想MHD的R²达到0.978±0.013，对理想加抗磁物理的R²达到0.966±0.009。该代理再现了谱形状和主导不稳定模式。校准诊断表明，后验不确定性对相对采集有用，但分散不足，不应被解释为校准预测区间。在单个CPU核心上预测全部32个输出约需20毫秒，而使用128个CPU核心进行相应BOUT++扫描约需21分钟，实现了6.3×10⁴倍的挂钟速度提升和8.1×10⁶倍的计算成本降低。
英文摘要：
Rapid prediction of pedestal linear stability is important for exploring tokamak operating space, uncertainty quantification, and future model-informed control, but mode-resolved magnetohydrodynamic stability calculations using BOUT++ are computationally expensive. We present a focused implementation of ELMO--the Edge Learning and Modeling Orchestrator--as an uncertainty-aware simulation-to-surrogate workflow integrating equilibrium generation, field-aligned mesh construction, large-scale BOUT++ calculations, automated campaign execution and data reduction, and Gaussian Process Regression (GPR). For a single DIII-D plasma shape, 3,869 of 7,992 requested configurations completed equilibrium reconstruction, mesh generation, stability calculation, and quality control. Each retained equilibrium was evaluated at sixteen toroidal mode numbers, $n=5$--80 with $Δn=5$, using ideal-MHD and ideal-plus-diamagnetic models, producing 123,808 mode-resolved calculations. Using eight pedestal features, the GPR surrogate predicts two sixteen-mode growth-rate spectra with latent posterior uncertainty estimates. Across five independent test realizations, the maximum-growth-rate prediction achieved $R^2=0.978\pm0.013$ for ideal MHD and $R^2=0.966\pm0.009$ for ideal-plus-diamagnetic physics. The surrogate reproduces the spectral shape and dominant unstable mode. Calibration diagnostics indicate that posterior uncertainties are useful for relative acquisition but are underdispersed and should not be interpreted as calibrated prediction intervals. Prediction of all 32 outputs requires about 20 ms on one CPU core, compared with about 21 min using 128 CPU cores for the corresponding BOUT++ scan, giving a $6.3\times10^4$-fold wall-clock speedup and an $8.1\times10^6$-fold reduction in computational cost.
