发表机构
Nikhef; Halmstad University; IFIC, CSIC-Universitat de València; Leiden Observatory, Leiden University（NIKHEF荷兰国家核与亚原子物理研究所; 哈尔姆斯塔德大学; 伊菲克研究所，西班牙国家研究委员会-巴伦西亚大学; 莱顿天文台，莱顿大学）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
Merlin是一个基于模拟推断的3x2点宇宙学分析流程，结合MNRE与cloelib预测，在50维参数空间中实现与嵌套采样器一致的后验，同时将计算成本降低两个数量级，并支持零额外评估的灵活分析选择。
AI 中文摘要
我们提出Merlin，一个基于模拟推断（SBI）的流程，用于对3x2点汇总统计量进行宇宙学分析：宇宙剪切、星系聚集及其互相关功率谱。我们的方法将边际神经比率估计（MNRE）与来自cloelib库的3x2点角功率谱预测相结合，尽管该流程易于扩展到其他宇宙学库。我们在一个代表第四阶段（Stage-IV）测光巡天的现实环境中演示了这一流程，其参数空间为50维，描述了宇宙学及广泛的系统效应。我们发现后验分布与嵌套采样器Nautilus的结果高度一致，同时将所需的CPU小时数减少了两个数量级。由于训练数据的生成与推断网络的训练是解耦的过程，单个模拟库可被重复用于在不同分析选择下进行推断，进一步提高了模拟器效率。我们通过应用尺度截断、改变巡天面积以及从数据向量中移除宇宙剪切（2x2pt）来展示这种灵活性，所有这些均在零额外模型评估下完成，而基于采样的方法则需要为每种情况付出高昂的重新运行成本。Merlin代码已在GitHub上公开提供。
英文摘要
We present Merlin, a simulation-based inference (SBI) pipeline to perform cosmological analyses of 3x2pt summary statistics: cosmic shear, galaxy clustering and the cross-correlation power spectra. Our approach combines Marginal Neural Ratio Estimation (MNRE) with 3x2pt angular power spectra predictions from the cloelib library, although the pipeline is readily extensible to other cosmology libraries. We demonstrate this pipeline on a realistic setting representative of a Stage-IV photometric survey, with a 50-dimensional parameter space describing cosmology and a wide range of systematic effects. We find posteriors that are in excellent agreement with the nested sampler Nautilus, while reducing the number of required CPU-hours by two orders of magnitude. Because the generation of training data and the training of inference networks are decoupled processes, a single simulation bank can be reused to perform inference under different analysis choices, further improving the simulator-efficiency. We illustrate this flexibility by applying scale cuts, varying the survey area, and removing cosmic shear from the data vector (2x2pt), all at zero extra model evaluations, whereas sampling-based methods need costly re-runs for each case. The Merlin code is publicly available on GitHub.
Comments15 pages, 9 figures, 1 table. The Merlin code is available at https://github.com/Alexandra-Wernersson/merlin
