Rubin Observatory Deploys AI to Cut Dark Energy Modeling to Minutes

鲁宾天文台应用AI将暗能量建模缩短至数分钟 神经后验估计将引力透镜建模时间从数天缩短至数分钟，使天文学家能够在数千个系统中扩展宇宙膨胀测量规模，同时通过分层推断消除总体偏差。 Phil Marshall 概述了 Vera C. Rubin 天文台将如何通过为期 10 年的空间与时间遗产巡天（LSST），应对包括宇宙加速和暗物质在内的核心宇宙学谜团。该望远镜配备了一台巨大的 32 亿像素相机，将对整个南天进行反复成像，编目 200 亿个星系、170 亿颗恒星以及数十亿个瞬态现象。为了应对海量数据洪流，该天文台通过 Rubin 科学平台利用云计算、通过 Zooniverse 开展公民科学众包，并利用神经网络的基于模拟的推断（Simulation-Based Inference）来快速分析时间延迟引力透镜。该项目从根本上推动了天文学研究的普惠化，为全球机构提供平等的数据访问权限，并促进广泛的 STEM 教育。 宇宙加速仍是物理科学中最大的谜团，要精确绘制其图景，需要使用 Ia 型超新星等标准烛光来测量宇宙膨胀历史。 (0:04) 将宇宙学测量规模扩大到 10 万颗超新星样本，可同时推动太阳系科学、星系结构和暗物质测绘等领域的广泛发现。 (3:00) 位于 Cerro Pachón 的 Vera C. Rubin 天文台开展为期 10 年的空间与时间遗产巡天（LSST），单次曝光即可发现数千颗小行星、变星、超新星和活动星系核。 (6:02) 在 SLAC 制造的汽车大小的 LSST 相机拥有 32 亿像素的焦平面，并在六个光学和红外滤光片上实现覆盖满月面积 45 倍的超宽视场。 (9:01) 在十年间，Rubin 天文台将拍摄 200 万次曝光，产生数百 PB 的处理后数据，编制一份涵盖 200 亿个星系、170 亿颗恒星和 600 万个太阳系天体的普查名录。 (12:59)

The NSF-DOE Vera C. Rubin Observatory is designed to investigate the accelerating expansion of the universe and survey the Southern sky through its decade-long Legacy Survey of Space and Time (LSST). According to project astrophysicist Phil Marshall, coupling the world's largest digital camera with cloud data infrastructure and simulation-based machine learning enables observational cosmology at unprecedented statistical scale. LSST Camera Specifications and Imaging Scale (9:01) Built at the SLAC National Accelerator Laboratory, the car-sized LSST Camera features a 3,200-megapixel (3.2-gigapixel) focal plane with a field of view covering roughly 45 times the apparent area of the full Moon. Outfitted with a carousel of six broadband optical and near-infrared filters, the instrument captures wide-field images capable of resolving immense cosmological detail. Project figures indicate that a native-resolution display of a single exposure would require 400 high-definition screens, while a printout at standard resolution would span two full tennis courts. A typical single exposure encompasses roughly 200 supernovae, 7,000 variable stars, 70 active galactic nuclei, and 400 solar system objects. Early commissioning over a brief four-night test run demonstrated this survey throughput by uncovering 2,000 previously uncataloged asteroids and serendipitously tracking the interstellar comet 3I/ATLAS without a targeted search. Petabyte-Scale Data Architecture and Cloud Access (12:59) Over its 10-year operational baseline atop Cerro Pachón in Chile, the observatory will accumulate approximately 2 million wide-field exposures across roughly 800 visits per pointing. The survey is slated to generate 30 petabytes of raw imagery, expanding into several hundred petabytes of calibrated, co-added data detailing 20 billion galaxies, 17 billion stars, and 6 million solar system bodies. Operating on a nightly data stream of 20 terabytes, processing hubs at SLAC and European partner facilities issue transient alerts within 60 seconds, deliver intermediate prompt products within 80 hours, and generate full annual reprocessed catalogs. To democratize data access across academic tiers, the Rubin Science Platform operates in the Google Cloud, allowing up to 10,000 astronomers to query petabyte-scale catalogs through browser interfaces without downloading data locally. The project plans to link with the Department of Energy's Genesis initiative to secure high-performance computing for intensive scientific modeling. Public engagement is integrated directly via Zooniverse projects such as 'Rubin Difference Detectives,' where citizen scientists classify image artifacts to provide labeled ground truth for machine learning filters. Testing Cosmic Acceleration with Gravitational Lensing (19:44) Cosmic expansion began accelerating roughly one billion years ago, contradicting expectations of gravitational deceleration and presenting what Marshall identifies as the primary mystery in physical science. In addition to measuring distances via Type Ia supernovae standard candles, cosmologists rely on strong gravitational lensing systems containing flickering active galactic nuclei (AGNs) to map cosmic history. When a massive foreground galaxy bends spacetime around a background AGN, light travels along paths of varying lengths and gravitational potentials, producing multiple images with measurable arrival time delays that yield geometric distances. Observational cosmology currently relies on approximately 10 well-characterized time-delay lens systems. By monitoring millions of galaxies across the southern sky, Rubin's survey cadence is projected to expand the time-delay lens sample into the thousands, significantly reducing statistical uncertainties on dark energy parameters governing cosmic acceleration. Simulation-Based Inference and Lens Modeling (21:47) Analyzing thousands of gravitational lenses presents a major computational bottleneck: traditional Markov Chain Monte Carlo (MCMC) methods require days of compute time per system to model the foreground galaxy's mass distribution. To address this, researchers apply Simulation-Based Inference (SBI) powered by Neural Posterior Estimation (NPE). Neural networks pre-trained on roughly 500,000 simulated lens systems estimate posterior parameter distributions in minutes rather than days, making population-scale analysis analytically feasible. Because systematic errors can emerge if the synthetic training distribution fails to represent real-world galaxy populations, the analysis uses hierarchical inference and normalizing flows. This framework enables astronomers to evaluate model sensitivity, correct for distribution shifts, and mathematically infer the accurate underlying prior distribution required to deliver unbiased cosmological measurements. Publicly funded by the NSF and DOE at an estimated average cost of roughly $5 per US taxpayer, the Vera C. Rubin Observatory also functions as an educational platform, logging approximately 80,000 curriculum uses shortly after operations began. The primary challenge moving forward remains validating neural network inference pipelines across incoming observational data to ensure systematic distribution shifts do not bias cosmological constraints.

SLAC国家加速器实验室的Phil Marshall介绍了由美国国家科学基金会（NSF）与能源部（DOE）联合资助的薇拉·鲁宾天文台（Vera C. Rubin Observatory）及其为期十年的“时空遗产调查”（LSST）。该项目旨在通过史无前例的大规模广角时域巡天与前沿机器学习推断技术，解决宇宙加速膨胀机制与暗物质分布等现代物理学核心谜题。 破译加速膨胀：超新星与引力透镜时延 (0:04) 现代宇宙学面临的一大核心困惑是宇宙膨胀在约十亿年前开始逆转常理、呈现加速状态，这一现象代表了当今物理学界最深奥的谜题之一。观测宇宙学的主要目标是通过精确测定遥远星系的距离与退行速度来绘制宇宙膨胀历史。传统方法依赖Ia型超新星作为“标准烛光”，由已知绝对光度推算天体距离。既往数据已明确证实了宇宙的加速膨胀，而鲁宾天文台计划积累约10万颗Ia型超新星数据，以空前庞大的统计样本大幅压缩观测不确定度。 除超新星巡天外，研究团队重点采用强引力透镜系统中的时间延迟测量法来约束暗能量。当包含闪烁活动星系核（AGN）的遥远背景星系被前景大质量星系弯曲的时空偏折时，会在不同路径上产生多个延迟成像。通过测定光变信号在各成像之间的微小传播时差，天文学家可直接计算透镜距离。科学界目前仅掌握约10个精测样本，LSST预计将这一样本规模扩展至数千个，从而将控制宇宙加速膨胀的暗能量参数约束精度提升数倍。 极限工程性能：智利峰顶与32亿像素相机 (6:02) 鲁宾天文台坐落于智利帕雄峰（Cerro Pachón）之巅，其巡天任务由SLAC耗时十年制造的LSST相机（LSST Cam）驱动。该设备尺寸堪比乘用车，焦平面分辨率高达32亿像素（3200兆像素），成像视场约为满月视面积的45倍，并配备涵盖紫外到近红外波段的6色滤光片转轮。以原生分辨率展示单幅全景图像需400台高清电视拼接，以72 DPI打印可覆盖两个网球场，其数字缩放精度如同从太空轨道俯瞰加利福尼亚州并逐级放大至识别地表独立建筑。 望远镜以规律的重访节奏系统性地拍摄南半球天空。单张广角曝光平均包含约200颗超新星、7000颗变星、70个活动星系核以及400个太阳系天体。在为期四天的早期调试巡天中，系统便发现了2000颗未知小行星，并在未预先设防的情况下捕获了星际彗星3I/ATLAS的运动轨迹，印证了其在短时标瞬变与运动天体探测上的敏锐性能。 百PB级数据流水线与云端普惠科研架构 (12:59) 在十年运行周期内，LSST将积累约200万次广角曝光，针对巡天范围内的每个天区进行约800次重访观测，记录约200亿次位置与亮度瞬变，最终编目200亿个星系、170亿颗恒星及600万个太阳系天体。这一过程将产生30 PB未压缩原始图像，经校准与图像叠加后衍生为数百PB处理数据。天文台构建了分级响应流水线：山顶观测站每晚回传20 TB裸数据，SLAC及英法计算中心在拍摄后60秒内发出自动化瞬变警报，80小时内交付即时数据产品，并逐年进行全量重新处理发布。 针对传统模式下海量数据下载的算力门槛，天文台构建了部署于谷歌云（Google Cloud）的“鲁宾科学平台”（RSP），可支持10,000名天文学家并发使用。研究人员直接通过网页浏览器和轻薄笔记本即可对数PB数据库执行复杂交互式查询，使小型学院学者享有与顶尖研究型大学同等的科研资源。此外，平台正计划接入美国能源部的Genesis集成研究基础设施，链接国家实验室的高性能计算中心以承载大规模机器学习推理需求。 算法突破：模拟推断加速与层次建模纠偏 (21:47) 尽管LSST提供了海量时域数据，但引力透镜系统的参数解算面临严峻的计算瓶颈。传统基于似然的马尔可夫链蒙特卡洛（MCMC）方法模拟单个透镜质量分布需耗费数天，无法处理数千个透镜的样本量。为此，研究团队引入基于模拟的推断（Simulation-Based Inference, SBI）与神经后验估计（Neural Posterior Estimation, NPE），利用约50万个逼真的模拟透镜预训练神经网络，将单系统后验参数分布估计耗时由数天大幅压缩至数分钟，使大规模群体分析在计算上变得可行。 在群体宇宙学推断中，算法面临的核心挑战是模拟与真实观测之间的“分布偏移”（Distribution Shift）。若模拟训练集中的星系质量分布未能完全覆盖自然界中的大质量星系群体，导出的宇宙膨胀速率便会产生系统性偏差。研究人员通过正规化流（Normalizing Flows）提升单体后验建模的表达能力，并引入层次群体推断（Hierarchical Inference）在观测中逆向推导修正先验。这种方法允许科研团队执行数百轮验证测试，兼顾秒级推断速度与无偏统计准确性。 公民科学协同与STEM教育普及 (18:37) 作为平均折合每位美国纳税人投入约5美元的公共科研设施，鲁宾天文台将公众参与深度嵌入日常科学运作。依托Zooniverse平台，天文台推出了“鲁宾差异侦探”（Rubin Difference Detectives）等项目，号召天文爱好者与公众志愿者审核差分图像切片，人工甄别候选瞬变是真实天体现象还是仪器杂音。这些众包标注构成了宝贵的真值数据集，能直接用于优化过滤每晚警报数据流的机器学习分类器。 此外，天文台专设教育与公众拓展（EPO）部门，为高中课堂开发涵盖天体物理与大数据分析的互动工具与教学大纲，在上线初期即记录了约80,000次使用。该项目利用公众对天文奇观的天然探索欲，打造了通往计算科学、大数据素养与跨学科工程技能培养的公共教育入口。 鲁宾天文台将高通量光学观测、云原生普惠数据平台与前沿机器学习推断融为一体，为解答暗能量与宇宙加速机制提供了跨越数量级的数据支撑。然而，由于群体参数测量高度敏感于先验假设，如何在未来十年的数据爆发期内确保仿真训练集与真实宇宙星系演化无缝对齐、彻底消除潜在的分布偏移系统误差，仍是该项目最终确证宇宙学模型的关键考验。

## Background 本内容由频道 Ai4 于 2026 年 8 月 17 日发布在 YouTube 平台上。视频简介显示，NSF-DOE 薇拉·鲁宾天文台（Vera C. Rubin Observatory）运行副主任 Phil Marshall 介绍了配备全球最大数码相机的西蒙尼巡天望远镜，并阐述了其未来十年开展时空遗迹巡天（LSST）所积累的数百 PB 级大数据，以及科研界如何借助 AI 与机器学习工具应对这一前所未有的数据挑战。 ## TL;DR 演讲首先从宇宙加速膨胀之谜切入，阐明了利用巡天观测与超新星标准烛光探索暗能量的科学构想。接着，Marshall 详细介绍了坐落于智利的鲁宾天文台及其核心仪器——由 SLAC 打造、拥有 32 亿像素的全球最大数码相机 LSST Cam，该仪器能够在十年内对南半球星空进行高频重访，产出数百 PB 的全天动态宇宙图景。为应对海量警报与观测数据，项目团队构建了云端 Rubin 科学平台，并积极联动 Zooniverse 开展公众科学。最后，演讲深入探讨了 SLAC 团队如何利用强引力透镜测量宇宙膨胀速率，展示了通过基于模拟的推断（SBI）与神经后验估计（NPE）将建模时间从数天缩短至数分钟，进而有效纠正分布偏移，并展望了天文学大数据对全球 STEM 人才培养的深远推动力。 ## Atmosphere & style 整场演讲充满科研探索的澎湃激情与幽默随和的大家风范，Marshall 以极具亲和力的口吻将艰深的宇宙学原理与前沿计算方法拆解得生动透彻。从抛出“凡是升起的必会落下”被宇宙打破的悬念，到逐级放大“谷歌地图级”深空星系时的惊叹，叙事节奏张弛有度、层层递进。观众不仅能沉浸在浩瀚星空的视觉震撼与尖端硬核工程的精巧之中，更能深切体会到“这项属于每一位纳税人的公共科学工程”所传递出的开放、包容与科研普惠情怀。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=_1H-yR25BO0&t=4s) 宇宙在近十亿年间的加速膨胀打破了经典物理直觉，是当前物理学界最大的未解之谜，鲁宾天文台的核心使命正是通过高精度观测宇宙膨胀历史来检验理论模型。 - [02:01](https://www.youtube.com/watch?v=_1H-yR25BO0&t=121s) 通过捕获 10 万颗 Ia 型超新星作为标准烛光，时空遗迹巡天（LSST）不仅能精准测定暗能量，还能同步解决太阳系未知小天体探测、银河系暗物质分布及大尺度纤维网结构绘制等关键问题。 - [05:08](https://www.youtube.com/watch?v=_1H-yR25BO0&t=308s) 坐落于智利帕穹山的薇拉·鲁宾天文台兼备广角（Wide）、敏捷（Fast）与极深（Deep）三大核心能力，将在 10 年内反复扫描覆盖整个南半球夜空。 - [07:32](https://www.youtube.com/watch?v=_1H-yR25BO0&t=452s) 单帧图像平均包含数百颗超新星和数千颗变星；早期试运行仅用 4 个夜晚就发现了 2,000 颗未知小行星，并成功捕获了人类探测到的第三颗星际天体 3I/ATLAS。 - [09:01](https://www.youtube.com/watch?v=_1H-yR25BO0&t=541s) SLAC 研制的 LSST Cam 拥有 32 亿像素，是人类历史上最大的数码相机，单次曝光视场达满月 45 倍，具备如同“Google Maps”逐级放大查看街景般的超高深空解析力。 - [12:59](https://www.youtube.com/watch?v=_1H-yR25BO0&t=779s) 10 年巡天相当于为南半球夜空拍摄一部宇宙电影，累计拍摄 200 万张图像、处理数百 PB 数据，完成对 200 亿个星系、170 亿颗恒星和 600 万个太阳系天体的全天体普查。 - [14:30](https://www.youtube.com/watch?v=_1H-yR25BO0&t=870s) 为应对天文学大数据在数据量、数据速率及多样性上的挑战，Rubin 科学平台构建在云端，为全球上万名天文学家提供平等的算力资源与 60 秒内发布的即时警报流。 - [18:37](https://www.youtube.com/watch?v=_1H-yR25BO0&t=1117s) 通过 Zooniverse 平台的“差异探测侦探”众包项目，普通公众的人工视觉核验标签被直接用于训练和优化机器学习算法，提升夜空变动警报的纯净度。 - [19:39](https://www.youtube.com/watch?v=_1H-yR25BO0&t=1179s) SLAC 团队利用强引力透镜中背景活动星系核（AGN）的光变时间延迟来精确测算宇宙膨胀速率，LSST 将使此类高精度样本数量从约 10 个扩增至数千个。 - [21:47](https://www.youtube.com/watch?v=_1H-yR25BO0&t=1307s) 引入基于模拟的推断（SBI）与神经后验估计（NPE），利用 50 万个模拟透镜训练神经网络，将单个透镜的参数建模耗时从数天缩短至数分钟。 - [23:40](https://www.youtube.com/watch?v=_1H-yR25BO0&t=1420s) 为防止大样本统计中的系统性偏差，研究团队采用标准化流与分层推断（Hierarchical Inference）校正“分布偏移”，确保暗能量参数测量的统计无偏与高准确度。 - [27:14](https://www.youtube.com/watch?v=_1H-yR25BO0&t=1634s) 鲁宾天文台作为公共资助的世界级装置，通过开放的 Skyviewer 与教育资源成为青少年的 STEM 引路石，推动新一代跨学科人才掌握大数据与人工智能技术。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=_1H-yR25BO0&t=4s) 宇宙加速膨胀之谜与观测宇宙学的使命 演讲者 Phil Marshall（国家科学基金会与美国能源部所属薇拉·鲁宾天文台运行副主任）在开场时打趣道，天文台主任给他定了一个“略带夸张色彩”的演讲标题——《有史以来最伟大的天文与天体物理发现机器》，但他会努力在接下来的半小时里向大家证明其实至名归。这是一座真正令人惊叹的顶尖科研设施。在进入技术细节前，Marshall 首先回顾了人类对宇宙演化历史的认知框架：宇宙始于极热的大爆炸，随后经历剧烈的暴胀阶段，随着时间推移逐渐形成恒星与星系，最终演化出如今我们所处、遍布螺旋星系的邻近宇宙。 然而，在这张宇宙演化图谱中，最令人费解与意外的现象发生在距今较近的宇宙演化阶段。图表中近期的曲线上出现了一个显著的“上扬拐点”。按照传统经典物理直觉，宇宙在膨胀过程中受自身引力牵引，星系之间相互拉扯，膨胀速度理应逐渐减慢。但实际观测表明事实并非如此——宇宙的膨胀正在加速，且这种加速状态已经持续了约十亿年之久。Marshall 指出，这是当前物理科学界最大的未解之谜。“我们从小接受的教育都是‘凡是升起的必会落下’（what goes up must come down）”，但在整个宇宙尺度上，这一定律却被彻底打破了。目前没有任何人确切知晓加速膨胀背后的原因。理论学家提出了各种假设模型，而像 Marshall 这样的观测宇宙学家，其核心使命就是通过测量越来越高精度的宇宙膨胀历史，去逐一检验并排除错误模型，最终揭示宇宙运转的深层机制。 ### [02:01](https://www.youtube.com/watch?v=_1H-yR25BO0&t=121s) 标准烛光测量与超级巡天的科学构想 天文学家确认宇宙正在加速膨胀的方法，是通过测量遥远星系的距离并对比它们的退行速度。在均匀膨胀的宇宙中，距离越远的天体退行速度越快。天文学家将 Ia 型超新星用作“标准烛光”（Standard Candles），因为这类超新星爆发时的绝对光度是已知的，通过测量它们在地球上观测到的视亮度便能推算出实际距离；再结合光谱测出的退行红移速度，即可精确绘制出宇宙的膨胀轨迹。过去几十年积累的观测数据（图表中的黄色观测点）极为明确地贴合了代表加速膨胀的绿色预测线，而不是代表减速并最终在大挤压中坍缩的蓝色预测线。 为了获得更高精度、更具决定性的测量结果，天文学界需要将数据量提升几个数量级——如果能够捕获 10 万颗超新星的数据，整个宇宙学研究将被彻底重塑。Marshall 描绘了一幅宏大的巡天设想：如果有一座望远镜能够长年累月凝视夜空、持续记录超新星爆发，不仅能获取数以亿计的精准数据点，还能顺带解决天文学中的一系列重大难题： - **太阳系内天体追踪**：在时序巡天图像中搜寻运动天体，测定太阳系内未知的小行星、微型行星，甚至搜寻海王星轨道外假想中的“第九行星”（Planet Nine）。 - **宇宙大尺度结构绘制**：在银河系之外，测定可观测天区内数以亿计星系的空间分布，拼接出宇宙纤维状网络结构，揭示大尺度结构是如何在引力作用下演化形成的。 - **银河系暗物质探测**：在银河系内部追踪恒星流（stellar streams），寻找暗物质团块穿过恒星流时所产生的细微扭曲与引力扰动特征。Marshall 在此展示了暗物质先驱薇拉·鲁宾（Vera Rubin）的历史照片与数据，正是鲁宾当年对仙女座星系及银河系恒星旋转速度的开创性测量，证实了星系内部存在大量看不见的暗物质。 ```mermaid flowchart TD accTitle: LSST 巡天望远镜能力与多领域科学产出 subgraph TelescopeCapabilities["鲁宾天文台核心观测能力"] A["Wide（广角大视场）<br/>3.2 亿像素相机 / 45倍满月面积"] B["Fast（快速全天扫描）<br/>敏捷指向 / 数天内覆盖全天"] C["Deep（深度极高集光）<br/>大口径反射镜 / 单帧与叠加极深"] end subgraph SurveyProcess["时空遗迹巡天 (LSST)"] D["10 年持续重复巡天监测"] end subgraph ScienceOutputs["多领域科学目标与产出"] E["宇宙学：10万颗 Ia 型超新星 / 暗能量与加速膨胀"] F["太阳系科学：小行星发现 / 星际天体 / 第九行星搜寻"] G["银河系与暗物质：恒星流微扰 / 银河系暗物质分布"] H["系外星系：超大质量黑洞 (AGN) / 宇宙大尺度结构演化"] end A & B & C --> D D --> E D --> F D --> G D --> H ``` ### [05:08](https://www.youtube.com/watch?v=_1H-yR25BO0&t=308s) 帕穹山上的鲁宾天文台与 LSST 巡天机制 要实现上述宏伟蓝图，对巡天硬件系统提出了三个近乎苛刻的要求：望远镜必须转向敏捷（Fast），拍完一处夜空后能迅速指向下一处；相机视野必须极其宽广（Wide），单次成像面积足够大以减少重复覆盖的开销；望远镜必须具备巨大的集光口径（Deep），确保每一幅单帧图像都能探探测到极其暗弱的天体。 最终建成的正是坐落于智利帕穹山（Cerro Pachón）山顶的薇拉·鲁宾天文台（NSF-DOE Vera C. Rubin Observatory）。Marshall 特别停顿强调，这座最新的大型天文台全部由公共资金资助建造，“它属于你们每一个人”。天文台的核心科学使命是执行为期 10 年的“时空遗迹巡天”（Legacy Survey of Space and Time, LSST）。望远镜就像铺设瓷砖一样用相机视野无缝覆盖整个南半球夜空，夜复一夜反复拍摄相同天区。这种重访机制带来了两重核心优势：一方面，比对前后不同时间点的图像可以捕获超新星爆发、瞬变源以及移动天体；另一方面，将同一天区的历史图像连续叠加后，等效于进行极长时间的深度曝光，从而揭示出以往因过于暗弱而无法被看见的古老星系与恒星。 ### [07:32](https://www.youtube.com/watch?v=_1H-yR25BO0&t=452s) 单帧图像的信息容量与试运行成果 鲁宾天文台单次拍摄的大画幅图像所蕴含的信息量惊人。平均而言，每一张单画幅图像中都包含： - 约 200 颗超新星； - 约 7,000 颗变星； - 约 70 个活动星系核（AGN，即遥远星系中心的超大质量黑洞）； - 约 400 个太阳系内天体（包括各类小行星与小天体）； - 以及大量人类目前尚未认知、有待巡天发现的未知动态天文现象。 巡天观测于 6 月份启动后，重复成像的策略迅速转化为了前沿科研成果。在早期的试运行（commissioning）阶段，团队仅开展了为期约 4 个夜晚的小规模测试巡天，便一口气发现了 2,000 颗此前未知的新小行星。此外，试运行图像中还捕获了人类历史上探测到的第三颗星际天体 3I/ATLAS（Atlas 3II）穿过太阳系的运动影像。Marshall 表示，团队并没有专门去定向搜索它，它只是自然而然地出现在了常规巡天图像中。据估计，在未来的 10 年正式巡天周期内，鲁宾天文台将发现大约 10 颗此类来自太阳系外的星际访客。 ### [09:01](https://www.youtube.com/watch?v=_1H-yR25BO0&t=541s) 史上最大数码相机 LSST Cam 与“谷歌地图级”深空解析力 实现这一切的核心仪器是由 SLAC 国家加速器实验室历时约 10 年打造的 LSST Cam——这也是人类历史上制造过的最大数码相机。 - **物理规格与传感器**：相机体积堪比一辆小汽车，其焦平面由高达 32 亿像素（3,200 megapixels）的感光阵列构成。 - **分辨率直观对比**：如果要以原生全分辨率完整显示单张原始图像，需要 400 台高清电视拼成显示屏墙；若以 72 dpi 的分辨率打印成纸质图像，其面积足以覆盖两座标准网球场。 - **视场与滤光系统**：单次曝光的视场面积约为满月视面积的 45 倍，相当于在手臂伸直时两根手指在夜空中遮挡的范围。相机传感器本身虽为黑白成像，但配备了 6 个不同波段的滤光片，覆盖从近紫外到近红外的光学波段，用以合成高质量的多波段彩色天图。 Marshall 展示了一张由相邻两块 LSST Cam 视野拼接而成的初光测试图像。仅仅这一小片天区就包含了 1,000 万颗恒星与星系。随着演讲中镜头不断逐级数字放大，画面中原本看似细小的噪点和微尘全部展现为真实的星系形态——有近邻星系、相互碰撞并合的相互作用星系、轮廓分明的螺旋星系，以及由黄色红星系组成的致密星系团。画面中的每一个光点都是一个由数百亿颗恒星构成的星系，其中绝大多数横跨了大半个宇宙的漫长时空。Marshall 总结道，拥有世界最大的数码相机带来了无与伦比的“数字变焦”（digital zoom）能力，其探索夜空的深度与细节层级，已经达到了天文学家在夜空中使用 Google Maps 逐级放大查看地球街景般的超高解析度。 ### [12:00](https://www.youtube.com/watch?v=_1H-yR25BO0&t=720s) 缩放宇宙视角与公众探索：Rubin Skyviewer 主讲人 Phil Marshall 形象地比喻了望远镜的分辨率与视场：仰望星空时所覆盖的天区，其广度相当于从太空俯瞰地球表面上的整个加州大小；而当镜头不断放大、精细到每一颗单独的恒星与星系时，其细节程度就如同从空中聚焦到地球表面的一栋栋独立建筑物。为了让公众能够亲身体验这种极致的探索乐趣，项目团队开发了名为 Skyviewer 的网页应用（`rubinskyviewer.app`）。Marshall 幽默地告诉在场观众，屏幕下方附有二维码，“这是属于你们的天文台，这是你们的数据集”，即使大家在接下来的演讲时间里只顾着在手机上滑动浏览首光（First Look）图像，他也完全不介意。 --- ### [12:59](https://www.youtube.com/watch?v=_1H-yR25BO0&t=779s) 十年巡天的数据洪流：宇宙级的“动态电影” Rubin 天文台的核心科学任务是为期 10 年的时域空间遗产巡天（Legacy Survey of Space and Time, LSST）。在这 10 年间，望远镜将在 6 个光学与近红外滤光片波段下累计拍摄 200 万张图像，捕捉夜空中约 200 亿次亮度与位置的变化。在数据规模上，30 PB 的原始图像经过校准与处理后将膨胀至数百 PB，且团队承诺绝不丢弃任何历史数据，而是随着巡天推进不断重新处理，叠加出更深、更透彻的宇宙图像。 ```mermaid flowchart TD accTitle: LSST 数据产出与科学处理流程 subgraph Capture["观测与采集"] A["10年巡天 / 6个滤光片"] --> B["200万张图像 (30 PB 原始数据)"] end subgraph DataProducts["数据产品生成"] B --> C["即时警报流 (60秒内发出 / 200亿次天体变动)"] B --> D["持续校准与堆叠处理 (扩展至数百 PB 深度图像)"] end subgraph Inventory["全天体普查清单"] D --> E["200亿 遥远星系"] D --> F["170亿 银河系恒星"] D --> G["600万 太阳系小天体"] end subgraph Science["科学分析与推断"] C --> H["全球数千名天文学家"] E --> H F --> H G --> H H --> I["对比物理模型 / 探索深空奥秘"] end ``` 这项巡天相当于为智利可见的南半球宇宙拍摄一部“宇宙电影”，对每个天区位置重复曝光约 800 次。在天文学家眼中，这意味着对数十亿个天体进行数万亿次自动测量，每个天体都拥有上千个观测历元（epochs）。来自世界各地的数千名研究人员将利用这些数据与物理理论模型进行对比，探索深空宇宙的演化。 - **巡天时长**：10 年连续观测 - **原始图像归档**：30 PB 原始像素数据，衍生处理后达数百 PB - **天体目录普查**：200 亿个星系、170 亿颗恒星、600 万个太阳系天体 - **观测频次**：单天区平均重访拍摄约 800 次，产生上千个时间历元 --- ### [14:30](https://www.youtube.com/watch?v=_1H-yR25BO0&t=870s) 天文学的 Big Data 挑战与 Rubin 科学平台 LSST 相机（LSST Cam）给全球天文学界带来了前所未有的“大数据难题”（Big Data Problem），并在三大维度上带来冲击： 1. **数据量（Volume）**：图像归档与自动测量目录规模远超以往任何天文项目； 2. **数据速率（Velocity）**：每晚实时向全球分发的夜空变动警报流（Alert Stream）以前所未有的速度扑面而来； 3. **数据多样性（Variety）**：海量测量数据促使学者采用参数更灵活、复杂度更高的物理模型。 为了应对这一挑战，天文台构建了 Rubin 科学平台（Rubin Science Platform），旨在为全球 10,000 名天文学家提供平等的云端科研环境。Marshall 将其生动地描述为“一台轻薄小巧的笔记本电脑直接连接在 Google Cloud 上的巨型数据库”。无论学者来自顶尖名校还是资源有限的小型教学学院，都可以凭借浏览器弹性调用数据与算力。 ```mermaid flowchart LR accTitle: 科学数据流向与计算架构 subgraph Facilities["数据处理与算力中心"] SLAC["SLAC 国家加速器实验室"] Intl["英国与法国数据中心"] Genesis["美国能源部 Genesis 计划<br/>(连接各大 DOE 实验室算力)"] end subgraph DataStream["数据分发时效"] PromptAlert["即时警报<br/>(曝光后 60 秒内)"] PromptProd["快报数据产品<br/>(80 小时内发布)"] AnnualRelease["年度完整重处理数据发布<br/>(每年一次全量重算)"] end subgraph Platform["用户工作区"] GCP["Google Cloud / Rubin 科学平台"] Users["全球 10,000 名天文学家<br/>(弹性 Jupyter/SQL 分析)"] end SLAC & Intl --> PromptAlert & PromptProd & AnnualRelease PromptAlert & PromptProd & AnnualRelease --> GCP Genesis -.->|补充前沿大模型与 ML 算力| GCP GCP --> Users ``` 虽然每晚产生约 20 TB 的原始数据，但科学家更关注的是处理后的衍生产品：60 秒内发布的即时警报、80 小时内发布的即时数据，以及每年一次全量重处理的年度数据发布。核心数据处理在 SLAC 以及英国、法国的数据中心完成。考虑到天文学家随着机器学习应用的深入对算力需求激增，Rubin 正在与美国能源部（DOE）的 **Genesis 任务**对接，将各大能源部国家实验室的超级计算集群打通互联，LSST 数据集作为核心数据输入，换取庞大的国家级算力资源支持后续分析。 --- ### [18:37](https://www.youtube.com/watch?v=_1H-yR25BO0&t=1117s) 全民科学参与：Zooniverse 上的“差异探测侦探” 除了专业天文学家，普通公众也能深度参与科学发现。Rubin 团队借助著名的公民科学平台 **Zooniverse** 发起了一系列众包项目。 - **差异图像辨别（Rubin Difference Detectives）**：邀请志愿者在线对差分图像进行人工视觉核验与分类，标注出图像中的变动天体是真实的天文事件（Real）还是图像噪点伪影（Bogus）。 - **优化机器学习分类器**：公众提供的这些高质量人工标注标签，将直接用作训练数据，训练并提升机器学习算法的分类精度，使望远镜每晚发出的公共警报流更加纯净准确。 --- ### [19:39](https://www.youtube.com/watch?v=_1H-yR25BO0&t=1179s) SLAC 核心科研：强引力透镜与宇宙加速膨胀 转入 SLAC 团队自身的前沿科研课题，科学家们重点关注宇宙膨胀历史以及宇宙加速膨胀背后的谜团。要解开这个谜团，核心在于高精度测量遥远星系的宇宙学距离，而**强引力透镜（Gravitational Lenses）**是关键突破口。 当一个背景星系中心包含活跃星系核（AGN / 超大质量黑洞吞噬气体产生剧烈亮度闪烁）且恰好位于一个前景大质量星系正后方时，背景光源发出的光线不会被挡住，而会被前景星系的引力场弯曲，形成多重成像。由于光线沿不同路径传播的光程与引力势存在差异，背景黑洞的闪烁信号在不同透镜图像之间会出现**时间延迟（Time Delay）**。通过测量这一时间延迟，即可精确测算出透镜星系的几何距离，从而约束宇宙膨胀速率（哈勃常数）。目前全球已高精度测量的透镜样本仅有约 10 个，而 LSST 将把这一样本量扩大至数千个，极大提高测量精度。 --- ### [21:47](https://www.youtube.com/watch?v=_1H-yR25BO0&t=1307s) 破解透镜建模瓶颈：基于模拟的神经网络后验估计 在数千个引力透镜中提取时间延迟并用于宇宙学研究极具挑战性。虽然 LSST 提供了高质量的时间序列数据，但最具计算瓶颈的环节在于**对透镜偏转天体的质量分布进行精确建模**。 传统基于似然（Likelihood-based）的马尔可夫链蒙特卡洛（MCMC）参数建模方法，对单个透镜后验概率分布函数（Posterior PDF）的采样往往需要耗费**数天**计算时间。面对数千个透镜样本，这种算力消耗无法承受。为此，团队引入了**基于模拟的推断（Simulation-Based Inference, SBI）**与**神经后验估计（Neural Posterior Estimation, NPE）**： - **模型先验与训练**：利用包含约 50 万个模拟引力透镜的数据集预先训练深度神经网络； - **极速推断**：当新观测数据输入时，训练完成的神经网络可在**数分钟内**直接推断出透镜参数与后验分布； - **效率提升**：经过过去数年的测试验证，建模耗时从“数天”直接缩短至“数分钟”，成功扫清了利用数千引力透镜进行宇宙学分析的技术阻碍，目前团队正致力于确保模型在大样本统计意义下输出结果的无偏性。 ### [23:40](https://www.youtube.com/watch?v=_1H-yR25BO0&t=1420s) 宇宙膨胀率测量中的精度与分布偏移挑战 在利用强引力透镜测量宇宙膨胀速率时，天文学家追求的不仅是极高的统计“精度”（Precision），更是结果的“准确度”（Accuracy）——必须保证得出完全正确的物理答案。为了验证这一点，研究团队投入了大量精力对训练数据进行严格的验证测试，确保这些数据在应用于真实巡天观测时能够给出可靠结论。研究表明，在纳入这一更大规模的强引力透镜样本后，描述宇宙加速膨胀的关键“暗能量参数”（Dark Energy Parameters）的测量精度与准确度能够提升数倍。 在展示的对比图表中，橙色与紫色的对比直观地揭示了准确度的重要性：紫色的结果精准对齐在黑色的十字准星中心，代表了准确的测量结果；而橙色则明显偏离，代表存在偏差的错误结果。这种误差并非源自单个透镜天体，而是来自透镜天体的整体“总体分布”（Population）。巡天观测中包含了从未知真实总体中抽取的数千个强透镜样本，倘若研究人员未能准确刻画该天体总体的特征、不理解大质量星系的平均物理演化规律，就会在宇宙膨胀速率的推断上得出错误的结论。因此，解决“分布偏移”（Distribution Shift）是当前的核心攻坚方向。研究团队目前正在利用标准化流（Normalizing Flows）来构建单个天体后验概率密度函数（PDF），并不断提高模型表达能力，以捕捉更复杂的总体分布特征，进而从分布偏移中恢复出真实的物理参数。 ```mermaid flowchart TD accTitle: 强引力透镜数据推断与分布偏移校正流程 subgraph SIM["高保真模拟与训练"] A["高复杂度天文物理模拟"] --> B["生成训练数据集"] B --> C["引入真实先验分布 Prior"] end subgraph NPE["神经后验估计与建模"] C --> D["标准化流 Normalizing Flows<br/>单天体后验 PDF 估计"] D --> E["分层概率推断<br/>Hierarchical Inference"] end subgraph OUT["科学输出"] E --> F["校正分布偏移 Distribution Shift"] F --> G["精确暗能量参数与宇宙膨胀率"] end ``` ### [25:19](https://www.youtube.com/watch?v=_1H-yR25BO0&t=1519s) 神经后验估计（NPE）与分层推断的力量 当前用于强引力透镜推断的概率图模型（Probabilistic Graphical Model），在底层科学逻辑上与科学家们在十至二十年前构想的完全一致；真正发生革命性变化的，是大规模数据处理与计算能力，而这正是神经后验估计（Neural Posterior Estimation, NPE）带来的核心突破。计算速度的大幅跃升带来了深远的科学效益：它不仅使分析海量样本成为可能，还允许科学家反复执行快速推断，从而全面排查各种模型假设背后的系统误差，确保测量的高准确度。 基于模拟的推断（Simulation-Based Inference, SBI）还带来了一项关键优势：在构建训练集的过程中，研究人员能够赋予模型一个真正可信的先验概率密度。通过在模拟中注入丰富的物理细节，使模拟宇宙越来越逼真、与真实宇宙的观测数据高度契合，从而有效降低分布偏移风险。更重要的是，“分层推断”（Hierarchical Inference）构成了应对分布偏移的通用解决方案——即如果在初始训练集中设定的先验分布存在偏差，科学家可以通过分层架构反向推断出真正期望设定的理想先验，从而在算法层面实现自适应纠偏。 - 科学逻辑的继承性：核心概率图模型二十年来保持不变，NPE 赋予了其处理 LSST 级别超大海量数据的规模化能力。 - 快速推断验证系统误差：毫秒级/秒级的推断速度支持反复测试不同物理假设下的系统性偏差。 - 分层推断纠偏机制：允许在总体层级动态推断修正先验分布，彻底解决训练数据与真实宇宙之间的分布不匹配问题。 ### [27:14](https://www.youtube.com/watch?v=_1H-yR25BO0&t=1634s) 天文学作为 STEM 的引路石与全民参与的鲁宾天文台 SLAC 国家加速器实验室以及全球诸多研究团队正在广泛采用神经后验估计技术处理天文大数据，这也预示着分析 LSST 数据集将推动物理学、天文学乃至数学和电子工程等交叉学科的新一代青年学者全面转向现代统计推断与机器学习。天文学常被称为通往 STEM（科学、技术、工程、数学）领域的“引路石”（Gateway Drug）——它在青少年心中播下对浩瀚星空的好奇种子，进而引导他们掌握处理海量大数据与人工智能的核心技能。 鲁宾天文台为此专门设立了“教育与公众普及部门”（Education and Public Outreach），为中学教师和学生提供丰富的教学资源，不仅讲授天体物理知识，还引导学生认识大数据与机器学习的实际应用。尽管空间与时间遗留巡天（LSST）才刚刚于 6 月正式启动，这些教育资源和互动教程已被使用超过 80,000 次。在接下来的十年里，数以万计的中学生将通过凝视夜空、探索 LSST 数据迈入大数据与人工智能的世界。讲者风趣地指出，听众未来在企业中招聘或合作的技术人才，很可能正是看着鲁宾天文台的数据成长起来的。此外，这座由纳税人资助的世界级科学装置属于每一个人（平均每位纳税人贡献了约 5 美元），讲者热情邀请公众访问 rubinobservatory.com、使用 Skyviewer 应用探索星空数据，或在 Zooniverse 上参与公民科学项目，共同开启为期十年的科学发现之旅。 - 教育普及影响力：配套的教育教程上线不久便已获得约 80,000 次独立使用，成为向高中普及大数据与 AI 知识的前沿阵地。 - 公众参与途径：大众可通过 `skyviewer.app` 浏览巡天图像，或通过 Zooniverse 平台直接协助科学家进行天体分类与数据标注。 - 全民共享的科学工程：该项目由公共资金支持，平均每位纳税人出资约 5 美元，真正实现了全民参与和数据共享。 ## Video description Phil Marshall - Deputy Director of Operations, NSF-DOE Vera C. Rubin Observatory Recently equipped with the world’s biggest digital camera, built at SLAC National Accelerator Laboratory, the Simonyi Survey Telescope at the Vera C. Rubin Observatory is built for high speed astronomical surveying. Over the next 10 years, scientists, engineers and developers at NSF NOIRLab and SLAC will be carrying out the Legacy Survey of Space and Time, repeatedly imaging the entire Southern sky in 6 different optical and near infrared bands, and accumulating hundreds of petabytes of pixel, catalog, and time series data for scientists all over the world to use. I will introduce the Rubin system and the data it produces, and show how our community is solving their LSST Big Data problems with AI and machine learning tools. [Source](https://www.youtube.com/watch?v=_1H-yR25BO0)

[0:04] Phil Marshall: the biggest mystery in physical science at the moment. [0:04] Phil Marshall: what goes up must come down [0:04] Phil Marshall: standard candles [9:01] Phil Marshall: first look [15:18] Phil Marshall: bringing the scientist to the data [18:37] Phil Marshall: Rubin Difference Detectives [18:37] Phil Marshall: bogus [21:47] Phil Marshall: getting the answer right on average [27:14] Phil Marshall: gateway drug for the STEM field

[0:04] Uh my director gave me this uh slightly hyperbolic  title. I'll try and live up to it for the [0:09] uh the next half hour or so. Um but this is a  truly amazing uh facility that we have built. I [0:15] want to tell you uh a little bit about it, what  we're doing and why and then in the second half [0:21] try and give you a few specific reasons for  why uh you might be interested in in the NSF [0:26] DOE Vera C Rubin Observatory and the LSST data  set that we're going to be producing. So first [0:34] some context just to remind you where you are.  Uh this is a picture of the universe and its [0:39] history. As you all probably know, uh the universe  began with a hot big bang. That's on the left. [0:45] um and then uh expanded in a rapid inflation  phase. Stars and galaxies started to form as time [0:52] goes towards the right and then you can see our  local universe full of familiar spiral galaxies [0:57] and things on the right. The the unusual thing,  the surprising thing about this picture is what's [1:02] going on recently. You see this uptick here. You  might have thought that an expanding universe [1:08] would slow down under its own gravity as galaxies  pull the universe back together again. Turns out [1:13] that's not what's happening. The expansion is  actually accelerating. It has been for the last [1:18] uh billion years or so. So this is uh I would say  uh the biggest mystery in physical science at the [1:26] moment. Uh we're taught from a young age that  what goes up must come down. But when it comes [1:30] to the whole universe, that's not the case. So  our program in observational cosmology is to try [1:38] and figure out what is going on. Universe is  expansion is accelerating. Nobody knows why. [1:44] Theorists will come up with ideas for what might  be happening. The observational cosmologists like [1:48] me, we're trying to make more and more accurate  measurements of the expansion history to try and [1:54] rule out models and try and get a handle on on  why the universe is doing this. Okay, so how do [2:01] we know that it's uh the expansion of the universe  is accelerating? Turns out you do it by measuring [2:06] distance to faroff galaxies and comparing their  distances with how fast they seem to be moving [2:12] away from us. In a uniform expansion, the things  that are further away appear to be moving moving [2:18] away far faster. So we use uh type 1A supernovi  as standard candles. We know how luminous they [2:25] are. We measure how bright they appear to be.  That tells us something about their distance. [2:30] We can measure their recession speed. And we  can trace the expansion history of the universe [2:34] in this way. So here's what all the data we've  been accumulating on this over the last couple [2:40] of decades. Uh the yellow points, those are the  observed supernova. You can see that they clearly [2:44] trace the green accelerating universe line, not  the blue non-acelerating collapsing back on itself [2:52] lines. This is how we know what the universe is  doing. Okay. So, if we want to make more accurate, [2:58] more precise, more accurate measurements of the  expansion of the universe, what do we do? Well, [3:03] we can add more points to this to this graph, we  can ask ourselves, what if we could see a 100,000 [3:10] supernovi? What if we sat uh watching the uh the  night sky and over time watch supernovi go off? [3:20] We could build up a much larger sample. If we did  that and put gazillions of points on that graph, [3:26] we could do other things as well. We could look  at uh in those images, we could look for moving [3:32] objects as well as as well as standard candles. We  could trace out asteroids and uh uh objects in our [3:39] solar system. We could look for planet 9 might be  out there in uh beyond the uh beyond Neptune. If [3:47] we did do a giant sky survey looking for 100,000  supernova over the whole sky as well as tracing [3:53] out moving objects, we could look beyond our solar  system outside our own galaxy. We could trace [3:58] out where all the galaxies in the universe are at  least in the observable patch of sky and we could [4:03] start to uh piece together the the network that  they that they form in and try and understand how [4:09] structure in the universe formed uh in our galaxy.  We could look for streams of stars and little [4:18] uh kinks and wiggles in those streams of stars  that you might get when a blob of dark matter [4:24] flies through them. Here's a picture of Vera  Rubin who discovered dark matter. Here she is [4:30] with some of her data showing that the Milky  Way or the Andromeda galaxy is more massive [4:35] than it looks. It must have some dark matter in  it to keep all those high-speed stars in it. We [4:41] could do some experiments like that in our galaxy  to try and understand what the dark matter is. [4:48] Okay, so we could do all of that if we just had  a big enough sky survey and we watched the sky [4:54] for long enough to see all the supernovi and  see all the moving objects and trace out where [4:59] all the galaxies are. We could learn a huge  amount about our universe by doing such a sky [5:03] survey imagined here. But what would it take  to do that? Well, if we really were to watch [5:10] the whole of the the observable sky uh for long  enough to catch a 100,000 supernovi and to see [5:16] all of these planets and galaxies, what would  we need? We would need a a telescope that can [5:22] move around the sky quickly to take an image  here and then move on to take an image here [5:26] and then an image here. We would need a camera  big enough that every time we took an image, [5:32] it was a big enough image that we didn't  have to go back so often. and all and we [5:38] want that telescope to be big enough to have a  big enough collecting area, a big enough mirror [5:43] to collect light that we could go deep in every  in every image. So, we need a telescope that can [5:48] go wide and fast and deep. Okay. So, you can  probably tell where this is going. I'd like [5:56] to show you what we've built so far and then  talk you talk you through it. So, let's see. [6:02] high amounts of if you go to the Reuben website  you can see that video for now though I'll wave my [6:11] hands some more on the top of a mountain in Chile  called Sarapan is this the Verus Rubin Observatory [6:18] uh uh our newest big astronomical facility um I  wanted to pause on this slide just to to point [6:25] out that um this was built with public money and  it does belong to you. this is your new biggest [6:30] uh astronomical facility. That's one reason  why you might be interested in it. Um but it [6:36] has been designed to do the kind of survey that  I was just talking about the the legacy survey [6:41] of space and time, the LSST. And so perhaps I can  show you a bit about uh about what it is doing. [6:53] So here we are. You can see it tiling  the sky with images from the camera. [6:59] uh building up a a a giant image of the  sky built out of many individual images [7:05] and then we go back night after night and  reimage it and reimage it and reimage it [7:11] so that we can look for changes between  the images. That's what we need to see the [7:14] supernovi go off. It's what we need to see  the the planets moving across the screen. [7:22] And when you add those images together, it's  as if you've taken a very deep long exposure [7:27] image. you can see uh galaxies, stars that you  couldn't see before, too faint to see before [7:32] and such a system was made. Okay. So, what kinds  of things can we see in one uh image taken with [7:40] Reuben? Each one of our uh uh large format images  uh contains about 200 supernovi. So, we're on the [7:47] way to that 100,000 that we needed. But it also  contains something like 7,000 variable stars, [7:53] uh 70 active galactic nuclear, that's super  massive black holes in other galaxies, [7:58] um 400 objects in the solar system,  so asteroids, uh minor planets, and [8:02] so on. And then a whole lot of other activity out  there, uh some of which we don't know about yet, [8:07] but we'll be discovering with this uh with  this survey. Uh we're already imaging the [8:14] sky repeatedly. We started the survey in June  and um that repeated images is already yielding [8:20] new science. On the left hand side uh you can  see some example um asteroids detected in the [8:27] um uh commissioning data. Uh we did a small  survey for for uh four nights or so and [8:33] discovered uh 2,000 new asteroids. And on the  right hand side you can see uh a little movie [8:40] uh made of uh an object called Atlas 3II which  is the third interstellar object ever detected. [8:47] Um we didn't go looking for it. It was just in  all the images that we were taking. There it [8:52] is. We expect to see about 10 more of these fly  through the solar system over the next 10 years. [9:01] Okay. So let me tell you a bit about the  the instrument that we're using to to do [9:04] this survey. Uh it's called LSST Cam. This is the  biggest digital camera in the world. It was built [9:09] at Slack, the national lab that I work at. Um it  took uh something like 10 years to build. Uh what [9:18] do I mean? It's the biggest camera in the world.  It's the size of a car and its focal plane has [9:22] 3,200 megapixels. Uh so what does that mean? It  means you need uh 400 HD TVs to display one image [9:31] at native resolution. uh you could print it out at  72 dots per inch and cover two tennis courts. Uh [9:38] so its field of view is something like uh 45 times  the size of the full moon if you can imagine such [9:43] a thing on the sky. Uh we have uh six filters  uh it's a black and white camera but we have [9:49] six filters to allow us to build up color imaging  uh from the ultraviolet through to the infrared. [9:57] Here's a picture of uh that filter set.  That's the um the camera just before it was [10:02] uh shipped to Chile in the tiny room slack.  Okay. So, to give you a sense of of how big [10:09] um that field of view is, uh here's two fingers on  the sky uh covering up a uh patch of test imaging. [10:16] You can see there the size of the LSST cam field  of view. Uh there's a patch of sky imaged with [10:23] roughly speaking two two LSST cam fields next  to each other. This was our first look image [10:29] set. And keep an eye on that green rectangle uh  because now we can uh uh zoom in and start to see [10:37] uh more of the detail that that is in the night  sky image to this depth. This is a preview of [10:42] what the 10-year survey imaging will look like.  There are 10 million stars and galaxies in this [10:48] image. Most of what you see here are galaxies.  You can see there's some nearby galaxies, [10:55] some interacting galaxies, some familiar familiar  uh spiral shapes. You can see clusters of [11:01] uh red galaxies that look yellow. And if you  keep zooming in, you can see that every um uh [11:12] every dot in that image really is uh a galaxy.  Each one containing billions of stars, most of [11:19] them halfway across the universe. And so there's  a huge amount of information in these images for [11:24] for learning about our universe uh where the  uh structures came from how they formed and [11:30] uh tracing back to to answer the question of how  fast is the the universe expansion accelerating. [11:38] So this is what you get for the big with the  biggest digital camera in the world. You get [11:42] a lot of digital zoom. You can keep going and see  more and more detail. Um, and it's now giving us [11:49] about the same digital zoom power that Google  Maps does. If you think about the night sky, [11:54] comparing it to the surface of the Earth, that  patch of um, imaging, the two LSST cam fields, [12:00] uh, looking up on the sky like that makes it about  the same size as California when looking down on [12:05] the surface of the Earth. And each individual star  and galaxy when you zoom in that far, it's like [12:11] zooming in to look at individual buildings on the  surface of the Earth. Okay, so hopefully uh this [12:19] is of interest to you. Uh we're trying to make  it easy for uh for everybody to enjoy the Reuben [12:24] imaging. So here's a website skyviewer um app.  It's underneath the subtitles here, but there's [12:31] a QR code. The sky viewer at Reubin Skyviewer.app  is designed to allow you to do this digital zoom [12:38] yourself and explore the the data that we make.  Why are we doing that? Well, because it's your [12:44] observatory. This is your data set. We want you to  enjoy it. So if you spend the rest of this talk uh [12:49] just scrolling around the first look imaging that  will be totally fine. I don't mind if you do that. [12:59] Okay. So some numbers for you. We're going to  survey the sky for 10 years. We're going to [13:04] uh end up taking uh 2 million images over six  filters. um we'll end up detecting something like [13:11] uh 20 billion changes in the night sky, changes in  brightness, changes in position of of everything [13:16] up there. Uh we'll build up an image archive  of 30 pabytes of of raw imaging. When we [13:23] um process that and calibrate it, it will  expand up to a few hundred pabytes. And [13:28] we won't throw any of it away. In fact, we'll  reprocess it continually through the survey to [13:32] make a deeper and deeper stacked image. and we'll  end up taking an inventory of the universe or at [13:39] least the southern half of the universe visible  from Chile. Uh so that's 20 billion galaxies, [13:44] 17 billion stars, uh just a wealth of data,  6 million solar system objects. And remember, [13:51] we're we're repeatedly imaging these night after  night, something like 800 times per pointing uh [13:58] and building up. You can think of it as a cosmic  movie. In practice, the scientists are thinking [14:03] about it in terms of trillions of measurements  of billions of objects. Each object with [14:09] um uh a thousand epochs and then many measurements  taken made automatically from those images. And [14:16] so across the world, thousands of astronomers  will be using these measurements and comparing [14:22] them to physical models in order to learn about  what's going on uh out there in in deep space. [14:30] So I like to uh uh prod my colleagues with the  thought that LSST cam gives all astronomers a [14:39] big data problem. This image archive, this  big catalog of automatic measurements, [14:44] is much bigger than anything they've worked  with before. The alert stream of everything that [14:49] changes in the night sky, which we put out to the  to the public every night, is data coming at them [14:56] faster than they've had to deal with before. And  all of these measurements, all of this imaging is [15:01] allowing them to probe the universe in more detail  with more flexible, more complex models than they [15:06] have before, which is to say they have a variety  problem as well. So all three aspects of a big [15:12] data problem is what is what you get for building  the uh the world's biggest digital camera. So it's [15:18] forcing all of us in in astrophysics to uh  to think hard about how we analyze our data, [15:23] how to to rethink it. And there are some aspects  of solving the LSST big data problem that um I [15:29] think will be familiar to many of you and just  wanted to tell you a little bit about how we're [15:33] how we're tackling it. So we're going to need to  enable data access to thousands of scientists to [15:39] work with this data set. We're going to need to  give them enough computing resources to do their [15:44] analysis and we're going to or they are going to  need to develop novel techniques for doing the [15:50] scientific inference that they want to do and  now get to do but on a much larger scale than [15:54] they have done before. So here's one part of  the solution. And this is something we've been [16:00] working on at the observatory to provide to these  thousands of science universe uh um scientists um [16:07] and that is to provide a science platform for  10,000 uh astronomers uh to work with the LSST [16:15] data from wherever they are. So imagine a small  laptop sitting on a giant database in the Google [16:21] cloud. That's the Rubin science platform and it's  a leveler. people at uh smaller universities uh [16:28] teaching colleges uh and beyond can now come to  the Reuben Science platform and work with this [16:33] data set uh perhaps in a way that they wouldn't  have been able to before. Here's some more detail [16:39] on that that Reuben science platform bringing  10,000 scientists uh to the data. Um uh you [16:48] can see how the raw data rate on the on the right  20 terabytes a night is actually not so relevant. [16:53] more relevant is the the data that they will  actually be working with. So the processed images [16:58] uh the prompt data products the the alerts coming  out within uh 60 seconds of the image being taken [17:04] uh prompt products released within 80 hours or so  and then roughly every year a big data release of [17:10] all the imaging we've taken so far reprocessed and  uh automatic measurements made. So the processing [17:18] happens at Slack. Uh also we have data facilities  in the UK and France. Um the prompt processing is [17:24] uh uh all done on a nightly basis and the users  are in the Google cloud on the Ruben science [17:30] platform um working in a in an elastic way.  We are thinking ahead through the survey and [17:38] thinking about how we will cope with increasing  amounts of user demand from the science community [17:43] because of course once they really get to grips  with their big data problem and start employing [17:47] machine learning problems to work with the data  at scale they'll get hungrier and hungrier for [17:52] computing resources. How are we going to keep keep  feeding them? One answer to that is coming in the [17:58] form of the the department of energy's genesis  mission. Roughly speaking, this this is the idea [18:04] of connecting together the uh computing facilities  at at all of the DOE labs and making them uh a a [18:13] single uh uh resource for scientists uh to use.  And so we're thinking about uh our contribution [18:20] to the Genesis mission being the LSST data set  and in return using all of these uh facilities [18:26] connected up and providing more computing  resources than we would have had otherwise. [18:30] So, keep an eye on the Genesis mission. I  think it's going to be uh exciting times. [18:37] Okay, remember this is your observatory. You  might want to help out with the science. And so, [18:42] here's another uh place to to go and spend  your time rather than listening to me. Uh [18:47] we're trying to crowdsource some of the um  the scientific activity on the Reubin data [18:54] uh using a citizen science platform called  Zunoveriverse, which some of you may have [18:58] heard of. This is a place where anybody can go  and help out with a science project that includes [19:04] uh visual inspection, visual classification  of uh objects. Um providing labels as you [19:10] might might think of it. So for example, here's  here's a project that we were running recently, [19:15] Reuben Difference Detectives, uh inviting  uh people to come and help label images [19:20] uh as real or bogus in order to help us train  better our machine learning uh classifiers to [19:26] make the alert stream cleaner. uh there'll be  more and more uh citizen science projects on [19:31] this platform over time and definitely encourage  you to come and help out if you have the time. [19:39] Okay, so I wanted to tell you a little bit about  what we're doing at Slack um with our science hats [19:44] on. All of that was me telling you about the  observatory that we've built, the camera that [19:48] we built um assembled in Chile and now and now  running. But in terms of the science at Slack, [19:55] at Slack, we're particularly interested in  this question of um the expansion history of [19:59] the universe. Why is it accelerating? So, we're  looking for ways to measure distance to galaxies [20:05] uh with greater and greater accuracy. And one of  the ways you can do it are with these galaxies [20:10] called uh gravitational lenses. It turns out that  if you have a distant background galaxy that has a [20:18] um uh uh a super massive black hole in it  where gas is falling into that black hole, [20:26] black hole flickers. So you can see it varying in  brightness. If that uh active galactic nucleus is [20:33] sitting right behind a massive galaxy somewhere  in the foreground. So imagine the AGN way off [20:39] in the distance, a massive galaxy here,  and you're over here. If the AGN is right [20:45] behind the massive galaxy, it doesn't get  blocked by the foreground galaxy. Instead, [20:49] the light can come round it this way,  also this way, like in the diagram, [20:55] and you end up seeing multiple copies of the same  background thing. And then you see it flickering, [21:00] but it flickers over here, slightly before over  here. And that time delay turns out you can use [21:06] that to measure the distance to the lens.  So bit convoluted but it's hard to measure [21:12] distance to galaxies and any way we can do  it is worth it. So here we are. We can use [21:17] um time delay lenses to measure distance in the  universe. Right now we have about 10 that are well [21:22] measured and they're giving a a fairly precise  probe of uh precise measurement of the expansion [21:27] rate of the universe. But think what we could  do with thousands of these things. you can make [21:32] a much more accurate measurement and pin down  uh what what's going on with the acceleration. [21:39] So [clears throat] what's that  going to take? So [clears throat] [21:47] modeling a thousand lenses measuring time delays  in in a thousand lenses uh is not going to be [21:52] easy. You can see here uh on the right hand side  a simulation of what um one of these time series [21:58] might might look like for a gravitational lens  where you can see four copies of the back of [22:02] the same background AGM. Uh but perhaps you could  see that there's enough variation there and enough [22:08] uh uh enough measurements to be able to to extract  time delay and use it for for cosmology. So we [22:14] have good time series data. The hardest part is  actually uh modeling the the distribution of mass [22:21] in the deflector so that you can do the distance  measurement. Uh at the moment if you do this with [22:28] a likelihood based uh method, it might take a  couple of days to characterize the posterior PDF [22:34] for the lens model parameters. We need to do that  much faster if we're going to be able to analyze [22:39] thousands of lenses. So what we're doing is using  um simulation based inference uh with um neural [22:47] posterior estimation and that's enabling us to  consider doing this measurement with thousands of [22:52] lenses instead of tens of lenses. So some examples  of how that work. We train the train our neural [22:58] network on um half a million or so simulated  lenses. That's the prior information about the [23:05] model that goes into the um uh into the system.  And then we apply that that train network to tell [23:11] us the uh the measurements of the same parameters  uh on on new data as they come in. So we've been [23:18] testing this for the last uh uh couple of years.  Uh it really does reduce um uh the time it takes [23:24] to model one of these lenses from uh from days to  minutes. So it makes the the sample accessible. [23:31] um we put a lot of effort into um uh how to get  the uh the answer right on average. I keep saying [23:40] that we're trying to make accurate measurements of  the expansion rate of the universe. We don't just [23:44] want high precision. We want to get the answer  right. We want accuracy. And so we we spend a [23:49] lot of time doing verification tests on the that  the training data we use is capable of giving us [23:55] the right answer when we apply it to real data.  Um, and it turns out that when you include this [24:01] this bigger sample, you can can indeed get uh  factors of a few better in in the measurement [24:06] of the the so-called dark energy parameters that  presumably govern this accelerated expansion. [24:17] So, the uh uh getting to the accuracy, uh  there's a plot here that shows um uh look [24:23] at the orange compared to the purple. [clears  throat] The purple is an accurate result. See [24:28] how it lines up with the black crosshairs? The  orange was an inaccurate result. And this is [24:33] in the the population um of the of the lenses  rather than the individual lenses themselves. [24:39] So we have thousands of lenses drawn from some  population. If we don't get that population right, [24:44] if we don't understand the how massive galaxies  work on average, then we get the uh the uh the [24:50] answers wrong about the expansion rate. So we care  very much about this this distribution shift and [24:57] uh current thrust is to try and make our uh  our models more and more expressive in order [25:02] to be able to capture that distribution  shift and allow us to recover from it. So [25:07] we're working with normalizing flows  for the individual uh posterior PDFs [25:11] and looking at more and more complexity and  and the modeling of the population as well. [25:19] So this inference that I just described for the  with the strong lenses, it's actually the same [25:26] one that we envisage doing 10 20 years ago. So the  probabilistic graphical model for the science that [25:32] we're trying to do hasn't changed. The only thing  that's changed is our ability to work at scale, [25:37] which is what the uh the neuroposterior estimation  enables. Um going to that uh uh greater speed [25:46] uh brings a few um uh benefits. Um it allows  us to uh do the inference fast enough to do to [25:54] analyze the sample in the first place. It also  allows us to do the inference fast enough that [25:58] we can redo it over and over again probing for  different systematics associated with different [26:03] assumptions that we make. And remember, we're  after accuracy, which means we have to we have to [26:08] do a lot of verification testing. The simulated um  uh simulation based inference is is nice for other [26:16] reasons though. In building that training set, um  it's basically allowing us to assign a prior PDF [26:22] that we really believe in. we can put a lot of  detail into the generation of that training set, [26:27] make the simulations more and more complex, more  and more realistic as a way to um uh reduce these [26:33] distribution shifts. And so there's a nice uh a  nice way of thinking about the science in terms [26:40] of making the simulations more and more realistic,  matching the real real observations from the real [26:46] universe better and better and in the process  learning more and more about the universe. [26:52] I did want to leave you with this notion of of  the hierarchical inference that you see here [26:56] as being the general solution to to distribution  shift. If you got your distribution wrong in the [27:01] training set, there's a way to correct for  it by uh thinking in terms of hierarchical [27:06] inference where you infer the prior that  you wished you had been able to assign. [27:14] Okay, so we're at Slack, we're doing this  kind of uh um inference with MPE. lots of [27:20] other groups doing it as well. I suspect that  many astronomers will be led down this path in [27:24] order to be able to analyze this LSST data set  that we're that we're building. And so, in fact, [27:30] there'll be a whole generation of students in  um physics departments, astronomy departments, [27:36] uh leaking out into maths and electrical  engineering and so on thinking about [27:40] uh the LSST data set and how to do inference at  this scale. Uh we like to say that um astronomy [27:46] is like a a gateway drug for the STEM field. It  gets people interested in in uh uh STEM early [27:52] and then they can learn about um working with uh  with big data by thinking about the LSST. In fact, [27:58] there's a whole department at Rubin devoted  to education and public outreach that produces [28:03] resources for high school teachers to use to  teach their students about the universe and [28:09] how astrophysical objects work. But it can also  be used to introduce them to big data and start [28:14] teaching them about machine learning and why we  do it. And so that education program is live. Um [28:21] uh it's had their their tutorials, their resources  have had something like 80,000 uh individual uses [28:28] already. And we only just started the survey in  June. So for the next decade, high schools will be [28:34] uh uh full of kids uh learning about big data  by starting to think about the night sky, [28:40] getting into LSST, and then going on from there.  So I wanted to leave you with that thought uh [28:46] that the uh people that you may soon be uh  seeing in your companies or hiring and so on [28:52] uh they may well know about Reuben and the LSST.  So you have something in common with them now to [28:56] talk to them about. Uh but most of all I wanted  you to encourage to to follow along and and join [29:01] us on this journey. This is your observatory.  You did help uh pay for it. Something like uh [29:07] $5 each of your taxpayers money on average. Uh so  thank you for that and do please follow along. You [29:13] can find us at rubenobservatory.com.  Uh you can use the skyviewer.app to [29:17] uh to explore the data. You can come and help  out with the science at the Zuniverse. And I [29:22] hope you uh have a have a great decade of LSST  science with us. Thanks very much for listening.