天文学的AI革命：在海量数据中寻找新天体，让微弱的宇宙信号被看见｜硅谷101年度线下大会

硅谷101

AI-assisted summary of 天文学的AI革命：在海量数据中寻找新天体，让微弱的宇宙信号被看见｜硅谷101年度线下大会 — original source. Check the original for full context.

Source published: 2026-10-28

Key ideas

随着南希·格蕾丝·罗曼空间望远镜和薇拉·鲁宾天文台等设施投入使用，天文学正迎来爆发式数据洪流，其中鲁宾天文台每晚将产生多达700万个瞬变源警报，逼迫学界全面转向AI处理管线。 (11:15)
清华大学团队研发的时域自监督去噪模型使詹姆斯·韦伯空间望远镜的观测深度提升约2.5倍，让其6.5米主镜发挥出相当于20米级虚拟口径的等效能力。 (16:45)
商业航天与民营资本正在重塑深空探索架构，包括采用现成商用部件开发的Lazuli项目，以及利用SpaceX星舰发射10面JWST级别反射镜的Nautilus阵列概念。 (18:25)
天文学公开透明且纯天然的光子数据为前沿AI提供了免疫于模型崩溃（Model Collapse）的真实物理世界数据流，成为前沿基础模型训练的理想土壤。 (20:45)

Detailed notes

## Background 本视频由科技媒体“硅谷101”于2026年10月28日发布于Bilibili平台。内容记录了在硅谷101 Alignment 2026年度大会上，西湖大学助理教授、前詹姆斯·韦伯空间望远镜（JWST）科学团队成员孙凤吾（Fengwu Sun）的主题演讲，分享AI如何助力天文学家应对海量观测数据、识别微弱宇宙信号并革新未来的太空探索模式。

## TL;DR 随着JWST的深场观测以及未来Roman望远镜、薇拉·鲁宾天文台和SKA射电阵列的运行，天文学数据正从TB级向PB级爆发，彻底超越了人工筛选的极限。面对这一数据洪流，AI已成为不可或缺的工具，不仅在引力透镜搜寻、超新星瞬变源筛选和望远镜自主调度中发挥关键作用，还通过自监督时域降噪算法将JWST 6.5米主镜的等效探测深度提升至20米级别。与此同时，以Project Lazuli和星舰星座为代表的商业航天正在打破传统航天研发周期，而天文学公开、纯净且无版权限制的真实光子数据，也为避免前沿AI出现“模型崩溃”提供了绝佳的基底。

## Atmosphere & style 演讲兼具前沿学术报告的严谨与硬科技大会的澎湃想象力，氛围开阔且引人入胜。孙凤吾以第一线观测者的视角娓娓道来，既有凝视百亿年前宇宙微光时的宏大与敬畏，也有面对每晚数百万条警报邮件时科研人员的真实焦虑与诙谐调侃。整场分享从硬核工程制造延伸至底层算法重构，充满理性思辨的同时洋溢着仰望星空的浪漫情怀。

## Bulletin - [11:15](https://www.bilibili.com/video/BV1QMxj6wE2N?t=675s) 随着南希·格蕾丝·罗曼空间望远镜和薇拉·鲁宾天文台等设施投入使用，天文学正迎来爆发式数据洪流，其中鲁宾天文台每晚将产生多达700万个瞬变源警报，逼迫学界全面转向AI处理管线。 - [16:45](https://www.bilibili.com/video/BV1QMxj6wE2N?t=1005s) 清华大学团队研发的时域自监督去噪模型使詹姆斯·韦伯空间望远镜的观测深度提升约2.5倍，让其6.5米主镜发挥出相当于20米级虚拟口径的等效能力。 - [18:25](https://www.bilibili.com/video/BV1QMxj6wE2N?t=1105s) 商业航天与民营资本正在重塑深空探索架构，包括采用现成商用部件开发的Lazuli项目，以及利用SpaceX星舰发射10面JWST级别反射镜的Nautilus阵列概念。 - [20:45](https://www.bilibili.com/video/BV1QMxj6wE2N?t=1245s) 天文学公开透明且纯天然的光子数据为前沿AI提供了免疫于模型崩溃（Model Collapse）的真实物理世界数据流，成为前沿基础模型训练的理想土壤。

## Key takeaways - [00:00](https://www.bilibili.com/video/BV1QMxj6wE2N?t=0s) 詹姆斯·韦伯空间望远镜（JWST）历经约25年建造与测试，运行在距地150万公里的第二拉格朗日点（L2），通过红外深场长曝光捕捉早期宇宙信号。 - [05:00](https://www.bilibili.com/video/BV1QMxj6wE2N?t=300s) 结合星系团引力透镜放大效应，科研团队定位了宇宙大爆炸后仅3亿年的早期星系，并成功探测到其恒星演化产生的氧元素信号。 - [05:00](https://www.bilibili.com/video/BV1QMxj6wE2N?t=300s) JWST发现了新型高亮度致密天体“小红点”（Little Red Dot），其超强能量被推测来源于超大质量黑洞对气体的剧烈吸积。 - [08:15](https://www.bilibili.com/video/BV1QMxj6wE2N?t=495s) 南希·格蕾丝·罗曼空间望远镜搭载3亿像素焦平面检测器，以每天1.4 TB（JWST的20倍）的数据产出速度巡天搜寻超新星，构建四维时空地图。 - [11:15](https://www.bilibili.com/video/BV1QMxj6wE2N?t=675s) 薇拉·鲁宾天文台配备人类最大的32亿像素数码相机，全速运转时每晚将产生约700万个警报，彻底打破了传统天文学家人工筛查的可能性。 - [15:00](https://www.bilibili.com/video/BV1QMxj6wE2N?t=900s) 机器学习算法在全天巡天中能一次性系统筛选约400个强引力透镜候选体，取代了过去人工挑选并发表论文的低效模式。 - [15:00](https://www.bilibili.com/video/BV1QMxj6wE2N?t=900s) AI正从辅助科研工具转变为自主运行智能体，自动承担环境监测、圆顶防险与超新星瞬变源自主指向等传统人工观测任务。 - [16:45](https://www.bilibili.com/video/BV1QMxj6wE2N?t=1005s) 清华大学合作团队利用自监督噪声模型分离残余探测器与背景噪声，算法层面将JWST的有效灵敏度提高了2.5倍。 - [16:45](https://www.bilibili.com/video/BV1QMxj6wE2N?t=1005s) 文献计量研究显示超过50%的天体物理学论文带有大语言模型写作痕迹，但论文中对LLM使用的正式披露率仅为0.8%。 - [18:25](https://www.bilibili.com/video/BV1QMxj6wE2N?t=1105s) 以Eric Schmidt资助的Project Lazuli为代表的民营资本项目，通过采购商用现成组件，将望远镜开发周期压缩至2到3年。 - [18:25](https://www.bilibili.com/video/BV1QMxj6wE2N?t=1105s) Nautilus概念探索利用SpaceX星舰货运空间一次性装载并部署10面6.5米轻量化反射镜，开启模块化低成本深空观测新模式。 - [20:45](https://www.bilibili.com/video/BV1QMxj6wE2N?t=1245s) 天文学所有原始光子数据均无版权限制地公开开放，且完全来自于真实的非地球物理信号，能有效解决合成数据递归训练引发的AI模型崩溃风险。

## Detailed walkthrough ### [00:00](https://www.bilibili.com/video/BV1QMxj6wE2N?t=0s) 詹姆斯·韦伯空间望远镜（JWST）的建造与深空观测

演讲嘉宾孙凤吾（Fengwu Sun）现任西湖大学助理教授，此前在亚利桑那大学（University of Arizona）担任博士后研究员，并曾是詹姆斯·韦伯空间望远镜（JWST）科学团队的成员。JWST 是人类历史上最大、最先进的空间红外望远镜。这项宏伟的计划早在 1996 年便已立项，历时近 30 年（其中建造与测试阶段花费了约 25 年），最终于 2021 年底由阿丽亚娜 5 号（Ariane 5）运载火箭从法属圭亚那的库鲁（Kourou）航天发射中心成功发射升空。

孙凤吾展示了望远镜在地面建造时的珍贵照片。在巨大的金色主镜面前，人类工程师显得如同蚂蚁般渺小。由于 JWST 主要在红外波段工作，为了避免望远镜自身热量产生的红外辐射干扰，它必须在接近绝对零度的极低温度下运行。孙凤吾的研究工作主要集中在“深场”（Deep Field）观测。深场观测的原理是让望远镜长时间凝视天空中看似虚无、没有亮星的区域，通过极长时间的曝光积攒光子，从而揭示百亿光年之外、包含数以十亿计恒星的极早期星系。

“and the telescope was designed back in 1996 and it take almost 30 years to 25 years to build the telescope, test everything and launch to the space with the Arianna 5.”

- **研发历程**：JWST 立项于 1996 年，历经约 25 年的建造与测试，耗资巨大。 - **发射细节**：由阿丽亚娜 5 号火箭在法属圭亚那库鲁发射，运行于距离地球 150 万公里的第二拉格朗日点（L2）。 - **深场观测**：通过长时间曝光，在黑暗的宇宙背景中寻找最古老、最微弱的星系信号。

---

[05:00](https://www.bilibili.com/video/BV1QMxj6wE2N?t=300s) 引力透镜、最遥远星系与神秘的“小红点”

天文学家寻找宇宙初期天体主要有两种方法：第一种是直接将望远镜对准黑暗虚无的深空进行长时间曝光；第二种则是更为聪明的“引力透镜”法。这种方法将 JWST 对准一个巨大的星系团（Galaxy Cluster），利用星系团巨大质量产生的引力场作为天然的“放大镜”。通过将空间望远镜与宇宙天然放大镜相结合，科学家得以窥见宇宙诞生初期极度微弱、遥远的天体。利用这种方法，孙凤吾所在的团队在宇宙深处成功定位了一个微小的红点，并通过光谱分析证实该星系存在于大爆炸后仅 3 亿年（当时宇宙的年龄仅为现在的 2% 到 3%）。更令人惊叹的是，在这个极早期星系的光谱中，科学家检测到了氧元素的信号，这意味着在宇宙诞生初期，恒星的演化与爆炸就已经开始产生我们今天呼吸所必需的氧气。

除了最遥远的星系，JWST 还发现了一类被称为“小红点”（Little Red Dot）的新型天体。天文学家在命名上非常直接，因为它们在图像中看起来确实既小、又红，且呈点状。这些天体虽然在特征上极像普通恒星，但由于它们距离极远且亮度极高，科学家相信其能量实际上来自超大质量黑洞吞噬气体时释放的引力能。此外，孙凤吾还展示了邻近星系中由超新星爆发吹出的美丽气泡，以及 JWST 拍摄的极具震撼力的木星（展现了大红斑与两极的极光）和土星图像。

“The galaxy has already produced oxygen by exploring stars at that time, and they produce all of the oxygen that we are breathing today in this room, starting from the 300 million years after the Big Bang.”

- **大爆炸时间线**：宇宙历史约为 138 亿年，新发现的遥远星系存在于大爆炸后 3 亿年。 - **小红点（Little Red Dot）**：一类由黑洞吸积气体提供能量的超亮、遥远致密天体。 - **行星观测**：JWST 拍摄的木星图像清晰展现了伟大的大红斑（Great Red Spot）以及南北两极的极光（Aurora）。

---

[08:15](https://www.bilibili.com/video/BV1QMxj6wE2N?t=495s) 走向四维宇宙与 Nancy Grace Roman 空间望远镜

尽管天文学家目前获取的多为二维图像，但宇宙本质上是四维的（三维空间加上时间维度）。为了理解宇宙的演化，天文学家正致力于为整个宇宙“拍摄电影”。例如，通过在两年内持续观测同一片天空，科学家可以识别出在第一年突然亮起、第二年又暗淡下去的超新星（Supernovae）。超新星在天文学中被称为“标准烛光”，通过寻找它们，科学家可以测量宇宙不同位置的距离，从而绘制出宇宙的三维地图。

为了实现这一目标，NASA 计划发射南希·格蕾丝·罗曼空间望远镜（Nancy Grace Roman Space Telescope）。虽然它的口径仅为 2.4 米（与哈勃望远镜相当，不及 JWST 的 6.5 米），但它配备了极其强大的焦平面检测器。该检测器由 18 个 4K x 4K 像素的芯片组成，单次曝光即可获得 3 亿像素的超大图像。其数据产生速度预计达到每天 1.4 TB，约为 JWST 的 20 倍。这些数据必须从距离地球 150 万公里的 L2 轨道传输回地面，对数据传输和处理带来了巨大挑战。

“And one of the main philosophers of Roman Space Telescope is try to make a movie of the universe by finding a supernovae, which by finding them, we sort of know they are like candles.”

- **焦平面配置**：Roman 望远镜拥有 18 个检测器组成的焦平面，单次成像达 3 亿像素。 - **数据率爆发**：Roman 每天产生约 1.4 TB 的数据，数据产生率是 JWST 的 20 倍。 - **科学目标**：通过大视场巡天寻找超新星，利用“标准烛光”效应绘制宇宙地图。

```mermaid flowchart LR accTitle: 天文学数据洪流演进路径 accDescr: 展示从JWST到未来SKA的数据量爆发式增长过程。

---

[11:15](https://www.bilibili.com/video/BV1QMxj6wE2N?t=675s) 薇拉·鲁宾天文台与即将到来的天文数据洪流

除了空间望远镜，地面上的薇拉·鲁宾天文台（Vera C. Rubin Observatory）同样在引发一场数据革命。该天文台建于智利的山区，配备了人类历史上制造过的最大的数码相机，分辨率高达 32 亿像素（3.2 Giga pixels），是 Roman 望远镜的 10 倍。鲁宾天文台拥有一面 8.5 米口径的主镜，目前每晚产生 10 到 20 TB 的海量数据。

鲁宾天文台的核心任务之一是进行“时域天文学”观测，通过对比不同时间拍摄的图像来寻找移动的天体（如小行星 Asteroid）或瞬变源。目前，它每晚会产生大约 100 万个警报（Alerts），当其全速运转时，这一数字将飙升至每晚 700 万个。面对如此庞大的警报数量，全球仅有的几千名天文学家根本无法人工处理——这意味着每位天文学家每天早上醒来都要处理数百封警报邮件，这显然是不可持续的。随着 2025 年后 Roman 和 Rubin 的全面运行，以及 2030 年左右平方公里阵列（SKA）射电望远镜的加入（届时每天将产生数 PB 的数据），天文学界正迎来前所未有的数据洪流，而 AI 将成为解决这一瓶颈的关键工具。

“So if all of us should work on the data, every night we have to reply 200 emails about the alerts. So it's definitely not scalable.”

- **超大相机**：薇拉·鲁宾天文台配备 32 亿像素相机，是人类历史上制造过的最大相机。 - **警报过载**：全速运转时，鲁宾天文台每晚将产生 700 万个瞬变源警报，远超人工处理极限。 - **未来展望**：预计到 2030 年左右，平方公里阵列（SKA）射电望远镜每天将产生数 PB（Petabytes）的数据，彻底改变天文学的研究范式。

[15:00](https://www.bilibili.com/video/BV1QMxj6wE2N?t=900s) Tackling the Data Flood: Gravitational Lensing, New Observing Modes, and AI-Driven Telescopes

Astronomers are currently confronting an unprecedented torrent of observational data. With new sky surveys generating petabytes of raw signals every single day, it has become physically impossible for human scientists to inspect even a fraction of what is collected. Fengwu Sun emphasizes that modern astrophysics is experiencing a severe data flood that leaves researchers completely out of human attention bandwidth, turning artificial intelligence from a luxury into an absolute necessity. To illustrate this shift, Sun points to the earliest deep-field images released by the James Webb Space Telescope (JWST), where massive galaxy clusters act as natural gravitational lenses, bending background light into curved, glowing arcs. In the past, finding such systems was an agonizingly slow, manual craft: a decade ago, a graduate student could identify just one or two gravitational lens systems in archival data, turn them into two peer-reviewed papers, and walk away with a PhD. Today, researchers must deploy machine learning pipelines capable of systematically scanning whole surveys to detect around 400 lens candidates at once—vital samples needed to measure precise cosmic distances and track how the universe expands over cosmic time.

```mermaid flowchart TD accTitle: AI in Astronomical Data and Operations accDescr: Diagram showing how AI processes raw space data and automates telescope tasking. subgraph DataOps["Observation & Noise Processing"] A["Deep Time-Domain Stacks"] --> B["Self-Supervised Noise Model"] B --> C["2.5x Deeper Signal (6.5m JWST feels like ~20m)"] end subgraph TelescopeControl["Autonomous Telescope Operations"] D["Sky Monitoring & Weather"] --> E["AI Decision Engine"] F["Transient Alerts (e.g. Supernovae)"] --> E E --> G["Autonomous Slewing & Task Scheduling"] end ```

Beyond finding hidden astronomical objects, AI is transforming how telescopes are designed to operate. Sun shares his own experience as a heavy user of tools like OpenAI Codex and Anthropic's Claude to analyze data taken in untargeted or wide-field observing modes on the James Webb Space Telescope. By training AI to predict where dispersed light and spectra fall across the detector, his team engineered a completely new JWST observing mode without needing hardware modifications. Concurrently, machine learning is taking over real-time observatory operations—a role human astronomers performed for four centuries. Rather than relying on a resident observer to pick targets, close the dome when clouds roll in, or scramble to repoint the mirror when a sudden supernova flares up, automated AI decision loops now handle dynamic scheduling, environmental monitoring, and immediate transient follow-up. Sun puts the urgency bluntly:

> “We have to rely on AI because we are running out of our attention.”

- **Gravitational lensing yields**: ML algorithms now identify roughly 400 strong gravitational lenses in survey volumes where human vetting previously yielded only one or two systems. - **Spectrum mapping**: Neural models predict the dispersion coordinates of slitless grism spectra on JWST detectors, unlocking novel observing techniques. - **Autonomous tasking**: Machine agents manage dome safety, cloud avoidance, and automated slewing toward transient explosive events like supernovae.

---

[16:45](https://www.bilibili.com/video/BV1QMxj6wE2N?t=1005s) Supercharging JWST Aperture with Deep Noise Filtering and LLMs in Scientific Publishing

A striking technical breakthrough discussed by Sun comes from his collaborators at Tsinghua University, who devised a self-supervised model tailored to temporal noise patterns. Standard astronomical imaging relies on stacking multiple exposures of the same patch of sky to construct a deep two-dimensional image. However, by analyzing the fluctuations of noise across the time domain rather than simply averaging pixels together, the Tsinghua team’s model learns to isolate and strip away residual detector and background noise far more effectively. The algorithm pushes the effective sensitivity of the data deeper by roughly a factor of 2.5. In astronomical terms, this computational boost makes the 6.5-meter primary mirror of the James Webb Space Telescope perform on par with an imaginary 20-meter space telescope—an aperture that current rocket fairings and space budgets cannot physically deliver to orbit.

At the same time, generative AI has thoroughly infiltrated the scientific community's day-to-day writing habits. Sun highlights a recent meta-science study examining the language patterns in preprint and published astronomy literature over the past two years. The investigation revealed that more than half of recent astrophysics papers show unmistakable stylistic traces of large language models. Curiously, while LLM-assisted drafting has become an open secret across academic departments, only 0.8% of papers explicitly acknowledged LLM usage in their methodologies or disclosures. Sun jokingly reassures the audience that astronomers are not deliberately deceiving one another, but rather struggling to adapt to a research paradigm shifting faster than professional norms can accommodate. Reflecting on this rapid, unstandardized transition, Sun remarks:

> “So I'm not saying that the astronomers are cheating to each other these days, but actually the field is changing so fast that we don't know what is the best way to proceed.”

- **Effective aperture gain**: Self-supervised temporal noise modeling yields an observational depth gain of ~2.5×, making JWST's 6.5-meter mirror behave like a ~20-meter aperture. - **LLM adoption in literature**: More than 50% of evaluated astrophysics publications exhibit statistical hallmarks of large language model text generation. - **Disclosure deficit**: Despite widespread LLM writing assistance, only 0.8% of published astronomy papers provide formal disclosures.

---

[18:25](https://www.bilibili.com/video/BV1QMxj6wE2N?t=1105s) The New Space Architecture: Private Capital, Project Lazuli, and Starship Constellations

Historically, space astronomy progressed at a glacial hardware cadence. While terrestrial computing power exploded over the last sixty years, flagship space missions remained tethered to multi-decade development lifecycles. Sun reminds the audience that the Hubble Space Telescope was originally designed during the 1970s and launched in 1990; yet astronomers in the mid-2020s are still competing for observation proposals on it, NASA continues studying life-extension missions, and the observatory will not retire until at least 2030. Today, however, that slow institutional model is being disrupted by private capital, rapid industrial prototyping, and commercial supply chains.

A prime example is Project Lazuli, funded by former Google CEO Eric Schmidt and led in partnership with 2011 physics Nobel laureate Saul Perlmutter alongside instrument teams at the University of Arizona. Designed as the world's first fully private orbital space observatory, Lazuli bypasses traditional aerospace contracting by purchasing commercial off-the-shelf components, developing under strict non-disclosure agreements, and aiming for launch within two to three years. Even more ambitious architectures are emerging to exploit super-heavy launch vehicles like SpaceX's Starship. Sun references the Nautilus Space Observatory concept conceived by University of Arizona researchers: instead of building a single bespoke multi-billion-dollar observatory, engineers can fold ten 6.5-meter segmented or lightweight mirrors into a single Starship fairing. Upon deployment in orbit, the array expands into ten JWST-class apertures simultaneously. Mass-manufacturing modular, 50-centimeter sub-elements drives unit costs down exponentially, unlocking rapid iteration cycles that mimic Silicon Valley hardware rather than traditional government space programs.

- **Hubble longevity**: Conceived in the 1970s and launched in 1990, Hubble remains an active frontline research tool through at least 2030. - **Project Lazuli**: Backed by Eric Schmidt, advised by Saul Perlmutter, and developed with commercial off-the-shelf parts to launch within 2–3 years. - **Nautilus concept**: Packs 10 deployable 6.5-meter mirrors inside a single SpaceX Starship payload envelope, drastically reducing average per-observatory production costs.

---

[20:45](https://www.bilibili.com/video/BV1QMxj6wE2N?t=1245s) Open Science and Real Photons: Why Astronomy Is Immune to Model Collapse

Wrapping up his presentation at the Alignment 2026 conference, Sun turns to the artificial intelligence researchers and industry leaders in the room, laying out why astronomy serves as a unique, symbiotic partner for frontier AI. Astronomical research provides massive, pristine empirical datasets: every single photon gathered by modern space and ground observatories is released to the public without paywalls, commercial barriers, or copyright claims. Because the universe does not label its own phenomena, astronomical archives represent an enormous playground of uncurated, unlabelled, real-world signals waiting for unsupervised and self-supervised learning algorithms to crack open.

Most compellingly, Sun points out that astronomy offers a structural cure to one of the biggest perils facing foundational models: model collapse caused by training iteratively on synthetic, recursively generated internet data. While web-scraped corpora increasingly cannibalize AI-generated text and imagery, observatories capture genuine, non-terrestrial photons emitted billions of years ago. Looking back across four centuries of observational science, Sun envisions AI not as a replacement for scientific inquiry, but as the perceptual engine that will allow humanity to process cosmic signals far beyond human imagination:

> “And if you believe that model collapse when trained on recursively regenerated data, then I will tell you that the astronomers are collecting fresh data every day and from nowhere from the planet.”

- **Unlabeled empirical streams**: Space telescopes provide pure, uncurated data devoid of synthetic contamination or human annotation bias. - **Total open access**: Astronomical institutions publicly release raw photon counts without copyright gates, offering open scientific ground for foundational model training. - **Historical milestone**: Modern astronomy enters its fifth century of systematic observation, using autonomous AI to push observational clarity past physical detector limits.

## Notable quotes > [15:00](https://www.bilibili.com/video/BV1QMxj6wE2N?t=900s) We have to rely on AI because we are running out of our attention. — Fengwu Sun

## Mentioned Fengwu Sun（孙凤吾）, 西湖大学, 亚利桑那大学, 清华大学, 詹姆斯·韦伯空间望远镜（JWST）, 阿丽亚娜 5 号（Ariane 5）, 哈勃空间望远镜（Hubble Space Telescope）, 南希·格蕾丝·罗曼空间望远镜（Nancy Grace Roman Space Telescope）, 薇拉·鲁宾天文台（Vera C. Rubin Observatory）, 平方公里阵列（SKA）, OpenAI Codex, Anthropic Claude, Eric Schmidt, Saul Perlmutter, Project Lazuli, Nautilus 空间望远镜, SpaceX Starship

## Video description 人类正在迎来一个全新的宇宙探索时代。新一代观测设施正以前所未有的速度，收集来自宇宙深处的数据。每天数TB甚至数十TB的图像和观测信号，远远超出了科学家人工分析的能力，AI正在成为天文学家探索宇宙的新工具。在硅谷101 Alignment 2026大会上，西湖大学助理教授Fengwu Sun介绍了AI如何帮助科学家识别引力透镜、寻找超新星，以及降低图像噪声，让微弱的天体信号更容易被发现。他也展望了AI参与望远镜自主观测和任务规划的可能性。当AI开始帮助人类探索宇宙，我们又将发现哪些过去难以看见的秘密？这期，让我们回到Alignment 2026大会现场，听听Fengwu Sun的分享。

大会嘉宾： Fengwu Sun，西湖大学助理教授

[Source](https://www.bilibili.com/video/BV1QMxj6wE2N)

中文简报

天文学的AI革命：在海量数据中寻找新天体，让微弱的宇宙信号被看见｜硅谷101年度线下大会随着JWST的深场观测以及未来Roman望远镜、薇拉·鲁宾天文台和SKA射电阵列的运行，天文学数据正从TB级向PB级爆发，彻底超越了人工筛选的极限。面对这一数据洪流，AI已成为不可或缺的工具，不仅在引力透镜搜寻、超新星瞬变源筛选和望远镜自主调度中发挥关键作用，还通过自监督时域降噪算法将JWST 6.5米主镜的等效探测深度提升至20米级别。与此同时，以Project Lazuli和星舰星座为代表的商业航天正在打破传统航天研发周期，而天文学公开、纯净且无版权限制的真实光子数据，也为避免前沿AI出现“模型崩溃”提供了绝佳的基底。随着南希·格蕾丝·罗曼空间望远镜和薇拉·鲁宾天文台等设施投入使用，天文学正迎来爆发式数据洪流，其中鲁宾天文台每晚将产生多达700万个瞬变源警报，逼迫学界全面转向AI处理管线。 (11:15) 清华大学团队研发的时域自监督去噪模型使詹姆斯·韦伯空间望远镜的观测深度提升约2.5倍，让其6.5米主镜发挥出相当于20米级虚拟口径的等效能力。 (16:45) 商业航天与民营资本正在重塑深空探索架构，包括采用现成商用部件开发的Lazuli项目，以及利用SpaceX星舰发射10面JWST级别反射镜的Nautilus阵列概念。 (18:25) 天文学公开透明且纯天然的光子数据为前沿AI提供了免疫于模型崩溃（Model Collapse）的真实物理世界数据流，成为前沿基础模型训练的理想土壤。 (20:45)

Reading summary

西湖大学助理教授、前詹姆斯·韦伯空间望远镜科学团队成员孙凤吾指出，伴随新一代空间与地面望远镜进入服役期，天文数据正迎来几何级爆发并击穿人类处理极限；人工智能已从辅助工具转变为不可或缺的基础设施，在突破物理口径限制、实现自主化观测的同时，天文数据也将以其纯净物理特质反哺AI基础模型。深空探测突破与早期宇宙极端天体 (0:00) 历经近30年立项与建造测试的詹姆斯·韦伯空间望远镜（JWST）于2021年底发射升空并定轨于第二拉格朗日点（L2）。通过深场长时间曝光与星系团引力透镜技术，天文学家得以捕捉宇宙诞生初期的微弱信号。孙凤吾团队利用星系团的“天然放大镜”效应，成功证实了一个存在于大爆炸后仅3亿年（当时宇宙年龄仅为现在的2%至3%）的早期星系，并在其光谱中检测出恒星演化爆炸产生的氧元素信号，证实早期宇宙已启动重元素合成。此外，JWST还探测到了被称为“小红点”（Little Red Dot）的新型致密天体。这类天体虽呈恒星样点状，但因距离极远且亮度奇高，科学家推测其能量源于超大质量黑洞剧烈吞噬气体所释放的引力能。JWST同时呈现了邻近星系超新星气泡以及木星大红斑与两极极光的高分辨率图像，展示了红外深空观测在极端天体物理研究中的关键突破。数据洪流爆发与望远镜自主化运维 (11:15) 天文学正从静态二维成像转向捕捉四维时域演化。计划发射的南希·格蕾丝·罗曼（Roman）空间望远镜配备3亿像素焦平面检测器，每天产生约1.4 TB数据（约为JWST的20倍），旨在通过寻找超新星“标准烛光”构建宇宙三维地图；地面上智利薇拉·鲁宾天文台搭载32亿像素相机，每晚产生10至20 TB数据与高达700万个瞬变源警报；预计2030年左右落成的平方公里阵列（SKA）日产数据更将达到PB级。全球数千名天文学家的人工审阅带宽已被海量数据完全击穿。为解决注意力过载，机器学习算法被全面引入：在巡天数据中，算法能一次性筛选出约400个强引力透镜候选体，而过去单名学者数年仅能手动辨识一至两个；在观测运行端，AI决策系统已开始接管天气监控、圆顶开闭及超新星瞬变事件的自主寻靶转向，重构了运行四百余年的传统天文观测流程。算法突破硬件限制与学术规范演变 (16:45) 在硬件受限的背景下，计算模型展现出直接倍增望远镜性能的能力。孙凤吾介绍，清华大学合作团队开发了一种针对时间序列噪声模式的自监督模型。该模型不再采用传统的像素平均叠加方式，而是深入分析噪声在时域维度的动态波动并予以剔除，使数据信噪深度有效提升约2.5倍。该算法突破在效果上使JWST现有的6.5米主镜等效于一台约20米口径的虚拟空间望远镜，大幅超越了现有火箭整流罩所能搭载的物理极限。与此同时，生成式AI工具已深度渗透天文学术研究。孙凤吾引用的元科学调研显示，近两年的天体物理预印本与期刊论文中，超过50%的文章呈现出大语言模型写作的典型统计特征，然而仅有0.8%的论文在方法或致谢中进行了明确声明。这种广泛应用与极低披露率的反差，表明学术规范与伦理共识的发展明显落后于科研人员采纳前沿AI技术的实际步伐。商业航天架构演进与抗模型崩溃价值 (18:25) 传统大科学工程耗时漫长，如哈勃望远镜从1970年代设计、1990年发射，服役期将至少延续至2030年。当前，商业资本与敏捷供应链正重构空间天文架构。前Google CEO埃里克·施密特资助的Project Lazuli跳过传统航天采购周期，利用商用现成组件力争两至三年内发射；亚利桑那大学提出的Nautilus概念则利用SpaceX星舰超大有效载荷，计划单次折叠发射10面由模块化子单元构成的6.5米轻量化镜面，以工业化量产大幅摊薄建造成本。孙凤吾强调，天文学同时为前沿人工智能提供了不可替代的反哺价值。全球天文机构公开释放全部无版权、未标注的真实光子数据，是无监督与自监督模型训练的理想土壤。更重要的是，在互联网语料因充斥递归生成内容而引发AI“模型崩溃”隐忧的背景下，天文望远镜每日从地球之外捕获的、未受合成污染的真实物理信号，为前沿基础模型演进提供了抵御崩溃的纯净数据来源。现代天文观测在跨入第五个百年之际，正经历硬件研制周期压缩、观测数据指数级暴增与研究范式智能化的深层转轨。孙凤吾总结认为，AI并非取代科学家对宇宙本质的思考，而是作为感知引擎拓宽认知边界；而天文学完全开放且源源不绝的纯净物理数据，也将在算法演进的关键节点为基础AI研究提供坚实的实证锚点。

Source quotes

[15:00] Fengwu Sun: We have to rely on AI because we are running out of our attention.

Original transcript

[0:00] 人类正在迎来一个全新的宇宙探索时代,新一代观测设施正在以前所未有的速度收集来自宇宙深处的数据,但一个新的问题也随之出现,这些望远镜每天可能产生数TB甚至数10TB的数据,它们远远超出了科学家人工分析的能力,而AI正在成为天文学家探索宇宙的新工具。接下来就让我们回到硅谷101Alignment 2026带回现场来看看这场精彩的演讲优优独播剧场——YoYo Television Series Exclusive大家好,我是弗乌尔,我身上的课程中的新北线上,中国的新北线上,中国的新北线上,也就是一个很高兴的地方。我之前是在大学的高党,我曾经在大学的高党,我曾经在大学的高党,是一位GWC的GWC中心理科目的工作。GWC中心理科目是最大的名字。那是最大的大型的船目,我们在世界中的船目中的船目。GWC中心理科目是最大的主要用于GWC中心理科目。今天我会将你所说的一些努力,包括了其他其他人的画面,以及以及我们在教我们学习的努力,100%一千万年 away from us, and each of the galaxies will have another billion of stars.and the telescope was designed back in 1996 and it take almost 30 years to 25 years to build the telescope, test everything and launch to the space with the Arianna 5.啊, in Guru基本上,我们是这些工人在工厅的工厅。你可以看到这些人会有很小的码石在手上,但它背景是很高的。看起来,它是300K维的黑白白,如果你想要看什么在深圳的背景,它背景是很高的。从 James West的空间它们叫做一个科目的科目因为你能看到看起来的颜色的颜色的颜色的颜色在另一方面,小星星的星星上有所在的方向, 而在其中的方向, 它是在其中的方向。而是如果您可以看出的画面, 例如有明显的画面, 例如有明显的画面, 例如有明显的画面。您可以看到了明显的画面。上方的画面, 例如有明显的画面。但是,在这里, 例如有明显的画面, 例如有明显的画面。但是,在这里, 例如有明显的画面, 例如有明显的画面。它能够尖尬地的毒更容易你会看到毒在营业,但不在营业所以那是能量到的力量。我讨论了一部分,我的研究是在DFIELD。我们的研究是我们的研究是我们的研究。我们是在研究中的第一个研究。 [5:00] Atronomic objects that form in the universe and where you can find them and how do they look like. And usually we do two ways. One, we just point telescopes to some place that do not have a bright star and just integrate as long as possible. Just take the data and stare at it. And that will be some image on the right. And there are a little bit smarter way that you actually point into a structure we call the galaxy cluster. gravitational potential of this galaxy cluster will form something like a magnifying glass. So you are actually pointing a magnifying glass, which is GWT, to another magnifying glass, which is a galaxy cluster, to find very faint, very far away. atronomic object at the beginning of the universe. And actually, we were quite successful with GWC. That's the whole purpose of why GWC was built. So this is one of the deepest images that our team received. And just in the middle of nowhere, we identify a small red dot. And we split the letter at different wavelengths, and confirmed this galaxy existed 300 million years after the Big Bang. Right now the universe has a history of roughly 13.8 billion years, and this is roughly 2 or 3 percent of the--at the 2 or 3 percent of the age when the universe just formed. Thank you. And most amazingly, if you observe at other different wavelengths, you actually get oxygen detection in the spectrum. And think about this. The galaxy has already produced oxygen by exploring stars at that time, and they produce all of the oxygen. that we are breathing today in this room. starting from the 300 million years after the Big Bang. And that was not-- when we discovered this object, this was the most distant galaxy we know. And right now, this is the most distant object that we have confirmed so far, a little bit more further. And still, in the vast universe with black background, you look through each of the objects, measure each of their distance, and you finally find one target through such kind of a large data set. Another interesting thing that the astronomers are paying attention these days is if you go to look back at the first image that the White House released, Actually, in this region, there's a small red compact object in this region. We were very curious about what that is when the data was released. And later, people, astronomers come up with a name for that. And you probably know that astronomers are very bad in giving names. And we just call them little red dot. because it's little, and also red, and also looked like a dot. These are a new type of star that the GwC identified. They look like a star, exactly like a star, in nearly all of the features. But actually, we believe, because they are so far away and so bright, they are actually powered by black hole. So it's considered that the black hole is creating gas, just just eating gas. And then when the gas fall into the black hole, it's emitting light, turning the gravitational potential into the light. uh if you go back to the the place that is near to us you will see beautiful galaxy like this and you see so many bubbles Each of the bubbles was blown away by a supernovae that the star died and explode. maybe a few hundred million years ago. And they also produce oxygen carbon that make all of the life on this planet. And speaking about my favorite image, this is actually Jupiter. So since 1609, When the first astronomer Galileo pointed his telescope into the sky, this is one of the first objects that he observed. And since then, every telescope actually observed Jupiter. And this is the best image of Jupiter I have ever seen. This is the great red spot. And on the top and bottom, that's actually the aurora light. So the northern light. If you stand on the Jupiter at the time, you will see aurora above your head. and This is the Saturn that you can see the ring and the gap. OK, so I'm showing a lot of images. They are 2D images, right? But we know that the universe is four dimensional, right? You have a space, a three-axis, and also it's a function of time. So it's 4D. And astronomers are making-- working very hard to try to make a movie of the whole universe. And perhaps that's a more informative for the world model that we just discussed, right? And the thing about this, we are actually working on this UNMIS, the James Webb Space Telescope. So you can look at the same region of the sky over two years, and you can identify some adult that goes off in one year. and fade away in the other year. And we call that supernovae, basically a star end of their life. [10:00] And one of the new space telescopes that NASA is launching, NASA launched in August 30th, is so-called Nancy Grace Roman Space Telescope. And one of the main philosophers of Roman Space Telescope. is try to make a movie of the universe by finding a supernovae, which by finding them, we sort of know they are like candles. You can put candles at different places and mirror the distance to it. So it actually helps us to map the universe. So this is the artist's impression of the Roman Space Telescope. It's not as large as the GWSC. It's only 2.4 meter across. But the power of the Roman Space Telescope has very beautiful detectors. So each of them, this is what we call a focal plan. They have 18 detectors, basically the same detector, similar detector as you use on a phone. Each of them is 4K times 4K pixel. So combine all of these 18 chips together, you are actually going to 300 million pixels at once. And consider if we keep observing the sky, the data that it will produce is roughly 1.4 terabytes per day. And that's roughly 20 times how GLC is producing data. And remember, these data do not happen on the ground. They happen at L2, which is 1.5 million kilometers away. And you have to transfer the data from the deep space to the ground. It's actually already taking data right now. So these are data that before the telescope is focused. is a star, but before the focus. So the team is working very hard in focusing the telescope and taking the real data, make sure telescope works. For us who are searching for the first light of the universe, we care about how Roman is observing the universe at the very beginning. So here I'm showing you a numerical simulation obtained by our astronomers. And all of the background image are the gas, the hydrogen is at that time. on the map. They are not randomly distributed, but they form structures. And all of these design circles represent one galaxy that should be detected by Roman. And think about it. This is the first image of the GWC that you just saw. And the GWC can only observe such a very small patch of sky. But Ruomen will get to... get a very large view of it. And basically, that shows how the data rate expansion in the astronomy is going on over the upcoming years. Another very important infrastructure in astronomy is so-called Viral Rubin Observatory. which is conducting a survey called the legacy survey of a space of time. So this is a telescope that is in the mountain of Chilean, and it's a ground-based telescope not launched in space. It's doing survey every night right now. The special thing with this telescope is the camera, is the biggest camera that a human being ever built. So it's 3.2 giga pixels, which is 10 times more than what they have in Roman. So it's actually producing data right now at the rate of 10 or 20 terabytes per day. And the telescope is 8.5 meter, so it's one of the largest telescopes that are built on the ground. And you can see how the mirror of the telescope, how large is this camera, this is the camera, compared with the engineering team on the ground. taken in last year. And this is the first image of the telescope release. You can see these beautiful galaxies, clusters, and each of the light you see is a galaxy. Each of the galaxies have 1 billion stars there at different distance. One way that Rubin is observing is observing, we call it, in time domain, to compare-- the difference of image. As you can see here, there are moving stars go across it, We call it asteroid. So one key thing in astronomy is to identify things move in the sky, because most of the things, they do not move in the sky. And that's how Roman is identifying asteroids within the first data set. And Rubin is actually producing roughly one million alerts, these kind of events every night. And when Rubin is run at full speed, it will go to produce seven million alerts every night. thousand astronomers. So if all of us should work on the data, every night we have to reply 200 emails about the alerts. So it's definitely not scalable. And I'm showing here how the data rate of different telescopes from major ground and space telescope as a function of time. Over 2025, there is a major data rate flood because of Rome and because of Rubin. And if the square kilometer away, which is a radio telescope launch, on the ground around 2030. This telescope will produce roughly two of data [15:00] PB of data every day. And basically, no astronomer will go to analyze all of the data. So what I will say is our field is ongoing, a major data flood. We have to rely on AI because we are running out of our attention. And I want to give you a few cases of what we are doing these days by using AI in our field. So the first thing is a great learning that you see the first image of the JLST. And this galaxy cluster is bending the light from the background galaxy into this beautiful arc like shown here. Maybe 10 years ago, one PhD student can identify one or two systems like this in the data, and then publish a paper. You publish two papers, you get the PhD degree. But these days, because of such a big data rate, you have to rely on machine learning method to go through all the data that people have been taking and publish a 400-ish, sorry, together. To identify these kind of things, you can measure the distance of galaxy, and that is very powerful for understanding how the universe is expanding as a function of time On my side I also a heavy user of Codex codex cloud these days So actually this is one of my research Actually I can use the data from GWC in an under observing mode and feed all my data into all the data to AI. And AI tell me where are the spectrum, the dispersed light, land on the detector, and help me to get a new observing mode, a new way to use the James Webb Space Telescope to observe the universe. and people have been using the AI to run the telescope. Basically, these jobs were done by human beings over the past 400 years. You have to pick a target and point your telescope to the target, and then if there is a cloud going on, you have to point your telescope to other places, or close your dome, or if you find something that is going off, like a supernova, you point the telescope to the exact place. All of these decisions change was by human beings, but these days you can do it with AI. And this is from my friends at Tsinghua. They develop a self-supervised model that actually, when we look at the stare when we point our telescope in a certain piece of sky we stack the data together So you have a very deep 2D image of it But if you look at the temporal, the time domain of the noise, actually, you can understand how noise is fluctuating. So you are getting deeper data than what you are getting from the traditional method. And these people, using this method, they can get the data deeper by roughly a factor of 2.5. So that means you are using GWC not only like a 6.5 meter telescope, but something like 20 meter telescope, that there is no way to launch that into space right now. And our field is being heavily impacted by a large language model over the last two years also. And there is a recent paper saying that more than half of the recent astronomy paper are written by large language models. And one interesting quote from the abstract is more than half of the recent paper are therefore show a lot of language model choice, but only 0.8% of them declare that in the paper. So I'm not saying that the astronomers are cheating to each other these days, but actually the field is changing so fast that we don know what is the best way to proceed And another interesting thing is I talking about the data but talking about the place that it generated which is telescope These days astronomers are building start to build telescopes like AI Star, which is very interesting, I want to share it with you. So when astronomers are computing, you know that the computer evolved very fast over the last 60 years. But when you talk about observing, let's say Hubble Space Telescope, this is telescope design in 1970s, launched in the space in 1990s. We are still writing the telescope proposal to apply for the time and use it right now. And it won't retire until 2030, at least. And NASA is still thinking about a mission to save Hubble, to make it run longer. And right now, the private fund is kicking in for building a telescope. So this is a project funded by Eric Schmidt, the former Google CEO. It's called Lazuli. This is the first private space observatory that will be launched to the space. They collaborated with people like a promoter who is a Nobel Prize laureate back in 2011, and also people at the University of Arizona to build all of the telescope and going to launch it in two or three years. Nearly all of the things on the telescope, they can buy it as a commercial product, and all of the product team signed the NDA before releasing the telescope idea to the people. [20:00] a little bit about this telescope. It's probably some closer telescope like OpenAI or Cloudy. But we believe this telescope will bring huge impact to our research in astrophysics. And another way to look at this is this mission concept called the Nautilus Space Observatory. which is by a professor at the University of Arizona. Think about the star shape. If you fold 10 mirror, each of the size is a 6.5 meter, to the starship. And when it gets launched, each of the mirrors get expanded when they go to the space. And basically, you will launch 10 telescopes like James Webb Space Telescope at the same time. And when you build 10 telescopes, you can do faster iteration. Okay, these years, people start to build telescopes like 50 centimeters, and they can do this kind of experiment. the price will be much cheaper than building-- the average price will be much cheaper than building just one, as you can imagine. Thank you. So this is my final slide. I'm very happy to report some of the recent progress that we made in the field of astronomy. And I would like to align with the AI researchers and leaders in this room. First of all, as an astronomer, I believe that we definitely need AI to prepare for the ongoing data flood. And this is already at a scale that no of the astronomical team can handle. And we definitely need that for the discovery and also the breakthrough, make a decision. And the other interesting thing with our data is our data comes with no label at all. algorithm. And astronomers release every photon that we collect to the public without any barriers. And the universe is giving us this gift. And another thing, very importantly, is if there's only one open science among the whole discipline, then that will be astronomy. Because astronomy is looking outward, not on anything on the ground in this planet. And if you believe that model collapse when trained on recrously regenerated data, then I will tell you that the astronomers are collecting fresh data every day and from nowhere from the planet. And astronomy is-- modern astronomy is a field that I already get 400 years of history. We collect a lot of data and start to identify things. Right now, AI is helping us to see more than what we can imagine. And that's all I want to share. Thank you very much.