利用 Claude Science 绘制首张完整的紫外光波段全天星图

The missing map of the sky

Oct 8, 2026

约翰斯·霍普金斯大学天体物理学家、Anthropic 研究员布里斯·梅纳尔在此介绍了他如何与 Claude Science 合作，绘制出首张完整的紫外光天空地图。这张地图将成为宝贵的教育工具，让学生能够详细观察以这一波长呈现的银河系丰富结构。Here, Brice Ménard, an astrophysicist at Johns Hopkins University and a researcher at Anthropic, explains how he worked with Claude Science to produce the first complete map of the sky in UV light. The map will be a valuable educational tool, allowing students to see in detail the rich structure of the Milky Way visible at this wavelength.

整个天空的紫外线图像，银河系中心位于中央（合成远紫外线 154 纳米和近紫外线 232 纳米数据）。这幅图约有三分之一的区域（包括银河平面的大片区域）是使用 Claude Science，按照本文介绍的方法预测的。地图的其他图层会将每个像素标记为“实测”或“预测”，并提供不确定性估计；你可以在这里查看。The entire sky, with the galactic center in the middle, in the ultraviolet (combining far-UV 154 nm and near-UV 232 nm). About a third of this map, including much of the galactic plane, was predicted with Claude Science using the method outlined in this post. Additional layers of the map label each pixel as “measured” or “predicted” and provide uncertainty estimates; you can see it here.

我讲授天体物理学时，常常会先让学生看到：观测波长不同，天空呈现出的样子也截然不同。每种波长都会凸显出不同的特征。在可见光下，人们主要看到的是恒星。切换到红外波段，就能看到微小尘埃颗粒发出的微光，它们沿着银河系聚集成带状。到了无线电波段，恒星和尘埃逐渐隐去，氢气——新恒星形成的原材料——则显现出来。而在 X 射线波段，人们可以看到剧烈的天文现象，比如恒星爆炸，或物质盘旋着坠入黑洞。这些特征共同帮助天文学家了解气体如何坍缩形成恒星、恒星如何为星系播撒尘埃，以及像我们这样的星系如何形成。When I teach astrophysics, I often begin by showing my students that the sky looks completely different depending on the wavelength you observe it in. Each brings out different features. In visible light, one mainly sees stars. Switching to infrared reveals the glow of tiny grains of dust concentrated in a band along the Milky Way. In radio waves, the stars and dust fade and hydrogen gas—the raw material from which new stars form—becomes visible. And in X-rays, one can see violent phenomena, like stars exploding, or matter spiraling into a black hole. Together, these features let astronomers understand how gas collapses into stars, how stars seed the galaxy with dust, and how galaxies like ours take shape.

每当谈到紫外线，以及它如何显现被星光照亮的尘埃——从年轻恒星周围的云团，到恒星爆炸留下的环状结构——我总是不得不先道歉。我唯一能展示的紫外线图像上到处都是空洞。When talking about UV light and how it reveals dust illuminated by starlight, from the clouds around young stars to the rings left by stellar explosions, I have always had to apologize. The only UV map I could show was full of holes.

为什么不存在完整的 UV 贴图Why no complete UV map existed

大气中的臭氧层会吸收紫外线，因此要观测紫外波段的宇宙，就必须进入太空。在过去的 50 年里，太空望远镜对部分天空进行了紫外观测，并公布了相关数据。但仍有许多区域尚未被观测。UV light is absorbed by the ozone layer of the atmosphere, so to observe the universe in the UV spectrum, one has to go to space. Over the past 50 years, space telescopes have observed parts of the sky in UV and published their data. But much remains unobserved.

最大的数据集来自 NASA 的 GALEX 任务，该任务于 2003 年至 2013 年运行，通过约 38,000 次独立观测拍摄了大约三分之二的天空。不过，GALEX 有意避开了存在极亮恒星的区域，包括银河系盘面（星体最为密集的星系扁平圆盘），因为这些恒星可能会损坏卫星的探测器。NASA 的 Swift 和韩国的 FIMS/SPEAR 等其他空间望远镜也采集了额外数据，但即使将这些数据合并起来，仍有不少空白。The largest dataset comes from NASA’s GALEX mission, which ran from 2003 to 2013 and imaged about two-thirds of the sky in some 38,000 separate observations. But GALEX deliberately skipped locations with very bright stars, including those on the plane of the Milky Way (the flat disk of the galaxy, where stars are most densely packed), because of the risk of damaging the satellite’s detectors. Other space telescopes, such as NASA’s Swift and South Korea’s FIMS/SPEAR, acquired additional data, but even in those combined datasets, there are a number of missing pieces.

有一些统计技术可以用来估算缺失数据，从而生成完整的天空紫外线图。但要妥善完成这项工作，需要数周时间，进行细致繁琐的操作，包括精确到单个像素的校准，以及复杂而反复的分析。可以理解的是，天体物理学家往往会把这类工作搁置一旁，优先推进更紧迫的研究。大多数领域都有一批类似的积压项目：它们本可以阐明某个关键概念，或帮助其他研究人员，却始终排不上任何人的待办清单，因而迟迟无法完成。有了 Claude，处理这类优先级较低的工作变得容易了。Statistical techniques exist to estimate the missing data and create a complete UV map of the sky. But doing this properly takes weeks of painstaking work, involving meticulous calibrations at the level of individual pixels and sophisticated, repeated analyses. Understandably, astrophysicists tend to put this kind of work aside in pursuit of more pressing research. Most fields have a backlog of similar projects, ones that would explain a key concept or help other researchers, but which never rise high enough on anyone’s list to get done. With Claude, it has become easier to tackle such lower-priority work.

构建缺失的地图Building the missing map

今年夏天，我着手用 Claude Science 绘制一幅完整的全天紫外线地图。这个任务说起来简单，做起来却不容易：收集所有可用的紫外线数据集，将它们统一到同一尺度，合并成一张地图，再补上紫外线望远镜从未观测过的每一片天空。This summer, I set out to make a complete UV map of the sky with Claude Science. The instructions were simple to state, but not to carry out: gather every available UV dataset, put them all on a common scale, merge them into a single map, and fill in every patch of sky that no UV telescope has ever observed.

为此，Claude 协调了一支 AI 智能体团队。首先，这些智能体在网上搜寻公开可用的天文紫外线巡天数据集；每个数据集都包含海量的独立数据点或图像——以 NASA 的 GALEX 任务为例，图像数量达数万张。这些数据是在数年间、不同条件下采集的。To do this, Claude orchestrated a team of AI agents. First, the agents searched the web for publicly available astronomical UV surveys, each comprising a massive collection of individual data points or images—tens of thousands of images, in the case of NASA’s GALEX mission—taken over several years and under varying conditions.

下载数据后，智能体必须使每次巡测的数据在内部保持一致，以便将一个时期拍摄的图像与另一个时期拍摄的图像进行比较。这需要特别关注亮星周围的区域。正如太阳或路灯会给地球上的照片增添眩光，恒星也会让太空望远镜拍摄的天空图像产生眩光。在精确测量恒星周围较暗的紫外光之前，必须先去除这些眩光。许多智能体并行处理天空中的不同区域，从而消除了这些差异。After downloading the data, the agents had to make each survey internally consistent, so that an image taken in one time period could be compared with an image taken in another. This required paying special attention to the regions around bright stars. Just as the sun or a streetlight can add glare to a photo here on Earth, stars can add glare to an image of the sky captured by a space telescope. That glare needed to be removed before the fainter UV light around the stars could be measured precisely. Many agents, working in parallel on different regions of the sky, were able to even out these differences.

接下来，必须将这些巡天数据合并起来。它们是由不同望远镜上的不同仪器采集的，每套数据呈现的紫外天空都略有不同，因此在合并成一张地图之前，必须先对它们进行相互交叉校准，以相同的分辨率重新绘制，并映射到统一的坐标系中。我向 Claude 提供了如何完成这些工作的高层次指示，随后 Claude 派出一组代理着手处理这些任务。Next, the surveys had to be combined. The datasets were collected by different instruments on different telescopes, and each one captures the ultraviolet sky slightly differently, so they had to be cross-calibrated against one another, redrawn at the same resolution, and mapped onto a common coordinate system before they could be merged into one map. I gave high-level instructions to Claude on how to do so, and Claude set a team of agents on the tasks.

接下来是最难的一步：填补地图上的空白。大约三分之一的天空从未经过紫外线观测。为了填补这些空白区域，我请 Claude Science 使用一种名为图像修复的技术——这是机器学习中一种常见的方法：模型通过数百万张照片进行训练，学会图像各部分与周围区域之间的关系，并利用这些知识修复图像中受损或缺失的部分。Then came the hardest part: filling in the gaps in the map. Roughly a third of the sky has never been observed in UV. To account for those blank spots, I asked Claude Science to use a technique called inpainting—a common process in machine learning, in which a model trained on millions of photos learns how each part of an image relates to its surroundings, and uses that knowledge to restore damaged or missing portions of an image.

Claude 将这种方法与我们掌握的其他信息结合起来。在缺少紫外线数据的地方，我们确实有其他波长的观测数据，包括可见光、红外线和射电波段。Claude 利用已经绘制出紫外线地图的那三分之二天空，学习紫外线亮度与这些其他波长之间的关系。随后，它将这种关系应用于缺少紫外线数据的那三分之一天空，估算每个位置的紫外线图像应是什么样，以及它对这项估算有多大把握。Claude combined this approach with additional information we had at our disposal. Where UV data is missing, we do have observations at other wavelengths, including visible, infrared, and radio. Using the two-thirds of the sky that has been mapped in UV, Claude learned how UV brightness relates to these other wavelengths. It then applied that relationship to the third of the sky for which no UV data exists, and estimated what it should look like at each point, and how confident it was in that estimate.

为了检验模型预测的准确性，我让 Claude 选取已有 UV 数据的区域，并故意遮住其中一部分，制造出图像修复技术旨在填补的空白区域。模型必须利用学到的知识，在无法获取真实数值的情况下补全图像。经过几轮优化后，它估算出的隐藏数据与真实 UV 测量值的差距约为 10%，这种差异几乎难以用肉眼察觉。To check how accurate the model’s predictions were, I asked Claude to take regions for which we already have UV data and deliberately hide parts of them, creating the kinds of blank patches inpainting is designed to fill. The model had to use what it had learned to complete the picture without access to the real values. After several rounds of refinement, it was able to estimate the hidden data to within about 10% of the real UV measurements, a difference almost imperceptible to the human eye.

最后，在这个经过弥散图像修复的背景之上，Claude 根据欧洲航天局盖亚卫星在可见光波段的测量数据，加入了对超过一亿颗恒星紫外线辐射的估算。首个完整的全天空紫外线地图版本由此完成。Finally, on top of this diffuse inpainted background, Claude added estimates of UV light from more than 100 million individual stars inferred from measurements by the European Space Agency’s Gaia satellite in visible light. The first version of the full UV map of the sky was complete.

反复试验Trial and error

Claude 并非第一次尝试就把一切都做对了。比如，有天晚上我查看处理后的图像时，注意到最暗的视野之一有些不对劲：一些微弱的圆圈，各自比周围的圆圈稍亮或稍暗。Claude did not get everything right on the first try. Looking through the processed images one evening, for example, I noticed something off in one of the dimmest fields: faint circles, each a touch brighter or darker than its neighbors.

这些是各次 GALEX 观测的覆盖范围。每幅 GALEX 图像都拍摄了天空中的一片圆形区域，其中都带有一丝来自地球大气层、并非完全均匀的紫外辉光。如果这种辉光没有被彻底去除，每个圆形区域看起来就会比周围区域略亮或略暗。Claude 在项目开始时就将此列为一个已知问题，但这张地图仍通过了其他智能体的两轮审查，都没有人发现这个问题。These were the footprints of individual GALEX observations. Each GALEX image captures a circular patch of the sky, and each contains a slight, not-quite-uniform UV glow from the Earth’s atmosphere. If that glow is not fully removed, each circle appears slightly brighter or darker than those around it. Claude had listed this as a known issue at the start of the project, but the map had still passed two rounds of review by other agents without the problem being caught.

我告诉 Claude：“我能看到圆盘上有单次观测留下的印记；你能修正一下吗？”智能体追查后发现，问题源自残留的大气辉光。随后，Claude 对全部 38,000 次观测中的这种辉光进行了修正。处理了几个小时后，那些圆圈消失了。I told Claude, “I can see discs with the imprint of individual observations; can you correct that?” The agents traced the problem back to the leftover atmospheric glow. Claude then corrected for that glow in all 38,000 observations. After a couple of hours of processing, the circles were gone.

最终地图The final map

我和 Claude 一起创建了十多个连续版本的地图，每个版本都比上一个更加完善。我们的合作持续了好几天，逐渐形成了一种可预见的节奏：我会在 Claude Science 中与它交流几轮，规划接下来的步骤；随后 Claude 会独立进行数小时的计算，而我则回去继续处理其他项目。Together, Claude and I created more than a dozen successive versions of the map, each more refined than the last. Our collaboration spread over several days and took on a predictable rhythm: I would have several exchanges within Claude Science to plan the next steps, then Claude would independently perform hours of computations, while I went back to working on other projects.

完成后的地图以紫外光呈现了整片天空，并展示了许多有趣的结构：明亮年轻恒星周围发光的尘埃云；巨大的尘埃环和弧圈，勾勒出由大质量恒星爆炸吹出的气泡状气体云。远离银河平面的地方，微弱的尘埃丝状结构横贯天空，映照着整座星系中所有恒星汇聚而成的星光。银河平面下方是大麦哲伦云和小麦哲伦云这两座邻近的小星系，年轻恒星使它们熠熠生辉。我觉得这幅图很美——它也可以与其他经典地图一道，帮助一代代学生领略我们银河系中各种物理过程的多样性。The finished map shows the entire sky in UV light, and displays many interesting structures: dust clouds glowing around luminous young stars, giant dust rings and loops tracing gas bubbles blown out by the explosions of massive stars. Far from the galactic plane, faint filaments of dust thread across the sky, lit by the combined starlight of the entire galaxy. Below the plane sit the Large and Small Magellanic Clouds, two small neighboring galaxies bright with young stars. I think it’s beautiful—and it can join the other canonical maps in helping generations of students appreciate the variety of physical processes at work in our galaxy.

这一次，我终于能够着手开展这样的项目，而不必牺牲原本可以用于研究的时间。我想，许多科学家都能想到一张地图、一幅图表或一项资源，因为同样的原因，他们一直迟迟没有着手去做。For once, I was able to pursue a project like this without sacrificing time I would otherwise have devoted to research. I suspect many scientists can think of a map, a figure, or a resource they’ve been putting off for the same reason.

致谢Acknowledgments

用于制作这张地图的测量数据来自 GALEX 和 Swift（NASA）、FIMS/SPEAR（韩国）、TD-1（欧洲），以及 Planck 和 Gaia（ESA）。我感谢参与这些任务并将数据公开的每一个人。Claude Science 中的许多智能体为绘制这张地图付出了耐心细致的劳动。我的贡献仅限于在这一过程中为他们提供指导。The measurements used to create this map are from GALEX and Swift (NASA), FIMS/SPEAR (Korea), TD-1 (Europe), and Planck and Gaia (ESA). I’m grateful to everyone involved in these missions and in making the data public. Many agents working in Claude Science contributed the patient labor of putting this map together. My contribution was limited to guiding them in the process.