内容来源：https://www.wired.com/story/ai-comes-up-with-bizarre-physics-experiments-but-they-work/

内容总结：

【前沿科技】AI助力物理学突破：从引力波探测到量子纠缠实验设计

LIGO灵敏度有望提升15%：AI设计出颠覆性干涉仪结构
美国激光干涉引力波天文台（LIGO）作为人类最精密的科学仪器之一，其4公里长的干涉臂可检测到相当于质子宽度万分之一的时空波动。加州理工学院物理学家拉纳·阿迪卡里团队利用AI软件重新优化LIGO设计时，AI提出在干涉臂末端添加3公里环形光路的反直觉方案。经数月验证，该设计可能基于苏联物理学家数十年前提出的量子降噪理论，有望将LIGO灵敏度提升10%-15%。多伦多大学量子光学专家埃弗拉姆·斯坦伯格评价："这是数千名科学家40年都未想到的突破。"

量子实验设计革命：AI重构纠缠交换方案
2021年，物理学家马里奥·克伦团队开发名为"PyTheus"的AI系统，通过图论建模量子光学实验。在重构1993年诺贝尔奖得主蔡林格的纠缠交换实验时，AI竟给出完全不同的设计方案——利用多光子干涉原理简化了实验结构。2024年12月，中国南京大学马小松团队成功验证该设计，实现无历史关联粒子的量子纠缠。克伦坦言："若学生提出这个方案，我肯定会认为荒谬。"

AI揭示宇宙暗物质分布新规律
威斯康星大学物理学家凯尔·克兰默团队通过机器学习分析暗物质团块数据，AI自主推导出比人工模型更精确的密度计算公式。加州大学圣地亚哥分校计算机科学家罗斯·余团队则让AI从大型强子对撞机数据中重新发现了爱因斯坦相对论的核心——洛伦兹对称性。研究人员指出，当前AI虽能识别复杂模式，但解释物理机制仍需人类智慧。不过随着大语言模型发展，克兰默预测："AI辅助构建物理假说的时代即将到来。"

（本文基于《量子杂志》授权报道，该刊由西蒙斯基金会资助，致力于推动公众对数学及自然科学的理解）

中文翻译：

本文原载于《量子杂志》（Quanta Magazine）。
在精密测量领域，激光干涉引力波天文台（LIGO）堪称巅峰之作。这座由双探测器组成的观测系统（分别位于华盛顿州汉福德和路易斯安那州利文斯顿）采用巨型L形结构，每条四公里长的臂腔内激光束不断反射。当引力波经过时，两条臂长的变化差异甚至小于一个质子的宽度——正是通过捕捉这种微小变化（其灵敏度相当于测量4.37光年外的半人马座α星时误差不超过一根头发丝的直径），科学家们才能做出重大发现。

这台机器的设计凝聚了数十年的心血，物理学家们必须将每个环节都推向物理极限。1994年启动建造后历时二十余载，期间还经历了为期四年的升级改造，最终LIGO在2015年首次探测到引力波：那对遥远黑洞碰撞产生的时空涟漪。

加州理工学院物理学家拉纳·阿迪卡里曾在2000年代中期领导探测器优化团队。他与少数合作者精心打磨LIGO设计，探索所有制约探测器灵敏度的边界。但2015年成功探测后，阿迪卡里开始思考如何突破现有设计——例如拓展可探测的引力波频段。这种改进将使LIGO能观测不同质量黑洞的合并，甚至可能发现意外现象。"我们真正渴望发现的是前所未见的宇宙奇观，"阿迪卡里说，"对宇宙的创造物不该抱有任何先入之见。"

研究团队转向人工智能，特别是物理学家马里奥·克伦最初为量子光学桌面实验开发的软件套件。他们先向AI提供所有可组合的光学元件，允许其设计任意复杂的干涉仪。初期AI不受约束，设计出的探测器跨度数百公里，包含数千个透镜、反射镜和激光器等元件。

"AI给出的设计完全超出人类理解范畴，"阿迪卡里回忆道，"它们过于复杂，像是外星科技或纯AI产物，毫无人类崇尚的对称美感和简洁性，简直一团乱麻。"研究人员经过整理才获得可解读的方案，但仍对设计原理困惑不已。"如果学生交来这种方案，我肯定认为荒谬绝伦。"但事实证明这个设计确实有效。

团队耗费数月才破解AI的思维。原来AI运用了反直觉的技巧：在主干涉仪与探测器之间增设三公里环形光路，使激光在离开干涉臂前循环传播。阿迪卡里团队意识到，AI可能运用了苏联物理学家几十年前提出的深奥理论来降低量子噪声，这些设想从未被实验验证。"要跳出传统思维框架需要极大勇气，AI确实不可或缺。"

阿迪卡里坦言，若当年建造LIGO时获得AI的洞见，"灵敏度本可提升10%到15%"。在亚质子级精度的世界里，这种提升堪称飞跃。多伦多大学量子光学专家埃弗拉姆·斯坦伯格指出："LIGO凝聚了数千人四十年的深思熟虑，AI能提出新方案，本身就证明它突破了人类思维的局限。"

虽然AI尚未直接催生物理学新发现，但已成为该领域的强力工具。除实验设计外，AI还能从复杂数据中发掘非平凡模式。例如算法已从瑞士大型强子对撞机数据中识别出自然界对称性（这些爱因斯坦相对论的关键要素虽非新发现，却验证了AI的潜力）；物理学家还借助AI推导出描述暗物质分布的新方程。"人类正开始向AI方案学习。"阿迪卡里表示。

量子纠缠新篇
经典物理中，台球等物体具有独立于测量的明确属性（如确定的位置和动量）。但量子世界截然不同——量子态只能给出概率性描述。更奇妙的是，多个量子物体可共享同一量子态。以光子为例，纠缠光子对即便分离仍共享量子态，测量其中一个会瞬时决定遥远另一个的状态。

数十年来物理学家认为纠缠必须始于同处。但1990年代初，后来因纠缠研究获诺奖的安东·蔡林格打破这一认知。他和同事设计实验：先制备两对无关联的纠缠光子（AB纠缠，CD纠缠），然后对B、C光子进行晶体-分束器-探测器组合操作。当B、C被测量湮灭后，原本无交集的A、D竟形成纠缠。这种"纠缠交换"现已成为量子技术基石。

2021年克伦团队用自研软件PyTheus（Python编程语言+斩杀米诺陶的忒修斯）设计新实验。他们将光学实验抽象为图结构（节点代表光学元件，边代表光子路径或相互作用），通过构建通用图模型寻找产生目标量子态（如无历史交互的纠缠粒子对）的实验配置。

当学生索伦·阿尔特用该方法优化纠缠交换方案时，得到的配置与1993年蔡林格设计毫无相似之处。"我们最初以为算法出错了，"克伦说。但分析显示，算法借鉴了多光子干涉理论，创造了更简洁的方案。2024年12月，南京大学马小松团队成功验证了这个AI设计的实验。

破解隐藏公式
物理学家还利用AI解析实验结果。威斯康星大学物理学家凯尔·克兰默形容："现阶段就像教孩子说话，需要大量人工引导。"但经真实与模拟数据训练的模型已能发现被忽视的模式。例如克兰默团队用机器学习预测暗物质团密度，AI推导的公式比人工模型更契合数据。"但它无法解释推导过程。"

加州大学圣地亚哥分校计算机科学家罗斯·俞团队训练模型从大型强子对撞机数据中发现洛伦兹对称性（爱因斯坦相对论核心概念）。"模型在不懂物理的情况下，纯靠数据重现了这一对称性。"俞解释道。

研究者指出，当前AI虽擅长发现模式，但解释机制仍具挑战。不过克兰默认为ChatGPT等大语言模型可能改变现状："它们在自动化假说构建方面潜力巨大，突破近在眼前。"斯坦伯格也认为："我们可能正跨向AI辅助发现新物理的激动时刻。"

本文经《量子杂志》授权转载，该刊由西蒙斯基金会运营，致力于通过报道数学、物理与生命科学进展提升公众科学认知。

英文来源：

The original version of this story appeared in Quanta Magazine.
There are precision measurements, and then there’s the Laser Interferometer Gravitational-Wave Observatory. In each of LIGO’s twin gravitational wave detectors (one in Hanford, Washington, and the other in Livingston, Louisiana), laser beams bounce back and forth down the four-kilometer arms of a giant L. When a gravitational wave passes through, the length of one arm changes relative to the other by less than the width of a proton. It’s by measuring these minuscule differences—a sensitivity akin to sensing the distance to the star Alpha Centauri down to the width of a human hair—that discoveries are made.
The design of the machine was decades in the making, as physicists needed to push every aspect to its absolute physical limits. Construction began in 1994 and took more than 20 years, including a four-year shutdown to improve the detectors, before LIGO detected its first gravitational wave in 2015: a ripple in the space-time fabric coming from the faraway collision of a pair of black holes.
Rana Adhikari, a physicist at the California Institute of Technology, led the detector optimization team in the mid-2000s. He and a handful of collaborators painstakingly honed parts of the LIGO design, exploring the contours of every limit that stood in the way of a more sensitive machine.
But after the 2015 detection, Adhikari wanted to see if they could improve upon LIGO’s design, enabling it, for instance, to pick up gravitational waves in a broader band of frequencies. Such an improvement would enable LIGO to see merging black holes of different sizes, as well as potential surprises. “What we’d really like to discover is the wild new astrophysical thing no one has imagined,” Adhikari said. “We should have no prejudice about what the universe makes.”
He and his team turned to AI—in particular, a software suite first created by the physicist Mario Krenn to design tabletop experiments in quantum optics. First, they gave the AI all the components and devices that could be mixed and matched to construct an arbitrarily complicated interferometer. The AI started off unconstrained. It could design a detector that spanned hundreds of kilometers and had thousands of elements, such as lenses, mirrors, and lasers.
Initially, the AI’s designs seemed outlandish. “The outputs that the thing was giving us were really not comprehensible by people,” Adhikari said. “They were too complicated, and they looked like alien things or AI things. Just nothing that a human being would make, because it had no sense of symmetry, beauty, anything. It was just a mess.”
The researchers figured out how to clean up the AI’s outputs to produce interpretable ideas. Even so, the researchers were befuddled by the AI’s design. “If my students had tried to give me this thing, I would have said, ‘No, no, that’s ridiculous,’” Adhikari said. But the design was clearly effective.
It took months of effort to understand what the AI was doing. It turned out that the machine had used a counterintuitive trick to achieve its goals. It added an additional three-kilometer-long ring between the main interferometer and the detector to circulate the light before it exited the interferometer’s arms. Adhikari’s team realized that the AI was probably using some esoteric theoretical principles that Russian physicists had identified decades ago to reduce quantum mechanical noise. No one had ever pursued those ideas experimentally. “It takes a lot to think this far outside of the accepted solution,” Adhikari said. “We really needed the AI.”
If the AI’s insights had been available when LIGO was being built, “we would have had something like 10 or 15 percent better LIGO sensitivity all along,” he said. In a world of sub-proton precision, 10 to 15 percent is enormous.
“LIGO is this huge thing that thousands of people have been thinking about deeply for 40 years,” said Aephraim Steinberg, an expert on quantum optics at the University of Toronto. “They’ve thought of everything they could have, and anything new [the AI] comes up with is a demonstration that it’s something thousands of people failed to do.”
Although AI has not yet led to new discoveries in physics, it’s becoming a powerful tool across the field. Along with helping researchers to design experiments, it can find nontrivial patterns in complex data. For example, AI algorithms have gleaned symmetries of nature from the data collected at the Large Hadron Collider in Switzerland. These symmetries aren’t new—they were key to Einstein’s theories of relativity—but the AI’s finding serves as a proof of principle for what’s to come. Physicists have also used AI to find a new equation for describing the clumping of the universe’s unseen dark matter. “Humans can start learning from these solutions,” Adhikari said.
Apart but Together
In the classical physics that describes our everyday world, objects have well-defined properties that are independent of attempts to measure those properties: A billiard ball, for example, has a particular position and momentum at any given moment in time.
In the quantum world, this isn’t the case. A quantum object is described by a mathematical entity called the quantum state. The best one can do is to use the state to calculate the probability that the object will be, say, at a certain location when you look for it there.
What is more, two (or more) quantum objects can share a single quantum state. Take light, which is made of photons. These photons can be generated in pairs that are “entangled,” meaning that the two photons share a single, joint quantum state even if they fly apart. Once one of the two photons is measured, the outcome seems to instantaneously determine the properties of the other—now distant—photon.
For decades, physicists assumed that entanglement required quantum objects to start out in the same place. But in the early 1990s, Anton Zeilinger, who would later receive the Nobel Prize in Physics for his studies of entanglement, showed that this wasn’t always true. He and his colleagues proposed an experiment that began with two unrelated pairs of entangled photons. Photons A and B were entangled with each other, as were photons C and D. The researchers then devised a clever experimental design made of crystals, beam splitters and detectors that would operate on photons B and C—one photon from each of the two entangled pairs. Through a sequence of operations, the photons B and C get detected and destroyed, but as a product, the partner particles A and D, which had not previously interacted, become entangled. This is called entanglement swapping, which is now an important building block of quantum technology
That was the state of affairs in 2021, when Krenn’s team started designing new experiments with the aid of software they dubbed PyTheus—Py for the programming language Python, and Theus for Theseus, after the Greek hero who killed the mythical Minotaur. The team represented optical experiments using mathematical structures called graphs, which are composed of nodes connected by lines called edges. The nodes and edges represented different aspects of an experiment, such as beam splitters, the paths of photons, or whether or not two photons had interacted.
Krenn’s team started by first building a very general graph, one that modeled the space of all possible experiments of some size. The graph had output features that represented some desired quantum state—say, two particles exiting the experimental setup that had never interacted but were now entangled.
The question, then, was how to modify all the other parts of the graph to produce this state. To figure this out, the researchers formulated a mathematical function. It took in the state of the graph and calculated the difference between the output of the graph and the desired quantum state. They then iteratively modified the graph’s parameters, which represented the experimental configuration, to reduce this discrepancy to zero.
When Krenn’s student Soren Arlt tried to use this approach to find the best way to do entanglement swapping, he noticed that the experimental configuration was unrecognizable—nothing at all like Zeilinger’s design from 1993. “When he showed it to me, we were confused,” Krenn said. “I was convinced that it must be wrong.”
The optimization algorithm had borrowed ideas from a separate area of study called multiphoton interference. By doing so, it created a simpler configuration than Zeilinger’s. Krenn’s team then did a separate mathematical analysis of the final design. It confirmed that the new experimental design would in fact create entanglement among particles with no shared past.
In December 2024, a team in China led by Xiao-Song Ma of Nanjing University confirmed it. They built the actual experiment, and it worked as intended.
Finding the Hidden Formula
Experimental design isn’t the only way that physicists are using AI. They’ve also put it to work parsing experimental results.
“Right now, I’d say it’s like teaching a child how to speak,” Kyle Cranmer, a physicist at the University of Wisconsin-Madison, said of the budding efforts to use AI to do physics. “We’re doing a lot of baby-sitting.” Even so, machine learning models trained on real-world and simulated data are discovering patterns that might otherwise have been missed.
For example, Cranmer and his collaborators used a machine learning model to predict the density of clumps of dark matter in the universe, based on observable properties of other such nearby clumps. Such calculations are necessary to understand the growth of galaxies and galaxy clusters. The system arrived at a formula to describe the density of dark matter clumps that better fit the data than a human-made one. The AI’s equation “describes the data very well,” Cranmer said. “But it’s lacking the story about how you get there.”
Sometimes it’s enough of a proof of principle to show that AI can rediscover things that people already know.
Rose Yu, a computer scientist at the University of California, San Diego, and colleagues have been training machine learning models to find symmetries in data. A symmetry implies that the data either remains unchanged or changes predictably and simply under a transformation. For example, a circle has rotational symmetry—it is invariant under rotation. Yu and her team applied their technique to data collected at the Large Hadron Collider and identified so-called Lorentz symmetries, which are crucial to Einstein’s theories of relativity. These are changes in perspective that leave the applicable laws of physics unchanged. For example, the rate of production of pairs of particles at the collider should not change at different times of day. If the rate varied, it would imply some dependence on Earth’s rotation and hence a preferential direction in space-time. “We showed that, without knowing any physics, the model can discover the Lorentz symmetry purely from data,” Yu said.
Cranmer and Yu point out that while such methods are good at discovering patterns, making sense of those patterns and coming up with hypotheses or the physics to explain them remains elusive for today’s AI models. But Cranmer thinks that the advent of large language models like ChatGPT could change that. “I think there’s a huge potential for language models to be useful to help automate that construction of hypotheses,” he said. “It’s kind of around the corner.”
Steinberg agrees that while AI has yet to invent new concepts, AI-aided discoveries of new physics could conceivably become reality. “We really might be crossing that threshold, which is exciting,” he said.
Original story reprinted with permission from Quanta Magazine, an editorially independent publication of the Simons Foundation whose mission is to enhance public understanding of science by covering research developments and trends in mathematics and the physical and life sciences.