人工智能模型可以胜任许多行业,，包括编写,、生成图像, 和创建 3D 模型。但当涉及到在不同环境中测试机器人或车辆设计时，它们’t没有帮助,，因为它们’对物理的理解不如对像素或文本的理解。

Artificial intelligence models are jacks of many trades, including writing, generating images, and creating 3D models. But they aren’t as helpful when it comes to testing robots or designs for vehicles in diverse environments, since they don’t understand physics as well as they do pixels or text.

很快,该项目可以帮助工程师预测车辆(如汽车和飞机),日常用品(包括椅子和容器),以及机器人如何响应各种物理元素,，例如风,水,和碰撞。研究人员相信，他们的工作也可能是迈向物理基础模型,的一步，这是一个经过大量数据训练的骨干系统，可以帮助人工智能工具泛化到不同的任务。

Soon, the project could help engineers predict how vehicles (like cars and planes), everyday items (including chairs and containers), and robots respond to various physical elements, such as wind, water, and collisions. The researchers believe their work could also be a step toward a physics foundation model, a backbone system trained on lots of data that can help AI tools generalize to different tasks.

“我们相信物理学是继文本和像素之后人工智能模型的第三种模式,” 麻省理工学院博士生和 CSAIL 研究员郭明浩, 是一篇介绍 GeoPT 的论文的共同主要作者。 “我们的通用模型具有多功能性，可以帮助构建物理世界模型。许多模型,（例如生成机器人数据和视频, 的模型）已经精通文本和视觉数据,，但凭借物理准确性,，他们 将获得更真实的结果。”

“We believe physics is the third modality for AI models, after text and pixels,” says MIT PhD student and CSAIL researcher Minghao Guo, a co-lead author on a paper introducing GeoPT. “Our general-purpose model has the versatility to help build a world model for physics. Many models, such as those that generate robotics data and videos, are already well-versed in textual and visual data, but with physical accuracy, they’ll get more-realistic results.”

要使用 GeoPT,，用户只需上传战舰,、客机, 和卡车, 等物体的 3D 模型，并指定他们想要模拟的力的方向和速度。结果是一种热图，显示对象在不同位置将如何受到影响。如果您知道您’想要模拟,的力的速度和方向(velocity)，您可以在GeoPT中捕获它。当您想要模拟汽车撞墙后的样子,、光线在物体周围反射的方式, 以及船是否在湍急的波浪中保持漂浮等情况时，这会派上用场。

To use GeoPT, users simply upload 3D models of objects like battleships, passenger airplanes, and trucks, and specify the direction and speed of the force they want to simulate. The result is a kind of heat map showing how the object will be affected in different places. If you know the speed and direction (velocity) of the force you’re looking to simulate, you can capture it in GeoPT. This comes in handy when you want to simulate things like how a car would look after crashing into a wall, the ways light bounces around objects, and whether a boat stays afloat over turbulent waves.

但是 GeoPT “ 如何将” 物理学搞得这么好? 它的知识来自于 “ 合成动力学,” 小粒子和复杂 3D 形状之间的一系列相互作用。 GeoPT 研究了 130 万个合成动力学, 样本，其中微小球体以不同速度和角度移动，直到停在物体上的某个点。

But how does GeoPT “get” physics so well? Its knowledge comes from “synthetic dynamics,” a series of interactions between small particles and complex 3D shapes. GeoPT studied 1.3 million samples of synthetic dynamics, in which tiny spheres moved at various speeds and angles until stopping at a certain point on the object.

这些粒子一旦接触,，基本上就会“粘”到物体上，而不是移动或弹开。 Picture learning about physical interactions using marbles and action figures — similarly, simulation models can use synthetic dynamics to gain a feel for physics before they train on labeled data.

These particles basically “stick” to an object once they make contact, instead of moving through or bouncing off. Picture learning about physical interactions using marbles and action figures — similarly, simulation models can use synthetic dynamics to gain a feel for physics before they train on labeled data.

研究人员发现，GeoPT 尤其擅长模拟工业场景,，因为它在基准测试中的表现优于最先进的模拟模型。公共线程: 它比其他工具, 更快地达到峰值性能，同时需要的标记数据显着减少。例如，在复杂 3D 形状及其对风流和表面压力, 的响应的数据集, 上，GeoPT 在速度, 精度, 和效率方面超越了最先进的模型。在捕捉战斗机对风的反应方面，它在速度和准确性方面也取得了类似的胜利。当 GeoPT 测试船体如何处理空气和波浪, 时，捕获物理力所需的标记数据减少了 60%，并且达到峰值精度的速度比顶级基线快四倍。

The researchers found that GeoPT was particularly skilled at simulating industrial scenarios, as it outperformed state-of-the-art simulation models across benchmarks. The common thread: It reached peak performance faster than other tools, while needing significantly fewer labeled data.

On a dataset of complex 3D shapes and their responses to wind currents and surface pressure, for example, GeoPT surpassed state-of-the-art models in speed, accuracy, and efficiency. It had similar triumphs in speed and accuracy in capturing how fighter jets responded to wind. When GeoPT tested how the hull of a boat handled both air and waves, it required 60 percent fewer labeled data to capture both physical forces and reached peak accuracy four times faster than top baselines.

该系统甚至成功地模拟了不同类型的汽车与另一个物体碰撞后的情况。它正确预测了 3D 车辆将如何变形，同时使用的数据少于最先进的基线。同样,，它对光线如何穿过本质上是一只玩具兔子的模拟是准确的,，尽管事先从未接受过该 3D 模型或光物理学的训练。 “如果您的模型在工业基准上表现良好,，这意味着它可以解决最困难的物理任务,” 麻省理工学院博士后和 CSAIL 研究员 Haixu Wu, 表示。 “GeoPT 在几秒钟内进行了超过 1 亿个网格点的高保真模拟。这使得该工具对于希望测试车辆蓝图而无需进行如此多的物理实验的工程师来说非常有帮助。”

The system even succeeded at simulating how different types of cars look after colliding with another object. It correctly predicted how 3D vehicles would deform while using less data than state-of-the-art baselines. Likewise, its simulations of how light would pass through what was essentially a toy rabbit were accurate, despite never training on that 3D model or light physics beforehand.

“If your model performs well on industrial benchmarks, that means it can solve the hardest physics tasks,” says co-lead author Haixu Wu, an MIT postdoc and CSAIL researcher. “GeoPT was making high-fidelity simulations with over 100 million mesh points in seconds. This could make the tool extremely helpful for engineers hoping to test out blueprints for vehicles without needing to run so many physical experiments.”

研究人员补充说，他们的系统只是他们’一直致力于的物理世界模型的预览。该团队希望将他们的系统, 训练扩展到更多形状并模拟更复杂的物理现象。例如, 更深入的方法可以帮助模拟天气模式, 测试不同的材料, 并生成逼真的视频。

The researchers add that their system is only a preview of the kind of physics world model they’ve been working toward. The team hopes to scale up their system, training on even more shapes and simulating more complex physical phenomena. For example, a more in-depth approach could help model weather patterns, test out different materials, and generate realistic videos.

“Using synthetic dynamics data is an exciting paradigm for imbuing physics into foundation models,” says Fei Sha, AI research scientist at Meta, who wasn’t involved in the research. “It challenges the traditional wisdom that physics and geometry are necessarily entangled in computation, and one must acquire costly and specialized data.在广泛的应用领域所取得的成功使我们到达了这个重要的关头: 我们已准备好立即快速构建物理基础模型,。" Wu 和Guo 与 MIT CSAIL 同事共同撰写了这篇论文，其中包括 Zongyi Li, 实验室博士后;zhiyang(Frank) Dou, CSAIL 附属机构和麻省理工学院电气工程和计算机科学博士生(EECS); Kaiming He, a principal investigator in the lab, associate professor of EECS, and a distinguished scientist at Google DeepMind; and senior author Wojciech Matusik, the Joan and Irwin M. (1957) Jacobs Professor of EECS and a CSAIL principal investigator.清华大学副教授龙明生也是该研究的合著者。 The team presented the paper at the International Conference on Machine Learning in July. The researchers work was supported, in part, by Neural Modular Physics Twin for Robotics.

“Using synthetic dynamics data is an exciting paradigm for imbuing physics into foundation models,” says Fei Sha, AI research scientist at Meta, who wasn’t involved in the research. “It challenges the traditional wisdom that physics and geometry are necessarily entangled in computation, and one must acquire costly and specialized data. The demonstrated success in a wide range of application domains leads us to this important juncture: We are ready to build physics foundation models, now and fast."

Wu and Guo wrote the paper with MIT CSAIL colleagues including Zongyi Li, a postdoc in the lab; Zhiyang (Frank) Dou, a CSAIL affiliate and MIT PhD student in electrical engineering and computer science (EECS); Kaiming He, a principal investigator in the lab, associate professor of EECS, and a distinguished scientist at Google DeepMind; and senior author Wojciech Matusik, the Joan and Irwin M. (1957) Jacobs Professor of EECS and a CSAIL principal investigator. Tsinghua University Associate Professor Mingsheng Long was also a co-author. The team presented the paper at the International Conference on Machine Learning in July.

The researchers work was supported, in part, by Neural Modular Physics Twin for Robotics.