If you’ve been keeping track of artificial intelligence (AI) developments, you’ll know AI models are versatile tools that significantly excel in writing, creating visuals, and generating 3D models. But there’s a limit to their proficiency. Where they fall short is performing tests on robots or vehicle designs across diverse environments. Essentially, they struggle with the laws of physics more than they do with pixels or text.
So, what has been obstructing the creation of an AI system that can faultlessly depict various physical scenarios? The answer lies in the incredible scale of physics data required which is, at this point, out of reach. Giving neural networks enough data points they can comprehend is a time-consuming endeavor. These networks rely on a type of algorithm known as “numerical solvers” to pin down physical properties across different areas of a 3D shape. While this painstaking method garners detailed results, it is rather slow, curtailing the data amount available. For instance, it can limit assessing the safety or aerodynamics of airplane designs.
MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and Tsinghua University researchers have developed a pioneering pre-training method called “GeoPT.” This approach allows simulation models to learn physics more effectively by virtually reenacting daily mechanical interactions in 3D. With GeoPT, models can grasp how particles interact with objects, intensifying their ability to accurately mirror real-world scenarios. What’s more, this model can outperform leading models by training on up to 60% less data and achieving optimum performance twice as fast.
Imagine the impact of GeoPT. Engineers could predict how vehicles, household items, and robots react to various physical elements such as wind, water, and collisions. According to the researchers, this revolutionary work could lay the foundation for a comprehensive physics system that assists AI tools in extending their abilities across various tasks.
Minghao Guo, a PhD student at MIT and a CSAIL researcher, explained the significant versatility of their ground-breaking model. “We believe physics is the third modality for AI models, after text and pixels,” Guo stated. This model will add a layer of physical accuracy in addition to handling textual and visual data, creating more realistic results.
The utilization of GeoPT is user-friendly and optimized for practical application. Users can upload 3D models of various objects such as planes or trucks and specify the direction and speed of the force they wish to simulate. The outcome is a heat map showing how the object will endure varying factors—a useful tool for simulating diverse scenarios like car crashes or boat buoyancy over heated debates.
GeoPT’s strength in understanding physics comes from “synthetic dynamics,” which encompasses the interactions between small particles and complex 3D shapes. After studying 1.3 million examples of synthetic dynamics, GeoPT developed a solid grip on physics. This was achieved by visualizing how tiny spheres moved at different speeds and angles before stopping on an object. This learning method is akin to observing physical interactions by using marbles and action figures, enabling simulation models to gain a physics understanding before training on labeled data.
GeoPT shows created promise towards industrial applications- from simulating fighter jets’ responses to wind and boats’ interactions with air and waves, to even accurately replicating car collisions and the behavior of light, despite lacking prior training on specific 3D models or light physics. Its ability to produce accurate, high-fidelity simulations with over 100 million mesh points in seconds turns it into an invaluable and efficient tool for engineers.
Looking forward, the research team intends to enhance their system further by training it on a wider range of shapes and simulating complex physical phenomena, paving the way for accurate weather pattern modeling, material testing, and more realistic video generation.
Fei Sha, an AI research scientist at Meta, applauds the use of synthetic dynamics data as an innovative approach that defies the traditional view that physics and geometry need specialized and expensive data. She asserts that we are well-positioned to quickly develop physics foundation models based on the success seen over various application areas.
Feel intrigued by this research? Find out more about it from the original news article. If you’re interested in incorporating AI automation into your business, explore the possibilities with implementi.ai.
This website uses cookies.