A new Computer program at Columbia University monitored physical processes and discovered key variables—a prerequisite for any physics theory. But, the variables revealed were surprising.
Energy, Mass, and Velocity Einstein's famous equation E=MC2 is made up of these three variables. But how did Albert Einstein become aware of these ideas in the first place? Before you can understand physics, you must first identify the factors that are relevant. Even Einstein could not have discovered relativity without first understanding the ideas of energy, mass, and velocity. Can variables like these, however, be detected automatically? This would significantly enhance scientific discoveries.
This is the query posed to a new artificial intelligence software by Columbia Engineering researchers. The AI programme was created to watch physical processes using a video camera and then attempt to find the smallest set of fundamental variables that adequately describe the observed dynamics. On July 25, the work was published in the journal "Nature Computational Science".
The picture demonstrates the motion of a chaotic swing stick dynamical system. Our research attempts to find and extract the smallest amount of state variables required to characterize such a system directly from high-dimensional video data. Credit: Yinuo Qin/Columbia Engineering.
The researchers began by providing the system raw film footage of physics phenomena for something they already had an answer. They loaded a video of a swinging double-pendulum with four "state variables"—the angle and angular velocity of each of the two arms, for example. The AI produced its result after many hours of analysis: 4.7.
"We considered this was a good enough solution," says Hod Lipson, director of the Creative Machines Lab in the Department of Mechanical Engineering, where majority of the work was done. "This is especially true given that all of the AI had access to raw video material and had no knowledge of physics or geometry. But we wants to determine what the variables were, not just how many there were."
The researchers then demonstrated the actual variables discovered by the algorithm. The variables themselves were challenging to extract since the software cannot express them in any intuitive way that people would understand. Following considerable examination, it was discovered that two of the variables chosen by the software related to the orientations of the arms, but the other two remain unknown.
“We tried to match other variables with everything we could think of, including kinetic and potential energies, linear and angular velocities, and variability in measurements,” explained Dr. Boyuan Chen. '22, who headed the research. "However, nothing seemed to fit exactly." Because the AI was making reasonable predictions, the team felt convinced that it had discovered a legitimate set of four variables, but "We still don't understand the mathematical language it speaks," he said.
Boyuan Chen describes how a new AI system observed physical processes and discovered significant variables, which is an essential step before developing any physics theory. Credit: Boyuan Chen/Columbia Engineering.
Following the validation of a number of different physical systems with known solutions, the researchers entered videos of problems for which they didn't know the explicit solution. An "air dancer" pulsated in front of a local used car lot in one of these videos. The system returned 8 variables after a few hours of analysis. Similarly, a footage of a Lava lamp generated eight variables. The application returned 24 variables when they gave a video clip of flames from a Christmas fireplace loop.
A particularly fascinating question was whether list of variables was unique to each machine or if it changed each time the application was run. "I often wonder if we have encountered intelligent extraterrestrial culture, if they have discovered the same physical rules as us, or if they have described the universe differently. I thought, "Lipson explained. "Some objects may look strangely complex because you're trying to understand them using the wrong variables."
The number of variables in the tests remained constant each time the AI ran, however the specific variables changed. So, yes, there are other ways to characterize the universe, and it's possible that our choices aren't ideal.
According to the researchers, this type of AI can assist scientists in discovering difficult phenomena for which theoretical understanding is falling behind the deluge of data—areas ranged from biology to cosmology. "While we employed media files in this work, any type of array data source might be used—for example, radar arrays or DNA arrays," noted coauthor Kuang Huang PhD '22.
The work is part of Lipson and Fu Foundation Professor of Mathematics Qiang Du’s decades-long interest in creating algorithms that can distil data into scientific laws. Previous software methods, such as Lipson and Michael Schmidt's Eureqa, could extract freeform physical principles from experimental data, but only when the variables were known ahead of time. But what if the variables aren't known yet?
Hod Lipson describes how the AI algorithm discovered new physical variables. Photographer: Hod Lipson/Columbia Engineering
Lipson, the James and Sally Scapa Professor of Innovation, says that researchers may be misinterpreting or failing to acknowledge numerous phenomena simply because they lack a good collection of variables to describe them. "People had known about objects moving fast or slowly for millennia, but it wasn't until the concept of velocity and acceleration was fully quantified that Newton could establish his famous law of motion F=MA," Lipson explained. Variables defining temperature and pressure had to be identified before thermodynamic rules could be defined, and so on for every other aspect of science. Variables are the foundation of any hypothesis. "What other rules have we forgotten because there are no variables?" says Du, who co-led the work.
Ishaan Chandratreya and Sunand Raghupathi, who assisted in data collection for the experiments, also contributed to the research. Boyuan Chen has been an assistant professor at Duke University since July 1, 2022. The research is part of a collaborative NSF AI institution for dynamical systems led by the University of Columbia, Washington and Harvard, with the goal of accelerating scientific discovery through the use of AI.
0 Comments