As one of science’s more rigorous and restrictive fields, physics is filled with complex calculations that must be done precisely in order to unveil its secrets.
All physical laws are described as mathematical relationships between state variables. These variables give a complete, non-redundant description of the relevant system.
These mathematical relationships describe nearly all physical laws in nature.
However, before even the most basic equation can be constructed, scientists have to figure out the variables of a system, which are an essential precursor to written equations.
Consider F=MA, Newton’s classic fundamental equation of force.
Newton first needed to comprehend the notions of acceleration, mass, and force in order to construct such an equation. Hod Lipson, a professor of engineering and data science at Columbia University, told VICE that this task lacks a clearly defined path to pursue.
“It’s an art, there is no systematic way,” Lipson explained. “It’s almost like, how do you discover the alphabet? It just happens organically.”
Lipson and his colleagues at the Creative Machines Lab aim to learn more about how this process of discovery works and how it may be enhanced using machine learning to find hidden, alternate physics that human scientists might have overlooked.
To do this, Lipson and colleagues created a machine learning algorithm that can analyze physical phenomena by “viewing” videos, like the swing of a double pendulum or the flicker of a flame, and generate the necessary amount of variables to explain the behavior.
The program was able to predict the appropriate amount of variables within 1 value (e.g. 2.05 variables to describe a single pendulum instead of 2) for known systems and it was even able to forecast variables for unidentified systems.
Although this algorithm is not the first to analyze data and attempt to draw a physical relationship from it, according to Lipson, this work stands alone because it is the first to not give the algorithm any information on the quantity or type of anticipated variables in a system.
This means that the system is not constrained to simply consider variables from a human perspective, which according to Lipson may be essential for revealing hidden physics in complex systems.
“It’s not that people are toiling away day and night to look for these variables and this can expedite the process,” said Lipson.
“It’s more that we are probably overlooking a lot of stuff,” he continued. “But so much is hinging on those variables that we thought if we could throw some AI power at this, maybe we’ll discover things that are super useful and will change the way we think.”
Boyuan Chen, who was the paper’s first author and is currently an assistant professor of engineering at Duke University, and Lipson sent videos of dynamic motion in various degrees of complexity to their algorithm to set it up for success. This comprised both well-known motions like swing sticks and double pendulums as well as motions that were yet unclear, such lava lamps, flickering fires, and inflatable air dancers.
After analyzing these videos, the AI tried to simulate the phenomenon a few steps into the future and come up with a list of smaller and smaller variables that were driving the activity. The AI would finally spit forth the bare minimum of variables needed for the system to correctly capture the motion.
The AI was largely effective in determining the appropriate number of variables, however, there is a significant flaw that will prevent it from being used in science labs any time soon. It can inform scientists that a system contains a particular number of variables, but it does not yet have the vocabulary to specify what those variables are. For instance, it reported eight variables for the “air dancer” and 24 variables for the fireplace.
For the time being, Chen isn’t really concerned about this.
“What we have right now is like a general framework,” Chen explained. “One thing that will be very interesting is to collaborate with experts who have data and an intuition about what that data is doing. What we want to do is to help them to discover what they do not know yet about the data.”
They anticipate that the algorithm will begin to exhibit patterns that make it easier for human colleagues to understand its findings in the future. This will be the next significant development in scientific research, according to Lipson.
“Humans have been doing this for 300 years, and it seems to me like we have kind of reached the end of what we can do manually,” Lipson said. “We need something to help us go on to the next level.”
The findings were published on July 25, 2022 in a study titled “Automated discovery of fundamental variables hidden in experimental data” in the journal Nature Computational Science.
Typos, corrections and/or news tips? Email us at Contact@TheMindUnleashed.com