As one of science's most rigorous and rigid disciplines, physics is filled with long equations and complicated measurements that must be just right to uncover their secrets. Nevertheless, before even the simplest equation could be put together, scientists first had to figure out a crucial precursor: a system's variables. The fundamental equation of force is F=MA. To compose such an equation, Newton needed to understand acceleration, mass, and force. Columbia University professor of engineering and data science Hod Lipson told Motherboard this is a difficult task. 

A systematic method isn't possible," Lipson says. As if you discovered the alphabet? "Organic."Lipson and colleagues in Lipson's Creative Machines Lab want to better understand how this process of discovery occurs and how machine learning can help uncover hidden, alternate physics not discovered by humans. By watching videos of physical phenomena such as a pendulum swing or a flame flicker, Lipson and colleagues have developed a machine learning algorithm that can generate the variables needed to explain the action. 

The algorithm correctly predicted the number of variables within one value and even the number of variables within an unknown system. A study titled "Automated discovery of fundamental variables hidden in experimental data" was published last week in Nature Computational Science. While Lipson's algorithm is not the first to analyze data and try to establish a physical relationship from it, it stands out since it reports no information to the algorithm about the expected number or type of variables. The system is not limited to looking for variables through a human lens, which Lipson says is crucial for uncovering hidden physics.

The process can be accelerated by not spending 24 hours a day looking for variables, says Lipson. He continues, "We're probably overlooking a lot."But so much depends on those variables that maybe we'll discover something super useful if we throw some AI power at it."Lipson and colleagues, including the paper's first author and now an assistant professor of engineering at Duke University, Boyuan Chen, fed their algorithm videos of dynamic motions of varying complexity. The motions included known ones, such as pendulums and swing sticks, as well as ones that are not yet understood, such as lava lamps, flickering fires, or inflatable air dancers. 

The AI modeled the phenomena a few steps into the future and created a list of increasingly smaller variables responsible for the action. Lastly, the AI would spit out the minimum number of variables required by the system to capture motion accurately. Though the AI did quite well at finding the right number of variables, there is one major catch that will keep it out of science labs. Currently, it does not have the language to describe what variables are in a system, for example, it returned eight variables for the "air dancer," and 24 for the fireplace.

AI systems have long been regarded as complex black boxes, making it difficult for scientists to reverse engineer any particular decision. We have a general framework right now, says Chen. “Collaboration with experts with data and intuition about the data will be very interesting. Our goal is to help them discover what they don't know about the data yet."Lipson says future research may examine systems beyond physics, such as disease evolution. The algorithm's patterns may help better communicate its findings to human collaborators in the future. Lipson believes this will be the next great scientific advance. The human race has been doing this for 300 years, and we seem to have reached the end of what we can do manually. To go on to the next level, we need something."