No, AI has not discovered a new type of physics

Can a computer algorithm discover something new in physics? It’s a great question. new research paper About the topic sexy title AI may have just invented ‘alternative’ physics.

The term “alternative physics” is a lot like “alternative facts”, but let’s investigate anyway. How does the performance of this computer program compare to that of an actual physicist? Or even an average student?

Newton’s mechanics

Isaac Newton It was Unparalleled genius. The English encyclopedist not only unified the studies of motion and gravity, but invented the mathematical language in which he describes them. The concepts of classical mechanics that Newton brought into being lie behind most of the physics that has since been invented. His concepts were later reformulated in a new mathematical language in the eighteenth century by the exceptional continental physicists Joseph Louis Lagrange and Leonhard Euler.

Newtonian mechanics requires analysis of the directional forces acting on massive bodies. If you took an introductory physics class in high school or college, you’ve seen these problems: boxes on inclines, pulleys, and buggies. You draw arrows going in different directions and try to balance the forces. It works well for small problems. As the problems become more complex, this method continues to work, but it becomes brutally boring.

Using Lagrange’s formula, if two aspects of the nature of a system can be determined, then the problem can be solved using only calculus. (Yes, ‘only’ calculus: squashing derivatives is much easier than solving very complex free-body diagrams where the arrows change at every position.)

The first thing to understand is the energy of the system, i.e. the (kinetic) energy of motion and the (latent) energy stored by the composition of the system. The second crucial thing is to choose the appropriate coordinates or variables for the movement of the system.

Subscribe to get unexpected, surprising and touching stories delivered to your inbox every Thursday

Imagine a simple pendulum, like that on an old-fashioned clock. A bob pendulum has kinetic energy from its oscillating motion and potential energy due to its location (altitude) within the gravitational field. The position of the pendulum can be described by one variable: its angle relative to the perpendicular. The Lagrangian solution for the motion of the pendulum can then be calculated relative ease.

Solving more complex problems in mechanics requires discovering the appropriate number of variables that can describe a system. This is easy in simple cases. In moderately complex cases, it is a student-level exercise. In very complex systems, the work of a professional may or may not be impossible. This is where the artificial intelligence “physicist” comes in.

Artificial intelligence physicists outperform undergraduates

The computer is set to analyze a problem Pendulum hanging on another pendulum. This problem requires two variables – the angle of each pendulum on the vertical axis – or four if a Cartesian (xy) coordinate system is used. If both pendulums pop hanging from the springs Instead of solid bars, the two variable lengths of the spring were added to get six variables in the Cartesian system.

The computer was asked to determine the number of variables needed to calculate the above problems. How did an artificial intelligence physicist do? Not great. For a solid pendulum on a pendulum, she gave two answers: ~7 and ~4-5. (The correct answer is 4 variables.) Similarly, I calculated ~8 and ~5-6 for the double-spring pendulum. (The correct answer is 6 variables.) Researchers praise smaller estimates as being close to correct answers.

But after looking into the details in the paper Supplementary materialHowever, the result begins to fall apart. The computer didn’t actually calculate 4 variables and 6 variables. Its best accounts were 4.71 and 5.34. None of these answers are rounded off to the correct answer. The four-variable problem is an intermediate university physics problem, while the six-variable problem is a more advanced university problem. In other words, an undergraduate physics student is much better than an AI physicist at grasping these problems.

The AI ​​physics world is not ready to install

Researchers continue to ask the program to analyze complex systems that not only contain an unknown number of variables, but it is unclear whether classical mechanics can describe systems at all. Examples include a lava lamp and fire. AI does an acceptable job of anticipating small changes in these systems. It also counts the number of variables required (7.89 and 24.70, respectively). The correct answers to these problems will be “new physics” in a sense, but there is no way of knowing whether or not AI is correct.

Using AI to analyze unknown systems is a great idea, but currently AI cannot get the right easy answers. Thus, we have no reason to believe that he corrects the difficult things.