快猫短视频

A machine-learning AI can solve physics problems by simplifying them to be more symmetric.

鈥淭here are many, many cases in the history of science where people thought things were more complicated than they actually were because they hadn鈥檛 found the most simple description of it,鈥� says at the Massachusetts Institute of Technology.

To create their AI, Tegmark and , also at MIT, used an artificial neural network, a computing set-up inspired by the way neurons connect in the brain. They programmed it to transform problems with complex coordinate systems 鈥� meaning how problems are described geometrically 鈥� to ones that show some form of symmetry. These simpler, more symmetric systems can often be easier for physicists to solve.

鈥淭o really deal with concrete problems, we have to resort to specific coordinate systems so we can write down some equations and aim to solve them,鈥� says Liu.

In the past, finding the simplest coordinate system hasn鈥檛 always been easy. An example is one of the first equations to describe a black hole, which was unable to explain what happens at the event horizon, the point at which light can鈥檛 escape the black hole鈥檚 gravity. 鈥淚t took 17 years for people to realise that there is a transformation that can simplify this system,鈥� says Liu. 鈥淏ut our tool only takes half an hour [to get to this stage].鈥�

Tegmark and Liu tested their system on six well-known physics problems, including the black hole one, that have already been solved, both from classical mechanics and general relativity, and it rediscovered the symmetry in each scenario.

To discover the hidden symmetries in each system, the AI is given a list of known symmetries, and then tries transforming the complex coordinate systems thousands of times until it finds a match with one of those known symmetries.

Solving problems that humans have already cracked proves that the model works, but the AI will be most valuable in helping to solve problems for which there aren鈥檛 currently solutions. 鈥淲hat I鈥檓 most excited about now is applying this to those areas of physics where you suspect that there might be some more simplicity to find that we haven鈥檛 found yet,鈥� says Tegmark.

Solving problems in physics using machine learning in this way is a top-down approach, using large amounts of pre-existing data, says at City, University of London. But there are other methods that use machine learning to solve maths and physics problems by building from the bottom-up, such as .

Both methods hold promise for discovering new maths and physics, says He. 鈥淚鈥檓 an optimist, so I think certainly in the near future, we will see significant new results in pure mathematics and theoretical physics which have been discovered by [taking] these two directions in tandem.鈥�