BEIJING, Dec. 5 ( Xinhua) — Since Newton’s time, how to predict the motion paths of three celestial bodies orbiting each other has been a headache for physicists, according tomedia reports. Now, with artificial intelligence, it’s a matter of time.
Newton first raised the The Three-Body Problem in the 17th century, but it proved extremely difficult to solve it in a simple way. Gravitational interactions between three objects (such as planets, stars, and satellites, etc.) form a chaotic system that is complex and highly sensitive to the initial state of each object.
Researchers have tried to solve the Three-Body Problem with software, but it can often take weeks, if not months, to complete the calculation. So the researchers decided to try the neural network. This is a regular type of artificial intelligence that roughly simulates how the brain works. And they’re building algorithms that solve them 100 million times faster than Brutus, the most advanced software program today. This would be a “priceless treasure” for astronomers who study the behavior of clusters and the evolution of the universe. If this neural network system works properly, the speed at which the answer is drawn will be unprecedented. In this way, we can further study deeper questions, such as “how gravitational waves are formed” and so on.
Neural networks must be trained by entering large amounts of data before they can be predicted. So the researchers first used Brutus software to generate 9,900 simplified versions of The Three-Body Problem Scenario for training neural networks. Next, the researchers tested the neural network with 5,000 new scenarios to see if they could accurately predict the evolution of these scenarios. It turns out that the prediction is not only very close to Brutus, but also completes in an instant. By contrast, the average calculation time for Brutus is nearly two minutes.
Programs such as Brutus are calculated so slowly because they use the “brute force calculation”, the method of poverty, which calculates every small step of the celestial trajectory. The neural network only analyzes the motion trajectory produced by these calculations and summarizes the corresponding laws from them to predict the future evolution of the system.
But for larger, more complex systems, it’s not that simple. This algorithm is currently only in the proof-of-concept phase, only learning some simplified version of the situation, but if you want to use more complex systems, even “quad system”, “five-body system” training, you first use Brutus to generate a large amount of data, which is not only time-consuming, but also expensive. This is the bottleneck that this neural network is currently experiencing.
To solve this problem, several researchers could first create a generic database using programs such as Brutus. But this requires a set of standard protocols to ensure that all data is standard and consistent.
In addition, the neural network itself has some problems to solve. For example, it can only run for a specified amount of time, but it is impossible to predict in advance how long a situation will take to evolve, so the algorithm may have “bounced out” before the problem was solved.
However, the researchers didn’t intend to single out the nervous system, arguing that it would be best for programs like Brutus to do most of the “hard work”, while neural networks were responsible for only the simulations that required complex calculations. (Leaf)