Scientists mark a milestone by using artificial intelligence to produce complex 3D simulations of the universe that are fast and accurate.
The simulation works so well even beyond its parameters, much to the surprise of the creators who have ended up baffled at the AI's incredible capabilities.
In a paper published in the journal Proceedings of the National Academy of Sciences, the astrophysicists shared the results of their project known as the Deep Density Displacement Model or D3M. Its speed and accuracy set it apart from the previous attempts of universe simulation.
According to study coauthor Shirley Ho, who is a group leader at the Flatiron Institute's Center for Computational Astrophysics and an adjunct professor at Carnegie Mellon University, their new AI can run simulations in just a few milliseconds, compared to other "fast" simulations that take several minutes to complete. It is also much more accurate.
In fact, the accuracy of D3M is astounding, leaving Ho and the rest of her team shocked. The model has been found to accurately generate a simulation of the universe with some of the parameters tweaked, such as the amount of dark matter in the cosmos — even though D3M never received training data on these parameters changes.
"It's like teaching image recognition software with lots of pictures of cats and dogs, but then it's able to recognize elephants," Ho explained. "Nobody knows how it does this, and it's a great mystery to be solved."
How It Works
The new D3M reportedly models how gravity shapes and affects the universe, which the researchers focused on since gravity is the most important force in the universe's large-scale evolution.
The most accurate simulations calculate how gravity shifts billions of individual particles over the universe's entire lifetime, which takes about 300 hours for each simulation. Faster simulations exist, but its accuracy is compromised in the process.
The D3M training involved feeding its deep neural network 8,000 different simulations from one of the existing high accuracy models. The neural networks run calculations on the training data and the researchers compare its results with their expected outcome. Further training allows these neural networks to adapt and produce faster, more accurate results.
After completing the training of D3M, the researchers ran simulations of a box-shaped universe extending 600 million light-years across. Then, they compared the results to slow-yet-accurate models and fast models. While the former took hundreds of hours to complete each simulation, the latter took just a couple of minutes. Impressively, the D3M completed the simulation in just 30 milliseconds. It's also incredibly accurate with a relative error of 2.8 percent when compared with the high-accuracy model.
Additionally, the D3M's unexpected ability to handle variations in its parameters could pave the way for further innovation on artificial intelligence and machine learning.
"We can be an interesting playground for a machine learner to use to see why this model extrapolates so well, why it extrapolates to elephants instead of just recognizing cats and dogs," Ho said. "It's a two-way street between science and deep learning."