The fastest supercomputers can also take hours to simulate complex natural phenomena, Science reported. As an algorithm that can simulate quickly, the emulator undoubtedly provides a shortcut. A study published in arXiv (PDF) shows that artificial intelligence can easily generate precise emulators that can accelerate simulations in all areas of science by billions of times.
The technique, known as Deep Simulator Network Search (DENSE), relies on a common neural structure search developed by Melody Guan, a computer scientist at Stanford University. It randomly inserts the compute layer between the input and output of the network, testing and training the generated lines with limited data.
If the added compute layer improves performance, it is likely to be used in future emulator changes, and repeating this process can improve the emulator. Using DENSE technology, the researchers developed 10 simulators for physics, astronomy, geology, and climate science.
DENSE’s emulators perform well, moving 100,000 to 2 billion times faster than other simulators. Moreover, these emulators are very accurate: the results of the astronomical emulator are more than 99.9% consistent with the full simulation, and the neural network simulator is much better than the traditional emulator in 10 simulations.