This space simulation model is so fast and accurate that its creators do not understand why it works - Photo 1.

This space simulation model is so fast and accurate that its creators don't understand why it works


For the first time in history, space physicists have successfully used artificial intelligence to create a complex three-dimensional model of cosmic simulations. The result is such a precise, complex and energetic structure, that its creators do not understand how the whole model works.

"We can operate this simulator model in just a few milliseconds, while other fast universe reconstruction models take several minutes to boot."The study co-author, Shirley Ho, head of the research team at the Center for Cosmic Physics Calculation under the Flatiron Institute, New York City, said. "Not only that, the model is much more accurate".

Speed ​​and accuracy of the Density and Depth Swap Model (abbreviated as D3M) is still not something that surprised the researchers. That must be the ability to accurately describe what the Universe would be like if it changed some of the basic elements of the Universe, such as how much dark matter existed in the Universe; must add D3M has never received any training data (like the way of putting data into artificial intelligence), only based on the basic elements inherent.

"Like an image recognition software that only knows cats and dogs, it suddenly identifies elephants"Professor Ho explained.

Ho and colleagues introduced D3M last June 24 at an event that took place at the National Academy of Sciences. The presentation was held by Siyu He, an analyst and researcher at the Flatiron Institute.

This space simulation model is so fast and accurate that its creators do not understand why it works - Photo 2.

Compare the accuracy of the two Universe models, with the color column on the left showing the error rate. New model on the left, D3M, both faster and more accurate than the current method used.

The simulator model is similar to D3M has become an indispensable tool in the construction of cosmological theories. Scientists want to know how the Universe evolved through periods, under different conditions, such as how dark energy affects the state of the Universe. Similar studies are time-consuming, when thousands of simulated programs must be run at a time. Having a fast and accurate simulation model is a remarkable milestone.

D3M created a model of how gravity creates the Universe. Often, researchers will only focus on the gravitational force of it, up to this point and with the doctrines we have, the most important force when considering the evolution of the infinite Universe.

The most accurate Universe simulation models calculate how gravity moves billions of individual particles that still float in space, throughout the long simulator time equivalent to the age of the Universe. To be accurate, it takes time to calculate, about 300 hours of machine running continuously for an emulator model. The faster the speed, the shorter the time, the lower the accuracy.

This space simulation model is so fast and accurate that its creators do not understand why it works - Photo 3.

The model of the Cosmic Universe was developed by the Kavli Institute of Cosmology, University of Chicago.

But new research uses the power of neural networks to make an accurate simulation model: D3M formed after the machine learning system received 8,000 different simulator models, with the most accurate calculations possible. Neural networks receive data and calculations to create the final model, then researchers compare expected results and actual results.

D3M surprised the researchers. To make a 600 million light-year universe, a slow-model – would probably take a few hundred hours, the "fast-paced" model would take several minutes. D3M takes 30 milliseconds, with amazing accuracy, with only 2.8% error; The rapid model has an error rate of up to 9.3%.

The ability to handle elements that are not in the input data makes D3M added an impressive part, making it an unprecedented versatile model of simulated Universe. And its application can also reach out to the industry that is supporting it successfully.

"We can build a common playground for machine learning systems, analyze why the new model can calculate such good extrapolation, why elephants can be recognized while only learning about cats and dogs"Professor Ho said. "This is a two-way street, which is beneficial for both aerospace science and deep learning".

Refer to Phys.org


Related Posts

Leave a Reply

Your email address will not be published. Required fields are marked *