The maps of the universe, faster and with the same accuracy

If you think the galaxy is large, compare it to the size of the universe: it is just a small point, along with a large number of other small points, forming groups that gather in super dishes, which in turn weave into the threads of the voids – which are a massive 3D skeleton in our world.

If that gives you the dizziness as you wonder how one can understand or even “see” something very wide, then the answer is: It is not easy. Scientists combine the physics of the universe with data from astronomical tools and building theoretical models, such as Eftoflss (the theory of the effective field of large -scale structure). They are fed with notes, and they describe these “cosmic web” models statistically and allow their main parameters.

Models like Eftoflss, however, require a lot of time and resource account. Since our astronomical data groups are increasing dramatically, we need ways to reduce the analysis without loss of accuracy. That is why the simulations are: it is “imitation” of how the models respond, but they work much faster.

Since this is a kind of “abbreviation”, what is the risk of loss of accuracy?

An international team, including, among others, has published Anaf (Italy), the University of Parma (Italy) and the University of Waterloo (Canada) in Magazine of cosmology and astronomical particle physics Simulator testing. JL, which they designed. The study is entitled “effort. JL: a fast and discriminatory emulator of the theory of the effective field of the universe’s effective structure.”

It explains that the effort. JL mainly provides the same right as the model that imitates it – sometimes even the fine details – while playing in minutes on a standard laptop instead of the super computer.

“Imagine the desire to study the contents of a cup of water at the level of their microscopic components, or individual atoms, or even smaller: in theory, you can. But if we want to describe in detail what happens when the water moves, the explosive growth of the required accounts makes it impossible in practice,” explains Marco Beni, a researcher at the University of Waterloo and the first author of the study.

“However, you can encrypt certain properties at the microscopic level and see its effect at the sample level, which is the movement of the liquid in the glass. This is what the effective field theory does, that is, a model like Eftoflss, where the water is perfect is the universe on very large scales and microscopic paradigms are small physical operations.”

The theoretical model statistically explains the structure that leads to the collected data: the astronomical notes of the symbol, which calculates “prediction”, are fed. But this requires time and a big account. Looking at the size of today’s data – and what is expected from the investigative studies that it has just started or will come soon (such as Desi, which has already released the first batch of data, and the tradition), it is not practical to do this comprehensively each time.

“This is the reason why we are now moving to the simulations like us, which can cut time and resources significantly,” Bonici continues.

The emulator mainly simulates what the model does: its essence is a nerve network that learns to link the input parameters to the forecasts of the already calculated model.

The network is trained on the outputs of the model, and after training, it can be circulated to groups of parameters that you have not seen. The emulator does not understand the same physics: it knows the responses of the theoretical model well and can expect what will lead to new inputs.

Votation. Asala JL is that it reduces the training stage by building in knowing the algorithm that we have already we have about how predictions change when parameters change: instead of making the network “re -learning”, it uses it from the beginning.

Votation. JL also uses gradients – IE, “how much and in any direction” change predictions if you adjust the teacher small – another component that helps the emulator to learn from much lower examples, cut account needs and allow them to operate on smaller machines.

Such a tool needs intense verification: if the emulator does not know physics, what is the extent of making sure that its shortcut leads to correct answers (i.e. the same model that the model will give)? The exactly published study answers this, indicating that effort. Girl – both both simulation and real data – in a close agreement with the model.

“In some cases, where you have to reduce part of the analysis to accelerate things, with an effort. JL was able to include these lost pieces as well,” Bonici concludes.

Votation. JL thus appears as a valuable ally to analyze data data coming from experiments such as Desi and Euclid, which is to deepen our knowledge of the universe on large scales.