Bowen is from The Temple of Convi
Qubits | Official account QbitAI
What were the seconds of the Big Bang just happening?
This is arguably one of the most complex problems in physics, in the case of a special form of existence in the universe within a few millionths of a second of the Big Bang.
This is a "perfect liquid state" at ultra-high temperatures, which is of great significance for exploring the structure and environment of the original matter of the universe.
In the laboratory, this pattern must be successfully simulated in a harsh environment of 150,000 times the core temperature of the sun.
To analyze or process this highly complex physical form, supercomputers need to approximate its form for a long time, and classical AI or CNNs are difficult to make meaningful explanations based on the physics concepts in it.
But now, a paper in the top journal of physics, PRL, proposes a new neural network structure called L-CNN, which solves the above problem very well:
How to handle canonical invariants
Before we dive into the structure of L-CNN, let's clarify a fact:
What exactly is the task that traditional AI and CNNs can't do?
Take, for example, the "perfect liquid" mentioned at the beginning, which refers to the fact that at extremely high energies and temperatures, protons and neutrons are dismantled and recombined into a new form of matter called quark gluon plasma (QGP).
(The entire universe before the original formation of matter was in this form.)
When ai is introduced to analyze and deconstruct QGP patterns, its Gauge Symmetry must be considered.
Gauge symmetry refers to describing the same event in different ways, for example, we can describe an electron-photonic system in terms of a pair of phases and electromagnetic field potentials, or we can describe it in another pair, and both descriptions should give the same physical substance.
Physical quantities, on the other hand, are norm-invariant, so particle fields and their interaction forces, which appear to be described mathematically in different ways, may actually be the same physical state.
Traditional CNNs have a hard time calculating or analyzing these canonical invariants, and naturally cannot get meaningful computer simulation results.
The new method mentioned at the beginning, the Lattice Gauge Equivariant neural network, is a completely new method that can calculate or analyze canonical invariants that traditional CNNs cannot handle.
The entire approach is based on lattice gauge theory.
At lattice points, canonical invariants are usually described in terms of wilson loops of different shapes.
Specifically, a new convolutional layer is added that can form wilson rings of arbitrary shape in a continuous bilinear layer while preserving gauge Equivariance.
The set of all shrinkable Wilson rings can be generated by the above method, and with the addition of topological information from non-shrinking loops, all canonical connections can be reconstructed in principle.
With such neural networks, it is possible to make predictions about complex systems of multiple physics.
Paper author Andreas Ipp also gives an example of quark gluon plasma:
For example, L-CNN can estimate what the quark gluon plasma will look like at a later point in time without having to calculate each intermediate step in detail.
At the same time, it ensures that the system produces only results that do not contradict the symmetry of the specification, that is, results that are meaningful, at least in principle.
This was difficult to do with all previous computational methods, and L-CNN undoubtedly provides a new way of thinking for simulating complex physical phenomena.
In the future, it will also provide more help for exploring the environment in which life existed at the first instant, understanding the origin state of matter in the universe, and the study of black holes and the grand unified theory.
About the author
There are four authors, all from the Institute of Theoretical Physics at the Vienna University of Science and Technology (TU Wien).
In the lower right corner is the corresponding author of the paper, David I. Müller, a postdoctoral fellow at the Institute of Theoretical Physics at the Vienna University of Science and Technology, whose main research areas are high-energy physics, lattice gauge fields, and machine learning.
thesis:
https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.128.032003
Reference Links:
https://www.eurekalert.org/news-releases/941106