Of the 100,000 equations that had to be solved, only four were left! This computational feat was achieved with artificial intelligence and now makes the solution to the famous problem of quantum physics, describing the behaviour of electrons moving in a grid more accessible. This achievement can help to develop materials with desired properties such as superconductivity.
Hubbard model
Thus, even with a small number of electrons, the problem requires enormous computing power because physicists have to deal with all the electrons at once. With more electrons, the rate of confusion becomes even higher and the computation becomes exponentially more difficult.
A machine "capable of detecting hidden patterns"
In order to ensure the most accurate display of the system, the management team must take into account all possible links between electrons. In addition, equations are particularly complex because each of them is a pair of interoperable electrons. So researchers from the Flatiron Institute in New York have decided to use a machine learning tool.
"", explains Domenico De Santa, a visiting scientist from the Flatiron Institute's Computer Quantum Physics Center, and co-author of this approach.
More specifically, the machine learning program creates connections within a normal-size renormalization group. Then the neural network regulates the force of these connections until it finds a small set of equations that generate the same solution as the original renormalization group. For De Santa, it's basically a machine that can detect hidden patterns.
The result of the program exceeded the team's expectations: the Hubbard model physicist was reduced to just four equations. This means that only four pixels will now be required to visualize the problem. In other words, the study of the emergent properties of complex quantum materials has now become much more manageable.
Potential applications in cosmology and neuroscience
Let's recall that superconductivity
Using superconductivity at more reasonable temperatures can result in much more efficient electrical networks and devices. For this reason, physicists try to predict using different models.
Like any machine learning algorithm, the algorithm used in this study should have been pre-trained in the data set; the training took several weeks. Modelling only covers a relatively small number of variables in the grid, but now that the program is trained, it can be adapted to work on other issues of condensed matter physics, says researchers. According to De Santa, this method can be beneficial in other areas that deal with renormalization groups such as cosmology and neuronauce.
The real test, the team notes, will be to check how well this new approach works on more sophisticated quantum systems, such as materials where electrons interact over long distances. Dee Sante and his colleagues are also studying that their algorithm actually "knows" about the system, which can provide additional information that physicists would find difficult to decipher.
In the meantime, this work demonstrates the possibility of using artificial intelligence to extract compact representations of correlate electrons, "", the team concludes in its article.