In the last years machine learning has attracted, also among physicists, great attention. It turned out that the algorithms, while having vast applications in the decision-making (for example in games such as chess or AlphaGo), pattern recognition, medical diagnosis and finance, can be applicable also in the area of condensed matter physics.
The study of many-body systems is a challenging task, as the size of the Hilbert space and consequently the amount of data to analyze, grow exponentially with the size of the system. Therefore, such systems seem to be a perfect 'material' for machine learning algorithms, which are especially well suited to deal with big and complex sets of data. Such an approach has already been successfully used, e.g., for the Ising model. By analyzing Monte Carlo-generated samples, deep neural networks were able to identify with quite good precision the transition temperature [1].
However, a problem arises when one has to deal with non-conventional phase transitions, such as the Berezinsky-Kosterlitz-Thouless (BKT) transition occurring without an ordered phase or symmetry breaking. As a result, the difference between the high- and low-temperature phases is much more subtle than in the case of Ising-type models. Therefore, unambiguous identification of the critical temperature is much more difficult. The BKT transition transforms the system from bound vortex-antivortex pairs at low temperatures to unpaired vortices and anti-vortices above the critical temperature. Thus, it is topological in nature. It was demonstrated in Ref. [2] that it is difficult to determine the BKT phase transition in the classical XY model with the help of neural networks.
In our study, we demonstrate how one can overcome some of these difficulties. We test our approach on the classical XY model but also on its quantum version as well as on the phase-fermion model [3], where the BKT transition also occurs.