Man versus machine: can AI do science?

Phase diagram produced by the machine. The phase boundaries previously determined by the scientists without the machine are given for comparison by dashed and solid lines.

Facial recognition in airports and autonomous driving are two daily life examples of the recent progress in artificial intelligence. But can machines also facilitate the work of a physicist by independently interpreting experiments or simulations?

A team of scientists based at the LOMA, the Okinawa Institute of Science and Technology (OIST) and the Ludwig Maximilian University of Munich have shown that machines can indeed help theorists, solving complex problems just as accurately as scientists, but considerably faster. The authors demonstrate that machine learning can successfully unravel the complex phase diagram found in classical Monte Carlo simulation of a highly frustrated magnet, identifying and correctly characterizing all of the different phases present, without training or supervision.

A key aspect is that the method used is not a “black box”, but provides interpretable results such as analytical expressions for the order parameters of each of the broken-symmetry phases. It also provides analytical expressions of the local constraints that characterise each spin liquid, and underpin their topological properties. This work shows machine learning can mature into a trustworthy tool for exploring exotic phases in many-body systems.

Bibliography :
Identification of emergent constraints and hidden order in frustrated magnets using tensorial kernel methods of machine learning,
J. Greitemann, K. Liu, L. Jaubert, H. Yan, N. Shannon and L. Pollet,
Phys. Rev. B 100, 174408 (2019),
DOI hal Summary