Constructing realistic effective spin Hamiltonians with machine learning approaches

The effective Hamiltonian method has recently received considerable attention due to its power to deal with finite-temperature problems and large-scale systems. In this work, we put forward a machine learning (ML) approach to generate realistic effective Hamiltonians. In order to find out the important interactions among many possible terms, we propose some new techniques. In particular, we suggest a new criterion to select models with less parameters using a penalty factor instead of the commonly-adopted additional penalty term, and we improve the efficiency of variable selection algorithms by estimating the importance of each possible parameter by its relative uncertainty and the error induced in the parameter reduction. We also employ a testing set and optionally a validation set to help prevent over-fitting problems. To verify the reliability and usefulness of our approach, we take two-dimensional MnO and three-dimensional TbMnO3 as examples. In the case of TbMnO3, our approach not only reproduces the known results that the Heisenberg, biquadratic, and ring exchange interactions are the major spin interactions, but also finds out that the next most important spin interactions are three-body fourth-order interactions. In both cases, we obtain effective spin Hamiltonians with high fitting accuracy. These tests suggest that our ML approach is powerful for identifying the effective spin Hamiltonians. Our ML approach is general so that it can be adopted to construct other effective Hamiltonians.