An international team of researchers has developed a method to tune a mathematical model of magnetic interactions that helps to realistically simulate and design materials with desired properties and predict their behavior before experimental testing. The research was published in Physical Review B.
Recently, much research has been focused on machine-learning interatomic potentials, which can ensure fast and accurate simulation of a material’s structure and properties. Quantum-mechanical methods, such as density functional theory, are very accurate but too time-consuming and computationally expensive. Machine learning accelerates the computations for large systems while offering virtually the same level of accuracy. However, physical reliability remains a major issue in machine learning applications.
In their new study, researchers from Skoltech, MIPT, and HSE University in collaboration with their international colleagues proposed an automatic learning algorithm for a machine-learning interatomic potential with magnetic degrees of freedom to accelerate the lengthy quantum-mechanical calculations in studying of paramagnetic materials without compromising accuracy.
Magnetic moments represent a new variable, which complicates the potential’s training. The simulation with a magnetic interatomic potential consists of two steps. The first step aims to optimize magnetic moments with fixed atomic coordinates and lattice parameters in order to minimize the system’s total energy. The second step involves molecular dynamics simulation, where magnetic moments are fixed while atomic coordinates and lattice parameters are varied based on magnetic interactions.
The training is also complicated by the magnetic moments present in the potential’s functional form. To solve this issue, the team developed an algorithm that automatically selects optimal configurations for the training dataset. The algorithm identified the configurations generated during the simulation and ran density functional theory calculations for the selected configurations. The results were then added to the training dataset.
“The standout feature of our potential is the ability to select configurations during the simulation, including the molecular dynamics simulation step. Since the potential is able to pick relevant configurations for further DFT simulation and refitting, the training dataset can be compiled automatically. The potential also takes into account the magnetic moments of the configurations selected during active learning,” said Ivan Novikov, an associate professor at the HSE Faculty of Computer Science, an associate professor at the MIPT Department of Chemical Physics of Functional Materials, and a senior research scientist at Skoltech AI.