Researchers are increasingly using neural networks as a machine learning method to analyze the atomic structure of complex materials. Although the method has been applied to phase transitions in single compounds, its transferability across compositions of binary compounds has never been tested.

Mangold et al. constructed a neural network potential (NNP) to train their machine learning model to analyze and predict the structural and vibrational properties of germanium (Ge) embedded with manganese-germanium (MnGe) nanostructures over a wide range of chemical compositions. Nanostructured MnGe has a number of thermoelectric properties, such as large magneto-thermopower, for potential applications in spintronics and thermoelectric energy conversion. However, its thermal conductivity is mostly unknown.

The main focus of their study was on phonons, or quantized lattice vibrations, which play an important role in the thermal transport of materials. Using three different chemical compositions of MnGe, they trained their system to reproduce electronic structure calculations based on density functional theory (DFT), which was also used to validate the NNP.

To extend neural network applicability to nanostructured materials, they added a superlattice with alternating layers of Ge and Mn5Ge3 to the training set. In addition, they tested the system’s transferability to Mn11Ge8, which was not included as part of the training set, and found that the neural network was able to reproduce its structural and vibrational properties with relatively good accuracy.

“From a fundamental point of view, we validated the neural network approach across diverse chemical environments,” author Davide Donadio said. “Our aim is to extend this research to even more complex systems, such as Ge with nano-inclusions of Mn5Ge3, addressing structural and thermal properties beyond the current limits of DFT calculations.”

Source: “Transferability of neural network potentials for varying stoichiometry: phonons and thermal conductivity of MnxGey compounds,” by Claudia Mangold, Shunda Chen, Giuseppe Barbalinardo, Jörg Behler, Pascal Pochet, Konstantinos Termentzidis, Yang Han, Laurent Chaput, David Lacroix, and Davide Donadio, Journal of Applied Physics (2020). The article can be accessed at