Evaluation of a Universal Neural Network Potential for Predicting Finite Temperature Properties Using Quasi-Harmonic Approximation

Computational materials design has been extensively studied with the development of DFT calculations and related optimization methods. However, material properties are usually evaluated at zero temperature due to the huge computational cost involved in DFT.<br/>As a result, the proposed material may not be experimentally synthesizable. For practical applications, it is also crucial to optimize the synthesis conditions to improve the material properties at an affordable cost. Therefore, it is necessary to establish a reliable computational procedure for evaluating the properties at a finite temperature.<br/>Predicting material properties at finite temperature requires accurate evaluation of thermodynamic quantities such as Gibbs free energy. There are two contributions to this accuracy: the accuracy of the potential energy calculated by DFT and the accuracy of the physical modeling of the phenomena. On the other hand, the calculation throughput, which is a compromise of the accuracy, is important for practical applications. Therefore, an appropriate method should be used to balance its accuracy and throughput.<br/>The Preferred Potential (PFP) implemented on Matlantis^TMis a recently developed graph neural network potential with the unique feature of universality[1]. PFP is trained on large DFT data sets, including not only stable crystals and molecules, but also surfaces and disordered structures. As a result, it is applicable to predict finite temperature properties of materials without compromising accuracy. In physical modeling of solids, the temperature dependence of thermodynamic properties can be well described by the Debye model under the quasi-harmonic approximation (QHA) of phonons, which takes into account the thermal expansion of the volume. QHA is suitable for predicting thermodynamic properties with moderate computational costs at the temperature range from room temperature to about two-thirds of the Debye temperature. By combining PFP and QHA, it is possible to efficiently predict material properties at finite temperatures with high accuracy.<br/>In this study, we have systematically validated the accuracy of PFP combined with QHA for predicting the thermodynamic properties at finite temperature. In particular, the calculated temperature dependence of the isobaric specific heats around room temperature, derived from the temperature derivative of the Gibbs free energy, is in good agreement with experiments. The temperature range in which the QHA is applicable has also been investigated with respect to the results obtained from molecular dynamics simulations. These results demonstrate that GNN potentials trained on large dataset well reproduce the potential energy curve around the local minimum, and are applicable to predict material properties at finite temperature.<br/>(1) S. Takamoto, et al. Nature Communications 2022, 13, 2991.