Sebastian Bichelmaier1,2,Jesús Carrete Montaña1,Georg K.H. Madsen1
Technical University of Vienna1,KAI GmbH2
Sebastian Bichelmaier1,2,Jesús Carrete Montaña1,Georg K.H. Madsen1
Technical University of Vienna1,KAI GmbH2
As computing power has constantly been increasing, first-principles calculations have been moving closer to applications and materials relevant to industry. However, studying the temperature-dependent behavior of strongly anharmonic solids is still methodologically challenging. In particular, molecular dynamics (MD) and stochastic sampling approaches are hindered, on the one hand, by the computational cost of accurate ab-initio simulations, and, on the other hand, the lack of accuracy and transferability of traditional force fields (FFs).<br/>We propose an automatically differentiable neural network FF coupled with a physically inspired potential for increased transferability. This neural network representation of the potential energy surface can be used to perform free energy studies using stochastic and MD methods, essentially allowing for a highly efficient and accurate exploration of a compound’s temperature-dependent properties.<br/>We discuss the example of HfO<sub>2</sub>, a material of high importance, both as a high-k dielectric and as a ferroelectric. Using the method above, we construct a neural network FF transferable to several different phases and employ it to study temperature-dependent phenomena arising from the complex and multi-faceted potential energy landscape of HfO<sub>2</sub>.