Deepfaking Low-Energy Interface Structures Using a Generative Adversarial Network

Understanding the properties of solid-solid interfaces is important to a wide range of applications, from energy-storage devices to electronic devices to surface coatings. However, interfaces are challenging to model explicitly because the interface structure is not limited to a single possible configuration, unlike the unit cell of a single phase which often has only one ground-state configuration. Certain interfaces can exhibit significant disorder due to atomic rearrangement, chemical mixing, variable composition, etc., where there is likely an entire distribution of structures that reflect the real interface. In such cases, an interface structure search, that extends beyond simply pressing two slabs of materials together, would be necessary to identify energetically favorable interface structures. A random search of low-energy structures by generating atoms randomly at the interface is computationally expensive and inefficient, leading to the need for a more intelligent search scheme. We will demonstrate that the interface structure search problem can harness the power of machine learning not only through machine learning interatomic potentials to accelerate relaxations, but also through generative modeling. By training a generative model on a dataset of low-energy interface structures, the generative model can learn the underlying distribution and generate new low-energy structures. The generative model we have developed is a generative adversarial network (GAN), applied to lithium halide solid electrolyte/cathode interfaces for all-solid-state battery application. We will show that the GAN is capable of learning from the low-energy structures to generate a large number of atoms (> 250 atoms), and outperforms random search.