Exploring Interface Structure Between Perovskite Oxides Using Evolutionary Structure Search and Automated Design of Deep Learning Potentials via Neural Architecture Search

Perovskite oxides have attracted enormous interest in producing next-generation magnetic and ferroelectric devices. Interfaces between perovskite oxides warrant particular attention as they present distinctive charge transfer reactions and magnetic properties that are often absent in the constituent materials. To harvest the special functionality of perovskite oxide multilayers, developing a mechanistic understanding of their interface phases is imperative to establish materials design principles. In this study, we leveraged evolutionary algorithm and neural architecture search methods to design deep learning potentials and first principles simulations to systematically examine interface structures of La_1-<i>x</i>Sr<i>_x</i>CoO_3−δ/La_1-<i>x</i>Sr<i>_x</i>MnO_{3−<i>δ</i>} (LSCO/LSMO) bilayer system as a function of oxygen deficiency (<i>δ</i>) and strontium concentration (<i>x</i>). Neural network potentials proposed in the literature are typically designed based on intuition and a particular neural network architecture can induce systematic errors into the potential. To overcome this problem, we use a state-of-the-art neural architecture search method to design potentials by optimizing network architectures and hyperparameters over a variety of neural network ensembles. The end products are a set of optimized potentials that then are used to search for interface structures using evolutionary structure search. Since total energies are calculated using an ensemble of potentials, our framework naturally allows us to incorporate uncertainty quantification. To generate our potentials, we use a set of about 800 training structures generated with first principles calculations by using evolutionary structure search. Next, we use these potentials to explore interface structures for different <i>δ</i> and <i>x</i> – this combination of evolutionary structure search and automated design of potentials allowed us to explore over 50,000 interface structures corresponding to 25 distinct <i>δ</i> and <i>x</i> at LSCO/LSMO interfaces. We validate the predictions made using our potentials by (a) performing first principles calculations of energetically favorable structures and comparing them with total energies predicted using our potentials, and (b) comparing predictions with a structure search performed using evolutionary algorithms with first principles calculations (i.e., without potentials). By statistically analyzing the spatial distribution of oxygen vacancies, Sr atoms, and the distortion of Co-/Mn-centered octahedral units, we determine the complex interplay of oxygen vacancy concentration, valence states of transition metal ions, and electronic properties at perovskite oxide interfaces. The fundamental understanding provides useful insights in designing perovskite oxide multilayers with tailored properties for device applications.<br/><br/>This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory (LLNL) under Contract DE-AC52-07NA27344. This work was funded by the Laboratory Directed Research and Development (LDRD) Program at LLNL (project tracking code 21-ERD-005). Computing support for this work came from the LLNL institutional computing facility.