Machine-Learning Molecular Dynamics Simulations of Defects and Domain Walls in Ferroelectric Potassium Niobate

Ferroelectrics are important piezoelectric materials extensively used in actuators and sensors. Among these, lead zirconate titanate (PZT) stands out as a commercially available piezoelectric material known for its excellent performance. However, due to increasing concerns about the environmental and health impacts of lead, there is a significant demand for developing high-performance lead-free piezoelectric materials. Various lead-free systems, including potassium sodium niobate (KNN), barium titanate, bismuth ferrite, and bismuth sodium titanate, have been explored. Among these, the KNN system exhibits relatively good comprehensive performance, making it a focal point in lead-free piezoelectric material research.<br/>Developing high-performance lead-free piezoelectric materials is highly challenging because the performance of ferroelectric piezoelectric materials is closely related to interactions among microstructures at different scales, such as phase structure, domain walls, defects, and grain boundaries. Understanding these complex interactions is crucial for designing high-performance piezoelectric materials. However, due to limitations in experimental characterization methods, it is difficult to distinguish the contributions of various microstructures to the performance directly. Theoretical simulation methods can potentially elucidate the physical mechanisms behind these phenomena. First-principles density functional theory calculations can simulate various properties of ferroelectric materials, but due to the complexity of this computational method, calculations are limited to the electron and atomic scales. Under current computational resources, this method still cannot effectively simulate microstructures on a large spatial scale and their evolution over a long time scale. On the other hand, molecular dynamics simulation is a very useful method for studying the macroscopic properties of ferroelectric materials, but it requires a high-precision interatomic potential.<br/>This study used a machine learning deep neural network model to construct an interatomic potential energy function for the lead-free ferroelectric material. By incorporating first-principles calculation data into the training dataset, the accuracy of the interatomic potential was ensured with approximate quantum mechanical precision. The results show that molecular dynamics simulations based on the machine learning interatomic potential maintain good consistency with first-principles calculations, accurately predicting multiple basic properties of the material, such as atomic forces, energy, elastic tensor, and phonon dispersion. Additionally, the method demonstrates satisfactory performance in the domain walls, defects, and phase transition simulations. This approach holds promise for constructing interatomic potentials for other ferroelectric systems, contributing significantly to research in the field of ferroelectric materials.