Beyond Tolerance Factor: Using Deep Learning for Prediction Formability of ABX3 Perovskite Structures

Abstract

Deep learning (DL) is a modern powerful instrument for multiple purposes, including classification. In this study, this technique is applied to the task of perovskites formability. A commonly known perovskite dataset is used to try to make an instrument superior to the ‘classic’ geometric approach. The authors found that the resulting models allow the finding of inaccuracies in the data and can successfully forecast perovskite formability with an accuracy of over 98% for the best case.

Recent progress and future prospects on halide perovskite nanocrystals for optoelectronics and beyond

“…Note that a reasonable t and μ are necessary but not sufficient for the formation of a stable 3D ABX 3 perovskite. In particular, when the organic cations are non-spherical, the assumption that the ions are unpolarizable hard spheres, especially for the iodide anion may not be valid ( Fedorovskiy et al., 2021 ). In 2016, Travis et al.…”

Unraveling the Correlation between the Interface Structures and Tunable Magnetic Properties of La1–xSrxCoO3−δ/La1–xSrxMnO3−δ Bilayers Using Deep Learning Models

“…Recently, neural network-based deep learning (DL) models mapping structure to properties, e.g., total energy or structural stability, have received increasing attention as they exhibit remarkable flexibility and excellent scalability, and are computationally more efficient compared to brute force DFT calculations. − Therefore, by combining a structure search scheme with DL methods, atomic structures of interfaces and their energies (for a wide range of compositions) can be determined efficiently, and this combined scheme can remarkably accelerate the exploration of phase space and allow for screening of materials with fewer computing resources. − However, most neural network models proposed in the literature use a single neural network architecture, and this lack of model optimization, especially of neural network architecture, can induce systematic errors and lead to overfitting or underfitting issues. To mitigate this problem, we propose a framework (see Figure ) that combines an efficient, genetic algorithm-based interface structure scheme with a state-of-the-art neural architecture search method to design ML models by optimizing neural network architectures and their hyperparameters over a wide variety of neural network ensembles.…”