Artificial Intelligence Simulation and Design of Energy Materials with Targeted Properties

Abstract

The discovery of new energy materials is fundamental to addressing numerous technological challenges. At the forefront of discoverable materials is the perovskite crystal structure, which is an emerging multifunctional material class with several attractive properties that are utilized in many engineering applications. They include high ionic conductivity for solar cells, good dielectric response for piezoelectric devices, and strong catalytic activity for batteries or hydrogen production. Considering both the chemical flexibility of perovskite stoichiometries and the possibility of polymorphism, the estimated number of perovskite compounds could potentially exceed ten million. Conventional discovery techniques often involve Edisonian-based synthetic chemistries and/or first-principles quantum mechanical calculations. Although such techniques have achieved substantial successes in the past, their application can be difficult, unpractical, uneconomical and computationally expensive for complex systems and extremely large chemical search spaces. In this regard, the current thesis explores Artificial Intelligence (AI) as potentially a more reliable, inexpensive and rapid alternative over conventional techniques. In the process of making such discovery, two design schemes are sequentially addressed, namely forward design simulation and inverse design simulation. The forward design is focused on simulating the structure-to-property relationship by accurately mapping perovskite materials to their deterministic target properties. Conversely, the inverse design simulation is aimed at simulating the property-to-structure relationship by discovering novel perovskite materials that possess deterministic target properties of interest. In the present research, the forward design is addressed by developing a hybridized Deep Learning (DL) framework comprised of a two-dimensional convolutional neural network (Conv2D) model and a support vector machine (SVM) model. The hybridized Conv2D-SVM approach is demonstrated to out-perform periodic benchmark representations in the Coulomb matrix, Ewald-sum matrix, and Sine matrix by about 70%, 75% and 66% on stability energy, formation energy and bandgap targets, respectively. In addressing the inverse design, three Deep Generative Modelling (DGM) pipelines are designed: (1) Target-Learning Variational Autoencoder (TL-VAE) model; (2) Evolutionary Variational Autoencoder for Perovskite Discovery (EVAPD) model; and (3) Lattice-Constrained Materials Generative Model (LCMGM). In total, 265 new perovskite materials are designed using the developed DGM pipelines, of which about 60% are newly discovered (i.e. unique and novel) chemical compositions, while the remainders are polymorphs of already known chemical compositions. The new materials are validated using Density Functional Theory (DFT) technique and are openly archived in materials database repositories. Overall, the current research demonstrates efficient AI innovative pathways for advancing the rapid search for perovskite energy materials. The newly discovered perovskites are the subject of ongoing follow-up research on their synthesization, characterization and testing.