FerroAI: A Deep Learning Model for Predicting Phase Diagrams of Ferroelectric Materials

TL;DR

FerroAI is a deep learning model that predicts phase diagrams of ferroelectric materials by leveraging a large dataset mined from research articles, aiding in the design of advanced ferroelectric materials.

Contribution

The paper introduces FerroAI, a novel deep learning approach trained on NLP-mined data to accurately predict phase boundaries in ferroelectric materials, surpassing traditional methods.

Findings

Successfully predicts phase boundaries in doped BaTiO3.

Identifies a morphotropic phase boundary at x=0.3 in Zr/Hf co-doped BT.

Guides discovery of a new ferroelectric material with high dielectric constant.

Abstract

Composition-temperature phase diagrams are crucial for designing ferroelectric materials, however predicting them accurately remains challenging due to limited phase transformation data and the constraints of conventional methods. Here, we utilize natural language processing (NLP) to text-mine 41,597 research articles, compiling a dataset of 2,838 phase transformations across 846 ferroelectric materials. Leveraging this dataset, we develop FerroAI, a deep learning model for phase diagram prediction. FerroAI successfully predicts phase boundaries and transformations among different crystal symmetries in Ce/Zr co-doped BaTiO$3$ (BT)-$x$Ba${0.7}$Ca$_{0.3}$TiO$_3$ (BCT). It also identifies a morphotropic phase boundary in Zr/Hf co-doped BT-$x$BCT at $x = 0.3$, guiding the discovery of a new ferroelectric material with an experimentally measured dielectric constant of 9535. These results establish FerroAI as a powerful tool for phase diagram construction, guiding the design of high-performance ferroelectric materials.