DeepTMC: A Deep Learning Platform to Targeted Design Doped Transition Metal Compounds

Abstract Exploring the influence of impurity incorporation on the stability and conductivity of crystal structures is crucial for energy materials and devices, which have catapulted to the forefront of material design. A remarkable feature of doped materials is the diversity of their compositions, which enables multiple substitutions of different element types and numbers to produce thousands of possible compounds and a virtually near-infinite number of multicomponent structures. Harnessing the full potential of these materials necessitates a rapid exploration of the multi-dimensional chemical space toward desired functionalities. However, trial-and-error experiments and high-throughput calculations cannot systematically and comprehensively study doped structures. Here, we present an efficient, accurate, and extensible deep learning model as a case study to predict the stabilities and conductivities of doped transition metal nitrides (TMNs), TM oxides (TMOs), and TM nitride oxides (TMNOs). We design and establish three high-precision models: a classification model (Chull) that can predict the phase stability with an accuracy of 89.8%, a classification model (Cgap) that can identify metals/semiconductors with an accuracy of 83.4%, and a regression model (Rgap) that can predict the band gap of semiconductors with a mean absolute error (MAE) of 0.35 eV. By integrating these three models, we develop a visual interactive software (DeepTMC) to targeted design doped transition metal compounds (TMC) with good stability and suitable conductivity. Importantly, the proposed method is not only applicable to the study of TMC structures, but also can be expanded to include more inorganic/organic databases, replacing DFT in qualitative studies of the doping effects and assisting the targeted design of energy materials.