Prediction of Li-ion conductivity in Ca and Si co-doped LiZr2(PO4)3 using a denoising autoencoder for experimental data

• Published in: APL Materials
• Main Authors: Yumika Yokoyama, Shuto Noguchi, Kazuki Ishikawa, Naoto Tanibata, Hayami Takeda, Masanobu Nakayama, Ryo Kobayashi, Masayuki Karasuyama
• Format: Article
• Language: English
• Published: AIP Publishing LLC 2024-11-01
• Online Access: http://dx.doi.org/10.1063/5.0231411

All-solid-state batteries composed of inorganic materials are in high demand as power sources for electric vehicles owing to their improved safety, energy density, and overall lifespan. However, the low ionic conductivity of inorganic solid electrolytes has limited the performance and adoption of inorganic all-solid-state batteries. The solid electrolyte LiZr2(PO4)3 has attracted attention owing to its high Li-ion conductivity. The ionic conductivity of LiZr2(PO4)3 changes with the crystalline phase obtained, which varies based on composition control through elemental substitution and process conditions such as sintering temperature. Traditionally, optimizing such parameters and understanding their relationship to physical properties have relied on researcher experience and intuition. However, a recent use of a materials informatics approach utilizing machine learning shows promise for more efficient property optimization. This study proposes a deep learning model to correlate powder X-ray diffraction (XRD) profiles with the activation energy (Ea) for Li-ion conduction, thereby enhancing the interpretability of the measurement data. XRD profiles, which contain information on crystal structure, lattice strain, and particle size, were used as-is (i.e., without preprocessing) in the deep learning model. An attention mechanism was introduced to the deep learning model that focuses on XRD crystal-structure information and visualization of important factors embedded in the XRD profiles. The highlighted areas in the output of this model successfully predict LiZr2(PO4)3 phases with low Ea (high Li conductivity) and high Ea (low Li conductivity). Moving forward, this deep learning model can offer new insights to materials researchers, potentially contributing to the discovery of new solid electrolyte materials.