MACHINE LEARNING ASSISTED PREDICTION FOR IONIC CONDUCTIVITY IN DOPED LLZO SOLID-STATE ELECTROLYTES USING FACILE DESCRIPTORS

• Main Author: Adhyatma, Abdurrahman
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/49775
• Institution: Institut Teknologi Bandung

Doped Li7La3Zr2O12 (LLZO) as solid-state electrolytes present itself as a possible solution for batteries with better safety requirements. A machine learning regressor is an effective method to peruse the search space for the best possible dopant combination, beginning with the prediction of ionic conductivity. Building a regressor starts with gathering data from published research that synthesized and characterized doped-LLZO. Subsequently, the data is explored and analyzed to gain insight concerning the data distribution and feature correlation. A regressor is built using polynomial regression and gradient boosted regression trees algorithms. Grid search cross-validation (CV) is used to optimized the polynomial regressor, while nested CV is used to optimize the gradient boosted regressor. After comparing the two finished models, gradient boosted regressor is the best model for ionic conductivity in LLZO. The model reached an R2 score of 0.739 on the training set, 0.630 on the test set, and 0.366 on 5-fold CV. Relationships between the features and the conductivity is examined using the model’s built-in feature importance. This research proved the potential of a machine learned regressor using facile descriptors to predict solid-state material properties. However, data with better quality and better engineered features are necessary to build a regressor with better predictive capabilities.