Prediction of stable perovskite quantum dots (QDs) using machine learning (ML) algorithms

• Main Author: Cheong, Ivan
• Other Authors: Martial Duchamp
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Materials
• Online Access: https://hdl.handle.net/10356/156298
• Institution: Nanyang Technological University

Perovskites are semiconducting material with many attractive physical and chemical properties such as electronic conductivity, ions mobility through the crystal lattice, photocatalytic, thermoelectric, and dielectric properties. However, they have not been widely studied, with applications limited to solar cells and catalytic activity, the full extent of possible applications are yet to be realised. At the nano regime the properties are often different and can be controlled, this is due to quantum confinement and larger surface to volume ratio, which can be adjusted by controlling the size, subsequently able to use for various application. With possible enhanced properties, having to synthesize and characterise the properties to understand if the application is feasible could be costly and time consuming, requiring equipment and expertise as well. Application of machine learning have seen wide adoption, researched on in many industries with many positive results. Use of this new technology in the materials industry is still in its infancy, with many studies incorporating it as a supplement instead of substitution. In this study, we aim to demonstrate that machine learning can assist in material discovery, accelerating the research process, substituting the laborious task of physically synthesizing and characterising an educated guess. With a focus on the promising optical properties of perovskites understanding its photoluminescence quantum yield (PLQY) and the possibility of implementation in photovoltaics applications The general strategy employed is to create a reusable database of as many elemental combinations of perovskites and their available properties from published literature, containing both simulations and experimental results. Extract the required data to train various machine learning models, use the model with the best fit to predict the possibility of synthesis and the estimated values of the material properties. After attaining the predictions, synthesize the hypothetical perovskite and measure the actual properties to validate the accuracy of prediction. Then adding the data into our database for more predictions, accelerating and simplifying material discovery. We created a database of over 400 perovskite’s experimental data, containing their composition, synthesis methods, passivation agents, dimensionality, nano sizes, structures and properties. We extracted a portion from the database, containing factors contributing to the photoluminescence quantum yield (PLQY), adjusted the sample data points for a better training of machine learning models through data cleaning, Box-Cox transformations and splitting into training or testing sets. Subsequently, trailed various machine learning models on the data, evaluated their suitability and inferred any correlations or relationships that may appear. We demonstrated that there is a relationship of the composition and ionic radius with its photoluminescence quantum yield (PLQY), however the highest R2 scores of 0.49 produced by the XGBoosting regression model were not good enough for running a useable prediction. Falling back to literature of known understanding of the photoluminescence property, suggested that the nanocrystal’s surface ligands interface has a substantial impact on optical properties. Hence finding a method to incorporate the string data type of synthesis methods and passivation agents to machine learning models who typically prefer numerical data types could improve the prediction abilities.