Intelligent design and discovery of novel 2D TMD alloys with machine learning

• Main Author: Tan, Nicholas
• Other Authors: Liu Zheng
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Materials
• Online Access: https://hdl.handle.net/10356/165736
• Institution: Nanyang Technological University

2D transition metal dichalcogenides (TMDs) are a novel class of nanomaterials with interesting properties due to their confined dimensions, distinct from their bulk counterpart. This enables fine-tuning of electronic and optical properties that can be utilized in various applications. However, the discovery of new nanomaterials with suitable properties is time-consuming and requires significant research effort. The purpose of this project aims to reduce the discovery time of new TMDs using machine learning techniques and predict important properties of these materials to aid in the synthesis of novel TMDs. Important properties and features of 2D TMDs were first collected from online databases before undergoing data cleaning, which were then used for model selection and training. The models utilised were extreme gradient boosting (XGB), multilayer perceptron (MLP), and support vector machine (SVM). Continuous improvements of the models were also done through feature engineering by looking at the feature importance of each model. Model evaluation was then done by looking at its R2, r, and MSE values. This project successfully trained five models to predict the target properties of Heat of Formation (HOF), Energy Above Convex Hull (EACH), Bandgap (BG), Direct Bandgap (DBG), and Magnetic Moment (MM). The MLP model was shown to be the best model for HOF, BG, and DBG, while the XGB model was the best model for EACH and MM, based on their R2 values. Utilising both models, we successfully predicted the properties of three TMD/TMD-alloys which were not present in the dataset. Specifically, the MLP model was used for BG prediction, and the XGB model was used for HOF prediction. In summary, successful machine learning models were created to predict five important targeted properties of TMD/TMD-alloys materials. The models were able to predict the HOF and BG of materials not found in the dataset. However, the model accuracy could be further improved by increasing the dataset, choosing more features, or improving on the chosen models.