Prediction of Mxene 2D Material's Electronic Property Using Machine Learning Tools

MXene, a family of 2D materials, offers excellent electronic and mechanic properties that could be used for multiple application, such as heterogeneous catalysis. However, little is known about how to accurately tune its properties for direct industrial application. Herein, a series of machine learning models are presented to predict the work function of MXenes, one of the key electronic properties for evaluating its performance as catalyst. Machine learning is a data analysis method capable of discovering complex patterns. These models are less computationally intensive and cheaper to predict properties, compared to first principles methods such as the density functional theory (DFT) widely used in academia. The model is trained from the 275 MXene datapoints from Computational 2D Materials Database (C2DB). Basic statistical analyses are conducted to evaluate the distribution and anomalies. Diverse machine learning tools were then utilised, such as compressed sensing method, AUTOFEAT feature generation method, and ensemble method. Some models exhibit a high accuracy, such as testing RMSE of 0.264 eV for kernel ridge regression at AUTOFEAT step 3. The results also exhibit that electronegativity of termination element is one of the dominant factors for predicting work function value. These models and insights can help experimentalists fasten their workflow and design a more efficient and versatile catalyst by quickly identify tailored MXenes of interest. For improvement, further research could be taken place, such as investigating how properties vary under different pressure and temperature and selecting more primary properties related to surface dipole moment for training of the machine leaning model.