MXenes, which are graphene-like two-dimensional transition metal carbides and nitrides, have tunable compositions and exhibit rich surface chemistry. This compositional flexibility has resulted in exquisitely tunable electronic, optical, and mechanical properties leading to the applications of MXenes in catalysis, electronics, and energy storage. The work function of MXenes is an important fundamental property that dictates the suitability of MXenes for these applications. We present a series of machine learning models to predict the work function of MXenes having generic compositions and containing surfaces terminated by O*, OH*, F*, and bare metal atoms. Our model uses the basic chemical properties of the elements constituting the MXene as features, and is trained on 275 data points from the Computational 2D Materials Database. Using 15 different features of the MXene as inputs, the neural network model predicts the work function of MXenes with a mean absolute error of 0.12 eV on the training data and 0.25 eV on the testing data. Our feature importance analysis indicates that properties of atoms terminating the MXene surface like their electronegativity, most strongly influence the work function. This sensitivity of the work function to the surface termination is also elucidated through experimental measurements on Ti3C2. We introduce reduced-order models comprising of ten-, eight-, and five-features to predict the work function. These reduced-order models exhibit easier transferability to new materials, while exhibiting a marginal increased mean average error. We demonstrate the transferability of these reduced order models to new materials, by predicting the work function of MXenes having surface terminations beyond the original training set, like Br*, Cl*, S*, N*, and NH*. Predicting electronic properties like the work function from the basic chemical properties of elements, paves the way towards rapidly identifying tailored MXenes having a targeted range of properties that are required for a specific application.
Read full abstract