Prediction of orthorhombic lattice constants using machine learning

Abstract

A crystal structure is composed of a unit cell repeating itself to occupy a space, forming what is known as a lattice. This arrangement is dictated by the structure's lattice constants. Lattice constants are integral for investigation into the properties of crystal materials. However, current methods to determine such constants may be computationally exhaustive and time consuming. In this study, we utilized a random forest machine learning model to predict the lattice constants of orthorhombic crystal structures. This model was trained using the various materials' structural properties. To quantitatively evaluate the quality of our our model, we compared the model generated lattice constants with the experimental values, and obtained the following coefficients of determination (R²): 0.860, 0.825, and 0.826 for the a, b, and c constants respectively, which we found to be similar with previous studies. Moreover, we found the resultant mean squared error and mean absolute error for each lattice constant to be minimal, further supporting the overall performance of our model. Furthermore, to illustrate the weight of each property on the training of the model, we calculated the feature importance across the three random forest regressors. We found the key features to be unit cell volume, crystal system type, mean atomic number, and total atomic number.