Machine Learning-Based Prediction of Elastic Properties Using Reduced Datasets of Accurate Calculations Results

In this paper, the applicability of machine learning for predicting the elastic properties of binary and ternary bcc Ti and Zr disordered alloys with 34 different doping elements is explored. The original dataset contained 3 independent elastic constants, bulk moduli, shear moduli, and Young’s moduli of 1642 compositions calculated using the EMTO-CPA method and PAW-SQS calculation results for 62 compositions. The architecture of the system is made as a pipeline of a pair of predicting blocks. The first one took as the input a set of descriptors of the qualitative and quantitative compositions of alloys and approximated the EMTO-CPA data, and the second one took predictions of the first model and trained on the results of the PAW-SQS calculations. The main idea of such architecture is to achieve prediction accuracy at the PAW-SQS level, while reducing the resource intensity for obtaining the training set by a multiple of the ratio of the training subsets sizes corresponding to the two used calculation methods (EMTO-CPA/PAW-SQS). As a result, model building and testing methods accounting for the lack of accurate training data on the mechanical properties of alloys (PAW-SQS), balanced out by using predictions of inaccurate resource-effective first-principle calculations (EMTO-CPA), are demonstrated.