Evolutionary algorithms provide an efficient method to explore the global energy landscape of materials and identify low-energy minima but usually rely on density-functional theory codes to perform such calculations at high computational cost. To accelerate the structure search, it is desirable to utilize the data obtained during structural relaxations through the learning of a surrogate model. We present a machine learning approach for the formation energies of crystal structures. We explore two kernel-based learning algorithms, kernel ridge regression and support vector regression. The efficiency of machine learning approaches relies on suitable data representations that encode the relevant physical information about the crystal structures. We test several physically motivated structure representations. We show that machine of the energy landscape using partial radial distribution functions for the data representation predicts the formation energies of Li-Ge crystals with chemical accuracy and a small prediction error of 20~meV/atom across the composition and structure space. The accuracy demonstrates the potential of machine-learning models to reduce the computational cost needed to identify low-energy candidate structures and improve the performance of the genetic algorithm for structure predictions.
Back to Workshop III: Surrogate Models and Coarsening Techniques