Exploring Materials Space with Machine Learning

Electronic structure calculations are too computationally expensive to thoroughly explore the composition space of any system. We use a surrogate model approach that constructs an interatomic potential from a small training set of electronic structure calculations. The surrogate model, called the Moment Tensor Potential (MTP)[1], automates and optimizes the creation of the training set for the interatomic potentials. The potential is then used to explore materials space and predict stable structures including those with new geometries not contained in the training set[2]. We use MTP to study the six promising candidates from a recent high-throughput search for ternary superalloys [3], namely MnNiSb, NiSbTi, NiSbSi, HfNiSi, CoTaV, and CoNbV. We analyze the phase diagram of each ternary system using a pool of 1.2 million structures to show the accuracy of these machine learned surfaces and predict new stable phases at a fraction of the computational time.
[1] Shapeev, A. V. (2016). 14(3), 1153-1173.
[2] Gubaev, K., Podryabinkin, E. V., Hart, G. L., & Shapeev, A. V. (2019). Computational Materials Science, 156, 148-156.
[3] Nyshadham, C., Oses, C., Hansen, J. E., Takeuchi, I., Curtarolo, S., & Hart, G. L. (2017). 122, 438-447.