Machine Learning Discovery of a New Descriptor for Topological Semimetal

The advent of expansive material databases provides an unprecedented opportunity to investigate predictive descriptors for emergent material properties. Particularly exciting possibility is to use expertly curated data to learn the descriptor as mechanism to bottle the human expert reasoning and intuition. For this, a reliable measurement based and expertly curated data and bench marking insight by expert researcher are critical. As a first step towards such program, we focus on the topological semi metal (TSM) among square-net materials as target property, inspired by the expert identified descriptor based on structural information: the tolerance factor [1]. We start by curating a dataset encompassing 12 primary features of 879 square-net materials, using experimental data whenever possible. We then use Dirichlet-based Gaussian process regression [2] using a specialized kernel [3] to reveal composite descriptors for square-net topological semimetals. The machine learned descriptors include a reproduction of the expert intuition (the t-ratio). Moreover, new descriptor that goes beyon the structural feature points to hypervalency as a critical chemical feature predicting TSM within square-net compounds. Our success with a carefully defined problem points to the “machine bottling human insight” approach as a promising approach for machine learning aided materials discovery.

[1] S. Klemenz, A. K. Hay, S. M. L. Teicher, A. Topp, J. Cano, and L. M. Schoop, Journal of the American Chemical Society 142, 6350 (2020).

[2] D. Milios, R. Camoriano, P. Michiardi, L. Rosasco, and M. Filippone, in Advances in Neural Information Processing Systems (2018) p. 11.

[3] F. Vivarelli and C. Williams, Advances in Neural Information Processing Systems 11 (1998).