FHI-vibes: A Toolkit for Finding and Analyzing Thermal Insulators

Identifying and designing materials with ultra-low thermal conductivities κ is one of the major challenges in the development of novel, more efficient thermoelectric “waste heat recovery” devices [1], which are pivotal for establishing a sustainable, green energy economy. The in-house developed ab initio Green-Kubo (aiGK) method is ideally suited to guide and accelerate the search for such materials, since –in contrast to perturbative methods– it captures the strong anharmonic effects determining the properties of thermalinsulators [2]. Due to the elevated computational cost of aiGK calculations, however, acomprehensive high-throughput search of materials space seems prohibitive, at first. We overcome this hurdle with a newly developed hierarchical, high-throughput framework that builds on the combination of the Atomic Simulation Environment ASE [3] with FireWorks [4], the AFLOW Library of Crystallographic Prototypes [5], Phonopy [6], and FHI-aims [7]. In this workflow, we systematically explore material space using a recently developed, accelerated geometry relaxation scheme that ensures the preservation of thetarget crystalline prototype [8]. The degree of anharmonicity is then quantified via anewly developed descriptor based on the statistical comparison of the energy and forces obtained in the harmonic approximation to the fully anharmonic ones obtained via ab initio MD and/or statistical sampling. This allows us to disregard materials with large κ in the early stage of the high-throughput search and thus efficiently focus on actual thermal insulators instead. For these materials with strong anharmonicity and thus low κ, the thermal conductivity is then computed with the aiGK method. The performed search that covers thousands of materials reveals that strongly anharmonic compounds with low κ are much more common than originally thought, although for different reasons. We shed light on these qualitative mechanisms determining the low thermal conductivity using representative examples from multiple material classes. For instance, we discuss the role of strong ionic interactions in Cu halides (e.g. CuCl, CuI); the influence of van-der-Waals interactions in layered chalcogenides like SnSe; and the role of highly mobile species in inorganic halide perovskites such as CsPbBr3 and CsPbI3. Eventually, we discuss how the developed approach can be further accelerated and guidedby data-mining the millions of performed first-principles calculations with artificial intelligence and active learning techniques such as SISSO [9].

References:

1. G. J. Snyder and E. S. Toberer, Nature Materials 7, 105 (2008).
2. C. Carbogno, R. Ramprasad, and M. Scheffler, Phys. Rev. Lett. 118, 175901 (2017).
3. A. S. Larsen, et al. J. Phys. Condens. Matter. 29, 273002 (2017).
4. A. Jain, et al. Concurrency Computat. Pract. Exper. 27, 5037 (2015).
5. M. Mehl, et al. Comput. Mater. Sci. 136, S1 (2017).
6. A. Togo and I. Tanaka, Scr. Mater. 108, 1 (2015).
7. V. Blum, et al. Comp. Phys. Comm. 180, 2175 (2009).
8. M. -O. Lenz, et al. to be submitted to npj Computational Materials.
9. R. Ouyang, et al. Phys. Rev. Mat. 2, 083802 (2018).