ICMCTF 2022 Session TS4-ThA: Big Data, Machine Learning, Artificial Intelligence and High-Throughput Methods

Exploratory theoretical predictions in uncharted structural and compositional space are integral to materials discoveries. A more recent addition to the family of laminated metal borides, so called MAB phases, is new types of chemically ordered quaternary borides, i-MAB (M´_4/3M´´_2/3AlB₂) and o-MAB (M´₄M´´SiB₂), in which the M-atoms are in-plane and out-of-plane chemically ordered, respectively. Both types of phases have been identified in high-throughput simulations followed by experimental synthesis, and they can be chemically exfoliated into 2D sheets, and selectively prepared in multilayer form, as delaminated single-layer sheets in colloidal suspension, or as additive-free filtered films. For example, Ti₄MoSiB₂o-MAB can be used to derive 2D TiO_xCl_y of high yield. (Mo_2/3Y_1/3)₂AlB₂ and (Mo_2/3Sc_1/3)₂AlB₂ i-MAB can be used for realization of so called boridene in the form of single-layer 2D sheets with ordered metal vacancies, Mo_4/3B_2-xT_z(T_z = -F, -O, -OH). The present talk will summarize the results to date of our predictive theoretical approach that have led to 2D materials synthesis from 3D quaternary metal borides, the mechanisms behind the realization of these 3D/2D materials, and evaluation of selected properties.

Refractory nitrides (TiN, Hf N , NbN etc.) nitride alloys (Ti_1-x Al_xN, etc.) or high-entropy alloys, such as TiZrHfTa are materials with high industrial relevance for hard coatings o f cutting tools [1] even for superconducting and plasmonic-based devices. M ajor objectives for their performance are the high-temperature the rmodynamic, dynamic and elastic properties. A recently developed combination of quantum mechanical calculations with machine-learning interatomic potentials (MLIP) [2], is utilized to calculate high temperature properties with high accuracy. On-the-fly training of moment tensor potentials allows us to perform the calculations with more than two orders less computational effort than using state-of-the-art ab initio molecular dynamics simulations. The calculated elastic constants are used to simulate surface acoustic waves and Brillouin light scattering (BLS) spectra. The results are compared with experiments. Furthermore, we investigate high-temperature bcc phase of titanium and predict very weak temperature dependence of its elastic moduli [3], called Elinvar effect, similar to the behavior observed for the so-called GUM metals . The effect in bcc-Ti is intrinsic and therefore unique.

[1] See, for example, F. Tasnádi et al., Phys. Rev. B 85, 144112 (2012); F. Tasnádi et al., Appl. Phys. Lett. 97, 231902 (2010); D. Holec et al. Phys. Rev. B 90, 184106 (2014); F. Tasnádi et al., Mater. Des. 114, 484 (2017); H. Huang et al., Adv. Mater. 5, 1701678 (2017). [2] I. S. Novikov et al., Mach. Learn.: Sci. Technol. 2 025002 (2021). [3] A. Shapeev et al., New. J. Phys. 22, 113005 (2020).

We present a systematic first-principles search for thermal insulators in materials space which covers hundreds of compounds, five lattice types and seven space groups, including simple rocksalt and zinc blende structures, up to complex perovskites. Using the high-throughput framework FHI-vibes [1] and a recently developed measure for the strength of anharmonicity [2], we identify 120 candidate materials with potential for low thermal conductivity at room temperature. We investigate the 60 most promising candidates with the ab initio Green Kubo method (aiGK) [3], enabling data-driven extraction of design principles for bulk materials with low thermal conductivity. The aiGK method provides an accurate framework to obtain thermal conductivities for materials, in particular strongly anharmonic ones such as thermal barrier coating ceramics like zirconia, since all anharmonic effects responsible for low thermal conductivity are included.We subsequently demonstrate how the first principles calculations can be complemented with the help of machine learning potentials to remove the computational bottleneck. For this task, we use message passing neural networks, a class of models that can accommodate implicit long-range interactions as well as directional information [4]. We present a systematic account of their performance for calculating the thermal conductivity of solid semiconductors and insulators and discuss implications for high-thoughput heat transport simulations and the discovery of novel thermal insulators.

[1] F. Knoop et al., J. Open Source Softw. 5, 2671 (2020)
[2] F. Knoop et al., Phys. Rev. Mater. 4, 083809 (2020)
[3] C. Carbogno, R. Ramprasad, and M. Scheffler, Phys. Rev. 118, 175901 (2017)
[4] K.T. Schütt et al., J. Chem. Phys. 148 241722 (2018)

Superhard materials with a Vickers hardness >40 GPa are essential in applications ranging from manufacturing to energy production. Finding new superhard materials has traditionally been guided by empirical design rules derived from classically known materials. However, the ability to quantitatively predict hardness remains a significant barrier in materials design. To address this challenge, we constructed an ensemble machine-learning model capable of directly predicting load-dependent hardness. The predictive power of our model was validated on eight unmeasured metal disilicides and a hold-out set of superhard materials. The trained model was then used to screen compounds in Pearson’s Crystal Data (PCD) set and combined with our recently developed machine-learning phase diagram tool to suggest previously unreported superhard compounds. Finally, industrial materials often experience tremendous heat during application; thus, we are building a method for predicting hardness at elevated temperatures.

Based on Machine Learning predictions, isovalue hulls and Pareto-optimal domains can be determined inside the composition space and used to identify the composition sets that show the best property compromises.