Crystal-graph attention neural networks have emerged recently as remarkable tools for the prediction of thermodynamic stability. The efficacy of their learning capabilities and their reliability is however subject to the quantity and quality of the data they are fed. Previous networks exhibit strong biases due to the inhomogeneity of the training data. Here a high-quality dataset is engineered to provide a better balance across chemical and crystal-symmetry space. Crystal-graph neural networks trained with this dataset show unprecedented generalization accuracy. Such networks are applied to perform machine-learning-assisted high-throughput searches of stable materials, spanning 1 billion candidates. In this way, the number of vertices of the global T = 0 K phase diagram is increased by 30% and find more than ≈150 000 compounds with a distance to the convex hull of stability of less than 50 meV atom-1 . The discovered materials are then accessed for applications, identifying compounds with extreme values of a few properties, such as superconductivity, superhardness, and giant gap-deformation potentials.

Heusler compounds attract a great deal of attention from researchers thanks to a wealth of interesting properties, among which is superconductivity. Here we perform an extensive study of the superconducting and elastic properties of the cubic (full-)Heusler family using a mixture of ab initio methods, as well as interpretable and predictive machine-learning models. By analyzing the statistical distributions of these properties and comparing them to anti-perovskites, we recognize universal behaviors that should be common to all conventional superconductors while others turn out to be specific to the material family. In total, we discover a total of eight hypothetical materials with critical temperatures above 10 K to be compared with the current record of T c = 4.7 K in this family. Furthermore, we expect most of these materials to be highly ductile, making them potential candidates for the manufacture of wires and tapes for superconducting magnets.

Graph neural networks trained on large crystal structure databases are extremely effective in replacing ab initio calculations in the discovery and characterization of materials. However, crystal structure datasets comprising millions of materials exist only for the Perdew-Burke-Ernzerhof (PBE) functional. In this work, we investigate the effectiveness of transfer learning to extend these models to other density functionals. We show that pre-training significantly reduces the size of the dataset required to achieve chemical accuracy and beyond. We also analyze in detail the relationship between the transfer-learning performance and the size of the datasets used for the initial training of the model and transfer learning. We confirm a linear dependence of the error on the size of the datasets on a log-log scale, with a similar slope for both training and the pre-training datasets. This shows that further increasing the size of the pre-training dataset, i.e. performing additional calculations with a low-cost functional, is also effective, through transfer learning, in improving machine-learning predictions with the quality of a more accurate, and possibly computationally more involved functional. Lastly, we compare the efficacy of interproperty and intraproperty transfer learning.

We present a comprehensive theoretical study of conventional superconductivity in cubic antiperovskites materials with composition XYZ 3 where X and Z are metals, and Y is H, B, C, N, O, and P. Our starting point are electron–phonon calculations for 397 materials performed with density-functional perturbation theory. While 43% of the materials are dynamically unstable, we discovered 16 compounds close to thermodynamic stability and with T c higher than 5 K. Using these results to train interpretable machine-learning models, leads us to predict a further 57 (thermodynamically unstable) materials with superconducting transition temperatures above 5 K, reaching a maximum of 17.8 K for PtHBe 3 . Furthermore, the models give us an understanding of the mechanism of superconductivity in antiperovskites. The combination of traditional approaches with interpretable machine learning turns out to be a very efficient methodology to study and systematize whole classes of materials and is easily extendable to other families of compounds or physical properties.

Crystal-graph attention networks have emerged recently as remarkable tools for the prediction of thermodynamic stability and materials properties from unrelaxed crystal structures. Previous networks trained on two million materials exhibited, however, strong biases originating from underrepresented chemical elements and structural prototypes in the available data. We tackled this issue computing additional data to provide better balance across both chemical and crystal-symmetry space. Crystal-graph networks trained with this new data show unprecedented generalization accuracy, and allow for reliable, accelerated exploration of the whole space of inorganic compounds. We applied this universal network to perform machine-learning assisted high-throughput materials searches including 2500 binary and ternary structure prototypes and spanning about 1 billion compounds. After validation using density-functional theory, we uncover in total 19512 additional materials on the convex hull of thermodynamic stability and ~150000 compounds with a distance of less than 50 meV/atom from the hull. Combining again machine learning and ab-initio methods, we finally evaluate the discovered materials for applications as superconductors, superhard materials, and we look for candidates with large gap deformation potentials, finding several compounds with extreme values of these properties.