Screening outstanding mechanical properties and low lattice thermal conductivity using global attention graph neural network

Abstract

Mechanical and thermal properties of materials are extremely important for various engineering and scientific fields such as energy conversion and energy storage. However, the characterization of these properties via high throughput screening at the quantum level, although highly accurate, is inefficient and very time- and resource-consuming. In contrast, prediction at the classical level is highly efficient but less accurate. We deploy scalable global attention graph neural network for accurate prediction of mechanical properties which bridge the gap between the accuracy at the quantum level and efficiency at the classical level. Using 10,158 elastic constants as training data, we trained the models on 5 mechanical properties, namely bulk modulus, shear modulus, Young's modulus, Poisson's ratio, and hardness. With the trained model, we predicted 775,947 data in search of materials with ultrahigh hardness. We further verify the recommended ultrahigh hardness materials by high precision first principles calculations, and we finally identify 20 structures with extreme hardness close to diamond, the hardest material in nature. Among those, two super hard materials are completely new and have not been reported in literature so far. We further recommend potential materials from bulk modulus prediction to search low lattice thermal conductivity, and we verify the thermal conductivity of 338 structures with first principles. Our results demonstrate that one can find materials with extreme mechanical properties recommended by graph neural network and low thermal conductivity material from bulk modulus prediction with minimal first principles calculations of the structures (only 0.04%) in the large-scale materials pool.

Keywords

• Graph neural network
• Machine learning
• Mechanical properties
• Ultrahigh hardness
• Lattice thermal conductivity
• DFT calculations