Exploration of novel boron nitride polymorphs: High-throughput screening combined with multi-task orbital crystal graph convolutional neural network (MT-OCGCN)

In response to the limitations of traditional material property prediction methods, we have developed a Multi-task Orbital Crystal Graph Convolutional Neural Network (MT-OCGCN). This innovative model allows for sharing parameters and computational resources, significantly improving prediction efficiency and accuracy. Experimental comparisons show that the model can achieve multi-task prediction with high accuracy for both strong and weak correlation properties, this offers an efficient approach for computational materials screening and design. Combined with high-throughput screening has resulted in the discovery of 13 new boron nitride polymorphs in the P 2 1 / c phase, among which 7 are characterized as wide direct band gap semiconductors. Furthermore, we have comprehensively analyzed the mechanical, thermal, and electronic properties of these new boron nitride polymorphs. It can be asserted that MT-OCGCN has the capability to precisely forecast the physical properties of novel structures that are absent from current databases. • A new deep learning model called MT-OCGCN was introduced for multi-task prediction. • The MT-OCGCN model can achieve multi-task prediction with high accuracy for both strong and weak correlation properties. • 13 new boron nitride polymorphs and their mechanical, thermal, and electronic properties were identified. • 7 wide direct band gap semiconductors were discovered.