Abstract Efficient and accurate crystal property prediction models play a crucial role in materials design by accelerating the screening of potential candidate materials. However, most existing methods focus on single properties, neglecting the inherent correlations between material attributes and failing to meet the demand for multi-property evaluation in practical applications. Furthermore, current crystal representation methods based solely on atomic features and atomic distances overlook the spatial relationships within the crystal structure, leading to prediction biases. To address these challenges, we propose the DTMP, a directional-aware graph transformer model for multi-property prediction. Specifically, DTMP integrates directional features, enabling the model to consider the complex spatial relationships between atoms while preserving the periodicity of the crystal structure. Additionally, DTMP employs a dynamic multi-task optimization strategy that adaptively allocates attention weights based on task characteristics, mitigating prediction biases caused by differences in property dimensions. Experimental results on the Materials Project and JARVIS datasets demonstrate that DTMP achieves efficient and accurate multi-property predictions, offering a low-cost solution for material design and screening, and making a significant contribution to accelerating the development of new materials.
Gao et al. (Fri,) studied this question.