Geometry-Driven Attention Model with 3D Molecular Features for Multi-Property Prediction of OLED Materials

As Organic Light-Emitting Diode (OLED) technology advances in applications such as high-end displays, medical devices, and VR/AR systems, the development of high-performance materials that improve energy efficiency and support environmental sustainability has become increasingly important. Traditional methods like Density Functional Theory (DFT) are accurate but computationally expensive, motivating the need for efficient prediction tools. This paper presents 3D-SAT-OLED, a deep learning model based on three-dimensional self-attention to predict key OLED performance parameters. The model integrates 3D molecular structures with a Gaussian kernel-based feature extractor (GPE) and spatial position encoding, while Dynamic Tanh (DyT) and Swish-Gated Linear Unit (SwiGLU) activations enhance stability and nonlinear modeling. Experiments show that 3D-SAT-OLED achieves state-of-the-art performance on QM9 and OLED QM properties data sets, significantly outperforming existing models across R², MAE, and other metrics, while demonstrating strong generalization and predictive accuracy.