Luminescence properties of fluoride laser crystals: a joint prediction model based on universal atomic embedded graph neural networks and transfer learning.

Predicting the luminescence properties of fluoride laser crystals in the 3-5 µm mid-infrared bands is challenging due to data scarcity and inefficient empirical methods. To address this, we propose a joint prediction model (UAEGNN-TL) that combines a Universal Atomic Embedded Graph Neural Networks with Transfer Learning. The UAEGNN component utilizes an enhanced universal atomic embedding strategy to represent the intrinsic features of constituent atoms within a crystal structure. This is synergistically combined with a transfer learning (TL) framework, which leverages knowledge from source tasks to mitigate the data scarcity limitation in the target task. Our results demonstrate that the UAEGNN architecture effectively captures the complex structure-property relationships governed by atomic interactions in fluoride crystals. Its predictive accuracy significantly outperforms mainstream models such as Crystal Graph Convolutional Neural Network (CGCNN), achieving a Mean Absolute Error (MAE) of 0.065 with a reduction of nearly 39% compared to the CGCNN model. Its coefficient of determination (