Multiscale correlation among composition, structure, and spectroscopic properties of laser glass

Laser glass serves as the key gain medium in high-energy laser fusion devices and fiber-optic communication systems. However, its development remains largely reliant on empirical trial-and-error methods, leading to prolonged research cycles and high costs. The multiscale correlation among composition, structure, and spectroscopic properties of laser glass has not yet been fully established. Although machine learning has been applied to target-oriented material development, its application in the study of laser glass spectral properties faces several significant challenges: (1) the lack of a large, consistent, homogeneous dataset of spectral data for training, (2) the difficulty of establishing an extrapolable and interpretable model that links glass composition to spectral properties, and (3) the complexity of glass structural characterization, particularly the local environment of rare-earth ions, which is more intricate than in crystals.In this work, the spectral database of Nd3+-doped Na2O-CaO-SiO2, K2O-BaO-P2O5, Na2O-BaO-P2O5, Al2O3-K2O-P2O5, and Al2O3-K2O-BaO-P2O5 glass systems was constructed using the phase diagram approach. Molecular dynamics (MD) simulations were performed with the LAMMPS software, yielding 19650 structural descriptors that closely matched experimental diffraction data. To minimize redundancy, low coefficient of variation screening and Pearson correlation analysis were applied, resulting in the selection of 12 key structural descriptors. For model development, two ensemble learning algorithms, random forest (RF) and extreme gradient boosting (XGBoost), were employed, using the 12 descriptors as inputs and the luminescent parameters from the phase diagram database as outputs. The predictive models for six spectral properties achieved R2 values exceeding 0.97, with XGBoost outperforming RF. Variable importance analysis revealed that alkaline earth metals, polymerization degree, and Nd-Nd clustering significantly influenced the refractive index, fluorescence branching ratio, and effective linewidth of laser glasses, respectively. For radiative lifetime and emission cross-section, the coordination competition between Nd and Si dominated in silicate systems, while in phosphate systems, Nd-Nd clustering exerted a greater effect. The model is expected to guide the design of high-emission cross-section Nd3+ laser glasses for high-power fusion devices, as well as provide theoretical insights into the thermal-optical performance balancing mechanisms and the regulation of nonlinear coefficients in laser glasses. For instance, appropriately increasing the content of P2O5 and K2O in phosphate laser glass can reduce the Nd-Nd clustering effect parameter (NNd-Nd), thereby enhancing the emission cross-section. The proposed framework integrates a phase diagram-based spectral database, MD analysis, and machine learning to link glass composition and microscopic structure with radiative properties, which provides insights for the design mechanisms and possible optimization pathways of commercial laser glasses such as LHG-8, NAP2, and LG-770. With future expansion of the database, this quantitative tool can be extended to other glass systems, facilitating the development of next-generation laser glass materials.