Prediction of Spectral Parameters in Er3+, Dy3+ and Nd3+ Doped Oxide Glasses via cGAN-Enhanced Hybrid Modeling

The Judd–Ofelt (J–O) intensity parameters and oscillator strengths are key to understanding the optical transition properties of rare-earth-doped glasses. However, the scarcity of experimental samples and the complex nonlinear relationship between composition and spectral properties pose significant challenges to accurate predictions. To address this, we propose a generalizable framework that integrates conditional generative adversarial network (cGAN)-based data augmentation with an attention-embedded artificial neural network (ANN)–support vector regression (SVR) hybrid model. The cGAN generates physically plausible virtual samples to enrich data distribution and enhance generalization in sparse compositional regions. The attention mechanism in the ANN identifies critical compositional features, which are then leveraged by SVR for robust regression of parameter trends. The framework demonstrates high predictive accuracy for Er3+-doped glasses, achieving R2 values above 0.93 for Ω2, Ω4, and Ω6, and exhibits strong generalization performance on independent Dy3+- and Nd3+-doped datasets without task-specific retraining, confirming its practical applicability across multiple rare-earth ions. The model maintains consistency across diverse glass host systems (tellurite, borate, phosphate, silicate/germanate, heavy-metal oxide), and the attention analysis reveals feature importance aligned with established glass chemistry principles. Demonstrated on Er3+, Dy3+, and Nd3+, with potential for a broader range of rare-earth ions through transfer learning and future dataset extensions, this approach offers a data-driven, physics-informed tool for the targeted design of rare-earth optical materials in next-generation optical sensors.