<p>Machine learning (ML) is playing an increasingly important role in optimizing photonic device modeling. AlGaN-based deep ultraviolet (DUV) laser diodes (LDs) have emerged as a critical class of photonic devices, with wide-ranging applications from disinfection to optical communication. While ML models offer a powerful solution for their design, the effectiveness of these models depends strongly on the quality and volume of the available data. For DUV LDs, obtaining high-quality datasets is challenging due to the complexity, high cost, and significant computational demands of both numerical simulations and device fabrication. To address these challenges, a self-augmentation strategy has been proposed that leverages four generative models: Conditional Generative Adversarial Networks (CGAN), Tabular Variational Autoencoder (TVAE), Bayesian Networks, and Gaussian Copula to synthesize data from limited samples. Optical power is used as the primary evaluation parameter to assess the performance of model. We have evaluated the fidelity and utility of the generated dataset using the mean absolute correlation coefficient (r) and mean absolute error (MAE), respectively, and assess robustness across multiple data augmentation scenarios. Among the models tested, CGAN achieved the highest fidelity of the data with an r value of 0.0202, closely matching the actual data (r = 0.0266). While TVAE provided the best utility of the model, reducing MAE to 0.73, an improvement of 8% over the actual dataset, and exhibiting stable performance under all augmentation scenarios. A five-layer deep neural network (DNN) is employed for model utility on hybrid dataset that is the combination of both actual and augmented dataset. We believe that the proposed framework offers a scalable and adaptable approach to enable accelerated design of next-generation DUV LDs while supporting in-the-loop data augmentation to address the data-scarcity challenge. This work constitutes the first comprehensive investigation of its kind in this field.</p>