A generative model for the bidirectional design of quantum wells in semiconductor lasers using diffusion model

Abstract The traditional design of quantum wells (QWs) requires the coupled solution of structural, electronic and optical parameters, making brute-force simulation-based screening costly and limiting rapid iteration during device design. In this paper, a conditional diffusion approach is introduced to generate QW band-gain signatures directly from structural parameters. Based on a fine-tuned Stable Diffusion model, a design generative model of QWs with a physics-constrained hybrid loss function is achieved. A Contrastive Language-Image Pre-training (CLIP) module is further applied as an image-to-parameter retrieval component. Compared with a classical GAN network and an unfine-tuned diffusion model, the fine-tuned diffusion model achieves higher image-level fidelity and lower PL-wavelength error on held-out QW designs. The physical simulator, dataset construction, split protocol and computational cost are reported to define the evaluation conditions. The results suggest that the framework is best interpreted as a fast surrogate for screening and visualization within the sampled design domain, rather than as a replacement for the underlying physics solver.