Integrating the strengths of cVAE and cGAN into cAAE for advanced inverse design of colloidal quantum dots

Colloidal quantum dots (QDs) exhibit unique structures, which often result in distinctive optical properties such as emission and absorption spectra. However, QDs with different structures can sometimes show very similar emission and absorption spectra, making it difficult to inversely design their precise structural parameters from a given target emission and absorption spectra. To overcome this so-called one-to-many mapping problem, this paper introduces a novel deep-learning-based generative model for the inverse design of QDs. In particular, we implement three types of conditional generative models: the conditional generative adversarial network (cGAN), the conditional variational autoencoder (cVAE), and the conditional adversarial autoencoder (cAAE). Each model is designed and trained to predict possible layer thicknesses of QDs that can provide a given target emission and absorption spectra, thus providing possible multiple solutions rather than a single deterministic outcome. This multi-solution approach not only increases the flexibility in QD structure design, but also enhances the accuracy and efficiency of the predictive process. According to calculation results, the cAAE stands out by effectively combining the strengths of both cGAN and cVAE. This integration allows cAAE to produce a more diverse and accurate inversely designed structures of InP/ZnSe/ZnS QDs.