A quantum-transformer hybrid architecture for polymer property prediction: Addressing data sparsity issues

Polymers have been used in various applications, and accurate prediction of polymer properties is important for their application and design. Although machine learning has demonstrated excellent performance, existing models still have limitations in dealing with complex nonlinear relationships and sparse datasets. This study proposes a novel solution - a PolyQT model that combines quantum neural networks (QNNs) with the Transformer architecture. The model aims to take advantage of quantum computing to enhance the modeling capability of complex nonlinear relationships while alleviating the data sparsity problem and improving the prediction accuracy of polymer features through its special structural design. Prediction experiments for six key features, namely ionization energy, dielectric constant, glass transition temperature, refractive index, crystallization trend and polymer density,show the significant advantages of the model: on the complete dataset, the R${}^2$ values of ionization energy, dielectric constant, glass transition temperature, refractive index reach and polymer density 0.85, 0.77, 0.85, 0.83 and 0.92, respectively, which are better than those of all the benchmark models; and the R${}^2$ value of crystallization trend is 0.27, which is not worse than that of most of the benchmark models. The R${}^2$ of the PolyQT model is always better than that of the classical Transformer under different data sparsity conditions, such as 40%, 60%, and 80%, indicating its superiority in sparse data processing. In addition, by comparing experiments with different numbers of quantum bits, we find that the model performs best with eight quantum bits, further exploring the critical role of the quantum mechanism in the model. Although quantum computing technology is still evolving, this study highlights the potential of quantum mechanics in predicting polymer properties, offers new insights for further understanding and optimizing these properties, and suggests promising directions for interdisciplinary research.