<p>The development of predictive tools based on machine learning (ML) and deep neural networks is increasingly pursued for materials discovery across various technologies, including thin films and optoelectronic devices. Although structure property relationships can be precisely determined using quantum–mechanical approaches, first-principles calculations but they are computationally intensive, which restricts their applicability in screening large numbers of candidate materials. Hence, in this work, we employ convolutional neural networks (CNNs) to construct a predictive model for the structural properties of double perovskite materials with the general formula A₂BB′O₆, a class characterized by an extremely large design space. We demonstrate that a carefully designed hierarchical ML framework provides higher predictive accuracy for the considered materials compared to conventional and straightforward modeling techniques. Within this architecture, each neural network component contributes a specific function in the estimation pipeline, enabling the prediction of complex features of perovskites, such as the lattice constant an important parameter for thin-film applications. Based on the proposed CNN model, the obtained evaluation criteria (RMSE = 0.017, MAE = 0.013, CC = 99.84%) show a significant improvement compared to previous studies. Furthermore, these findings validate the robustness and reliability of the model for accurate lattice-constant prediction, even when operating under limited or scarce data conditions.</p> Graphical abstract <p></p>