Phthalonitrile is a widely applied resin due to its outstanding high-temperature resistance and versatility. However, the high melting point of its monomer limits its range of applications. Due to the scarcity of data, traditional deep learning models face challenges in accurately predicting the properties of phthalonitrile monomers. In this work, we propose a novel deep learning model based on techniques such as molecular similarity weighting, pre-training parameter averaging, and hybrid neural networks to address the issue of data sparsity. Our model exhibits exceptionally high accuracy for predicting the melting point of phthalonitrile monomers, e.g., the mean absolute error on the experimental dataset is 15.3 °C. To explore novel low-melting monomers, we develop a high-throughput molecule generation software and use our model to filter the virtual molecules, and ultimately obtain candidates that meet low-melting and synthesizability requirements. Moreover, we have successfully synthesized one candidate molecule, and the high consistency between our predictions and experimental measurements again demonstrates our model’s accuracy. Meanwhile, molecule dynamic simulations are conducted to provide explanations for the reduction in melting point. This study not only provides guidance for designing phthalonitrile monomers, but also proposes an innovative approach for exploring of new materials under conditions of sparse data.