<p>Large-scale foundation models, including universal neural network interatomic potentials (NIPs) in computational materials science, have demonstrated significant progress. However, despite their success in accelerating atomistic simulations, NIPs still face challenges in modeling certain property classes. Machine learning (ML) offers alternatives for structure-to-property mapping but different ML approaches present distinct trade-offs: feature-based methods often lack generalizability, while deep neural networks require significant data and computational power. To address these trade-offs, we introduce HackNIP, a two-stage pipeline that leverages pretrained universal NIPs. This method first extracts embeddings from universal NIP foundation models and then uses these embeddings as fixed-length feature vectors to train downstream learners for structure-to-property predictions. This study investigates whether such a hybridization approach, by “hacking” the NIP, can outperform end-to-end deep neural networks, determines the dataset-size regime in which this approach surpasses direct fine-tuning of the NIP, and identifies which NIP embedding depths yield the most informative features. Together, this work demonstrates a hybridization strategy for overcoming trade-offs in structure-to-property predction in materials science, thereby lowering barriers to practical materials-design workflows.</p>