A Large-Scale Nanocrystal Database with Aligned Synthesis and Properties, Enabling Generative Inverse Design

Nanocrystal synthesis has been highly dependent on trial and error due to the complex correlation between synthesis parameters and physicochemical properties. Although deep learning offers a potential methodology to achieve generative inverse design, it is still hindered by the scarcity of high-quality data sets that align nanocrystal synthesis routes with their properties. In this work, we developed NanoExtractor, a large language model (LLM) with well-designed data augmentation strategies to extract structured synthesis routes and corresponding properties from unstructured literature. NanoExtractor achieves a weighted average score of 92% on the test set, significantly outperforming other chemistry-specialized (9%) and general-purpose LLMs (57%). With this model, we constructed a large-scale nanocrystal synthesis-property (NSP) database containing nearly 160 000 aligned entries. On the basis of this database, we further developed NanoDesigner, an LLM for generative inverse synthesis design, achieving an F1 score of 0.85. The applicability of NanoDesigner was experimentally validated across multiple nanocrystal systems, including MgF₂, CsPbBr₃, and PbS. Notably, NanoDesigner recommends a critical nonstoichiometric precursor concentration for synthesizing MgF₂ nanocrystals, which was experimentally proven to be essential for suppressing byproduct formation. In all, our work bridges the gap between unstructured literature and data-driven synthesis, providing a human-AI collaborative paradigm for accelerating material discovery.