Introduction: Automated nanoscale synthesis plays a pivotal role in materials science, yet the existing methods often struggle to balance efficiency and mechanistic understanding, particularly for complex systems like gold nanoparticles (AuNPs). Large language models (LLMs) show promise in enhancing synthetic workflows, but their alignment with physicochemical principles remains underexplored. This study addresses this gap by integrating Retrieval-Augmented Generation (RAG) with domain-specific expertise to develop an expert-level model for AuNP synthesis.
Methods: We curated a vector database from 62 high-impact research papers focusing on AuNP synthesis, emphasizing their mechanistic insights and experimental conditions. The RAG framework leverages Deepseek as the base LLM, augmented with retrievals from the vector database to contextualize the responses. The evaluation employs the confidence-based score (c-score) proposed in a prior study, which quantifies the model’s certainty in selecting correct answers based on physicochemical mechanisms, alongside the traditional accuracy metrics.
Results: The RAG model demonstrates significant improvements over the baseline Deepseek model, achieving a c-score of 0.78 and an accuracy of 82% on a benchmark of 775 multiple-choice questions derived from AuNP synthesis experiments. These metrics surpass the performance of prior LLMs, indicating a deeper grasp of the underlying mechanisms rather than superficial pattern matching. The case studies reveal the model’s ability to resolve ambiguities in the synthesis pathways, such as the ligand-induced growth directionality and surface energy effects.
Conclusions: This work establishes RAG as a robust framework for automated nanoscale synthesis, combining domain knowledge with advanced reasoning. The integration of expert-curated literature and mechanism-focused evaluation ensures reliable predictions, paving the way for AI-driven discovery in materials science. Future directions include expanding the database to other nanomaterials and refining the retrieval strategies for real-time synthesis optimization.