RESUMEN
This study explores the use of large language models (LLMs) in interpreting and predicting experimental outcomes based on given experimental variables, leveraging the human-like reasoning and inference capabilities of LLMs, using selective catalytic reduction of NOx with NH3 as a case study. We implement the chain of thought (CoT) concept to formulate logical steps for uncovering connections within the data, introducing an "Ordered-and-Structured" CoT (OSCoT) prompting strategy. We compare the OSCoT strategy with the more conventional "One-Pot" CoT (OPCoT) approach and with human experts. We demonstrate that GPT-4, equipped with this new OSCoT prompting strategy, outperforms the other two settings and accurately predicts experimental outcomes and provides intuitive reasoning for its predictions.
RESUMEN
Mining the scientific literature, combined with data-driven methods, may assist in the identification of optimized catalysts. In this paper, we employed interpretable machine learning to discover ternary metal oxides capable of selective catalytic reduction of nitrogen oxides with ammonia (NH3-SCR). Specifically, we devised a machine learning framework utilizing extreme gradient boosting (XGB), identified for its optimal performance, and SHapley Additive exPlanations (SHAP) to evaluate a curated database of 5654 distinct metal oxide composite catalytic systems containing cerium (Ce) element, with records of catalyst composition and preparation and reaction conditions. By virtual screening, this framework precisely pinpointed a CeO2-MoO3-Fe2O3 catalyst with superior NOx conversion, N2 selectivity, and resistance to H2O and SO2, as confirmed by empirical evaluations. Subsequent characterization affirmed its favorable structural, chemical bulk properties and reaction mechanism. Demonstrating the efficacy of combining knowledge-driven techniques with experimental validation and analysis, our strategy charts a course for analogous catalyst discoveries.