From Utrecht University, Shanghai University of Electric Power, East China University of Science and Technology, National Center for Magnetic Resonance in Wuhan Chinese Academy of Sciences
Xu, B., Li, G., Wang, B., Bian, J., Pan, H., Min, Y., Qi, G., Xu, J., Deng, F., Ju, F., Ling, H., & Wang, Z. (2025).
Knowledge graph for methane selective conversion: Revisiting and predicting product selectivity and methane conversion. Advanced Science, e14601.
Outline
⭐Motivation
⭐Constructing the Methane Selective Conversion Knowledge Graph
❖ Article retrieval and preprocessing
❖ Definition of 11 entity types and 32 relation types
❖ Entity extraction using Large Language Models
⭐A user-friendly, code-free interface supports intuitive querying and visualization of CH4-KG: http://139.224.202.44:3000/
providing researchers, regardless of AI expertise, with a comprehensive, structured overview of methane selective conversion.
⭐Overview of the DNN-KG Model
The model’s input consists of six entity types: three text-based entities (Catalytic Material, Oxidizer, andTarget Product) and three numerical entities (Reaction Pressure, Reaction Temperature, and Reaction Time). The text-based entities are transformedinto 1536-dimensional vectors using the text-embedding-ada-002 model, while the numerical entities are used directly as values. These together formthe input features for the DNN-KG model. The model is composed of a four-layer pre-trained section and a two-layer re-trained/fine-tuning section. Byleveraging the knowledge graph, the model uses cosine similarity to search for data similar to the initial input vectors, thus constructing a re-trainingdataset and forming localized knowledge graphs.
Predicted CH3OH selectivity and CH4 conversion at 500.0 K under a total reaction pressure of 10.0 bar.
Zeolites- and MOFs-based catalysts (orange dots), transition metal- and metal salt-based catalysts (blue dots) and noble metal-based catalysts (pink dots) overlaid on the non-catalytic selectivity-conversion trade-off line described by Equation (1) at 500 K under 10.0 bar using ΔGDFT (black line) including a ±1 σ error (the shading area between blue and pink dash lines).