CoChat
Sign Up
Cited by 19Ammonia Synthesis and Nitrogen ReductionMachine Learning in Materials ScienceElectrocatalysts for Energy Conversion

Automating structure-activity analysis for electrochemical nitrogen reduction catalyst design through multi-agent collaborations

Open access
Hu, Xu, Chen, Suya, Chen, Letian, Wang, Huijuan, Zhang, Xu, Zhou, Zhen·2025·National Science Review
DOI 10.1093/nsr/nwaf372PMC PMC12586994
Open this paper in CoChat — free accountOpen PDF

Abstract

The electrochemical nitrogen reduction reaction (eNRR) offers sustainable ammonia production, yet elucidating structure–activity relationships (SARs) is challenging. We introduce eNRRCrew, a novel multi-agent framework integrating large language models (LLMs), machine learning and automated tools to advance eNRR research. By analyzing 2321 papers, eNRRCrew constructed a comprehensive database of electrocatalyst properties, conditions and performance. The framework employs a random forest classifier for eNRR yield prediction, with model interpretability analysis revealing key factors like space group number and elemental electronegativity difference. Additionally, clustering analysis identifies distinct Faradaic efficiency patterns. eNRRCrew's five LLM agents enable natural language interaction for novel catalyst recommendation, performance prediction, data analysis and literature insights. This approach surpasses traditional methods in extracting SARs and guiding rational catalyst design, offering a scalable platform for various electrocatalysis domains and a new paradigm for LLM-driven scientific discovery.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper introduces eNRRCrew, a multi-agent LLM framework that mines eNRR literature, builds a catalyst-property database, and uses machine learning to predict yield and Faradaic efficiency while recommending novel electrocatalysts. It reports high extraction accuracy and demonstrates natural-language catalyst recommendation, including one computationally validated and one experimentally validated candidate.

Key findings

  • eNRRCrew analyzed 2321 eNRR paper abstracts and extracted 14 key NH3 synthesis parameters with GPT-4o, achieving average precision 0.97, recall 0.95, and F1 0.96 against 50 manually extracted abstracts.
  • A random forest classifier best predicted high vs. low eNRR yield, with training F1 0.84 and test F1 0.63; ROC AUC was 0.90 for training and 0.64 for test. SHAP highlighted space group number, elemental electronegativity difference, Na2SO4 electrolyte, and weak-acid pH as important features.
  • FE regression performed poorly (R² 0.24 training, 0.12 test), so agglomerative clustering identified six FE clusters; Cluster 3 had significantly higher median FE than other clusters.
  • In 10 independent recommendation runs, eNRRCrew proposed 13 unique catalyst systems: 11 predicted high yield/low FE, 1 low yield/low FE, and 1 high yield/high FE (Mo–W dimer on Ti2NO2 MXene), whose thermodynamic stability was supported by AIMD. A recommended MoFeNC catalyst was recently experimentally validated by the authors’ group. Against six recent publications, yield prediction achieved F1 0.67, and FE prediction separated high-performing catalysts (median FE 22.5%) from lower-performing ones (median FE 16.14%).

Methods

  • Literature corpus: 2321 Web of Science abstracts retrieved with an eNRR search query; LDA topic modeling, t-SNE visualization, and OpenAI text-embedding-3-small similarity analysis. Data extraction used ChemPrompt, few-shot prompting, and structured tabular output with GPT-4o; validation used 50 manually annotated abstracts across 14 parameters.
  • Machine-learning dataset: 1837 eNRR entries filtered to electrochemical N2-to-NH3 reactions, excluding missing FE/yield; after preprocessing, 159 descriptor variables and 1117 entries. Magpie generated elemental descriptors; structural and reaction-condition descriptors were one-hot encoded; pH descriptors were constructed from electrolyte information. Models included random forest, support vector classifier, XGBoost, artificial neural networks, and KNN baseline, with RFECV, Optuna hyperparameter tuning, 80/20 train-test split, SHAP, and Ward-linkage agglomerative clustering for FE.
  • eNRRCrew framework: built on AutoGen with a Chainlit interface; five agents are an orchestrator (GPT-4o), yield predictor, FE predictor, GraphRAG retriever, and CSV file handler, with task agents driven by GPT-4o-mini; an additional code execution agent runs code in Docker. The GraphRAG retriever uses a knowledge graph constructed from structured tabular data.

Limitations

  • Stated: reliance on abstracts may omit granular experimental details; scaling analysis shows performance improves with more data; extending to full text and multimodal data is needed; quantitative stability metrics are limited by scarce standardized, machine-readable data and publication bias against negative results; future work should transition to regression models.
  • Evident from the text: test-set predictive performance is modest (yield F1 0.63, AUC 0.64; FE regression R² 0.12), extraction validation used only 50 abstracts, practical testing used only six recent publications, and the final dataset contained 1117 entries with right-skewed yield distributions and generally low FE/yield values.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1093/nsr/nwaf372

Cites

53 of 53 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 39 cited sources checked · 0 supported · 0 flagged

  • Findingreasoning

    A greater difference in electronegativity among various elements in the electrocatalyst is associated with a higher eNRR yield.

    “Also, a greater difference in electronegativity among various elements in the electrocatalyst is associated with a higher eNRR yield.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The SHAP feature importances from the random forest model reflect causal structure–activity relationships rather than only model-specific associations.
  • Findingown data

    The eNRR yield distribution in the dataset is non-normal and right-skewed for both yield metrics.

    “Contrary to an ideal normal distribution, it reveals that both yield metrics exhibit a non-normal, right-skewed character.”
    Quote in the paperNot a substantial claim
  • Findingown data

    The GPT-4o-based extraction framework achieved average precision, recall and F1 scores of 0.97, 0.95 and 0.96, respectively, across 14 key NH3 synthesis parameters.

    “Our framework, based on the GPT-4o model, demonstrated exceptional performance with average precision and recall, and F1 scores of 0.97, 0.95 and 0.96, respectively (Fig. S3).”
    Quote in the paperSubstantial claim
  • Findingown data

    After data preprocessing and feature engineering, the dataset comprises 159 descriptor variables and 1117 entries.

    “After data preprocessing and feature engineering, the dataset comprises 159 descriptor variables, as listed in Table S2, and 1117 entries, which we utilized for the subsequent statistical analysis and development of machine-learning models (see METHODS for details).”
    Quote in the paperSubstantial claim
  • Findingown data

    The random forest classifier performed best among the considered machine-learning models.

    “Using Optuna [37] to find the optimal hyperparameters, we discovered that the random forest classifier performed the best, as demonstrated in Table S3.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [37] Akiba, 2019, Optuna: a next-generation hyperparameter optimization framework source

      Not checked in the free run.

  • May combine two claims. 39
  • The quote only partly states this claim. 1
  • Classified as probably not a substantial claim. 35

  • May combine two claims.

    Findingown data

    The yield in unit mass has a median of 26.15 μg h−1 mgcat−1 and is heavily concentrated at lower values with a long tail extending towards higher yield.

    “The yield in unit mass (median 26.15 μg h−1 mgcat−1) is heavily concentrated at lower values with a long tail extending towards higher yield”
    Quote in the paperSubstantial claim

  • May combine two claims. The quote only partly states this claim.

    Findingreasoning

    DeepSeek-R1 resulted in lower accuracy and significantly slower processing times for structured extraction than GPT-4o.

    “ultimately resulting in lower accuracy and significantly slower processing times for this structured extraction task.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Cluster 3 exhibits significantly higher median FE values compared with other clusters, with its 25th and 75th percentiles also surpassing those of the remaining clusters.

    “Notably, Cluster 3 exhibits significantly higher median values for FE compared with other clusters, with its 25th and 75th percentiles also surpassing those of the remaining clusters.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    'Electrolyte without concentration (Na2SO4)' and 'pH (weak acid)' show high SHAP feature importance.

    “‘Electrolyte without concentration (Na2SO4)’ and ‘pH (weak acid)’ also show high SHAP feature importance”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The SHAP feature importances from the random forest model reflect causal structure–activity relationships rather than only model-specific associations.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    t-SNE visualization of papers based on topic distribution revealed different clusters, each highlighting a specific focus ranging from efficiency optimization, synthesis processes, materials properties, molybdenum-based catalysts, experimental methods and theoretical studies to general performance enhancements.

    “The resulting t-SNE plot revealed different clusters; each cluster highlights a specific focus, ranging from efficiency optimization, synthesis processes, materials properties, molybdenum-based catalysts, experimental methods and theoretical studies to general performance enhancements.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The random forest classifier achieved a prediction F1 score of 0.84 for the training set and 0.63 for the test set.

    “The confusion matrices in Fig. 3a reveal that the classifier achieved a prediction F1 score of 0.84 for the training set and 0.63 for the test set, demonstrating its effectiveness in distinguishing between high and low yields.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    After recursive feature elimination, a total of 152 features were retained.

    “After performing recursive feature elimination, a total of 152 features were retained.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Topic modeling with LDA identified an optimal number of seven topics, achieving a peak coherence score of 0.50.

    “The results reveal an optimal number of seven topics, achieving a peak coherence score of 0.50.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    There is no significant correlation between yield and various elemental property descriptors, with absolute correlation coefficients for unit in area and unit in mass both below 0.26.

    “Additionally, it is evident that there is no significant correlation between yield and various elemental property descriptors, as the absolute values of the correlation coefficients for unit in area and unit in mass are both below 0.26, as shown in Figs S9 and S13.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The yield in unit area has a median of 12.26 μg h−1 cm−2 and displays an even more pronounced skew.

    “while the yield in unit area (median: 12.26 μg h−1 cm−2) displays an even more pronounced skew.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingown data

    For the FE dataset, regression proved ineffective, and agglomerative clustering techniques were applied.

    “For the FE dataset, where regression proved ineffective, we applied agglomerative clustering techniques.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Of the 13 recommended systems, one was predicted to have both low yield and low FE.

    “One system was predicted to have both low yield and low FE”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    Apart from the orchestrator, each agent can perform tool calling to accomplish tasks that users request.

    “Apart from the orchestrator, each agent can perform tool calling to accomplish tasks that users request.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    Clustering techniques further elucidated factors influencing FE.

    “Clustering techniques further elucidated factors influencing FE.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Scaling analysis (Fig. S32) confirms that though the current dataset is effective, model performance consistently improves with more data.

    “Our scaling analysis, as shown in Fig. S32, further confirms that though our current dataset is effective, model performance consistently improves with more data, highlighting a clear path for future enhancement.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingown data

    In a demonstration, eNRRCrew's orchestrator selected the CSV file handler to address a user query about the relationship between FE and applied electrode potential; the CSV file handler proposed a plan and generated Python code that was executed to produce a scatter plot.

    “As shown in Fig. 5c, the user wanted to explore whether there is a relationship between FE and the corresponding applied electrode potential. In response to this query, the orchestrator in eNRRCrew selected the next speaker as the CSV file handler based on its understanding of the user's question.”
    • Also stated in: § eNRRCrew, a multi-agent collaboration framework · plain
      “The CSV file handler, after considering the user's inquiry, proposed a plan to resolve the issue by creating a scatter plot of the two variables and generating the corresponding Python code. This code was executed by the code executor, automatically producing the plot shown in Fig. 5d.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Findingnone given

    The crew automatically executes relevant tools to transform user-provided text information into CSV files usable by the Yield Predictor, and executes this function asynchronously, automating data preprocessing, feature engineering and predictive model execution, enabling end-to-end catalyst recommendation and performance prediction through natural language.

    “The crew automatically executes relevant tools to transform user-provided text information into CSV files usable by the Yield Predictor. The crew then executes this function asynchronously, automating the processes of data preprocessing, feature engineering and predictive model execution, enabling end-to-end catalyst recommendation and performance prediction through natural language.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Findingreasoning

    The orchestrator allocates tasks through a dynamic, context-aware process driven by an LLM rather than by fixed rules or keyword matching.

    “The orchestrator allocates tasks through a dynamic, context-aware process driven by an LLM rather than by fixed rules or keyword matching.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    eNRRCrew excels in recommending novel electrocatalytic systems and forecasting their performance, facilitated by a seamless natural language interface and automated dataset construction using techniques like ChemPrompt and few-shot prompting.

    “Notably, eNRRCrew excels in recommending novel electrocatalytic systems and forecasting their performance, facilitated by a seamless natural language interface and automated dataset construction using techniques like ChemPrompt and few-shot prompting.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingcitation

    eNRRCrew is primarily built upon AutoGen.

    “eNRRCrew is a multi-agent collaboration framework for eNRR research, primarily built upon AutoGen [50].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [50] Wu, Autogen: enabling next-gen LLM applications via multi-agent conversation framework

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    Users can interact with eNRRCrew through natural language.

    “Within this framework, users can interact with the crew through natural language.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    AIMD simulations confirmed the high thermodynamic stability of the Mo-W dimer on Ti2NO2 MXene, providing computational validation for eNRRCrew's top prediction.

    “To directly probe the viability of this novel candidate, we performed ab initio molecular dynamics (AIMD) simulations. Our results confirmed its high thermodynamic stability, providing crucial computational validation for eNRRCrew's top prediction.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Findingnone given

    The final dataset used for model construction contained 1117 entries after data cleaning and preprocessing.

    “The final dataset used for model construction contained 1117 entries after data cleaning and preprocessing.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    In a demonstration, when a user asked which catalysts exhibit a high yield per unit mass in eNRR, the orchestrator assigned the query to the GraphRAG retriever, which retrieved pertinent data from a knowledge graph and formulated a detailed response emphasizing high NH3 yield rates of various catalysts, such as 3D COFs with Fe-N4, supported by specific data references.

    “as shown in Fig. 5e, when a user asks which catalysts exhibit a high yield per unit mass in the eNRR, the orchestrator assigns the query to the GraphRAG retriever, which retrieves pertinent data from a knowledge graph to formulate a detailed response. The answer emphasizes the high NH3 yield rates of various catalysts, such as the 3D COFs with Fe–N4, supported by specific data references.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Of the 13 recommended systems, eNRRCrew anticipated that 11 would exhibit high NH3 yield but low FE.

    “Of these, the crew anticipated that 11 systems would exhibit high NH3 yield but low FE.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    eNRRCrew is a multi-agent collaboration framework consisting of five LLM-based agents (orchestrator, yield predictor, FE predictor, GraphRAG retriever, CSV file handler) using GPT-4o and GPT-4o-mini.

    “we have developed a multi-agent collaboration framework called eNRRCrew, as depicted in Fig. 5a”
    • Also stated in: § eNRRCrew, a multi-agent collaboration framework · plain
      “This crew consists of five agents, each based on LLMs, specifically using GPT-4o and GPT-4o-mini in this work.”
    • Also stated in: § CONCLUSIONS · plain
      “In this work, we introduced eNRRCrew, a multi-agent collaboration framework leveraging LLMs and machine learning to address challenges in eNRR research.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    If a code execution error occurs, eNRRCrew allows the code executor to return the error message to the CSV file handler, which can regenerate the code until the task is successfully completed.

    “if there is a code execution error, the crew allows the code executor to return the error message to the CSV file handler, which can then regenerate the code until the task is successfully completed.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Random forest classifiers effectively predicted eNRR yield, with SHAP analysis identifying critical descriptors such as space group number and electronegativity differences, offering vital insights for electrocatalyst design.

    “Random forest classifiers effectively predicted eNRR yield, with SHAP analysis identifying critical descriptors such as space group number and electronegativity differences, offering vital insights for electrocatalyst design.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Findingreasoning

    The orchestrator's LLM evaluates the entire conversation history and predefined roles/capabilities of all available agents, reasons which agent's expertise is most relevant, and selects the most suitable agent to speak next.

    “At each turn, the orchestrator's underlying LLM evaluates the entire conversation history along with the predefined roles and capabilities of all available agents. Based on this holistic analysis, it reasons which agent's expertise is most relevant to the current state of the task and selects the most suitable agent to speak next.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingnone given

    The orchestrator is based on GPT-4o, while the other agents executing specific tasks are driven by GPT-4o-mini.

    “Specifically, the orchestrator is based on GPT-4o, while the other agents executing specific tasks are driven by GPT-4o-mini.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingown data

    The knowledge graph underpinning the GraphRAG retriever is constructed from structured tabular data with a multi-step, automated pipeline that includes entity disambiguation, description summarization and efficient incremental updates.

    “The knowledge graph, which underpins the GraphRAG retriever agent, is constructed from structured tabular data with a multi-step, automated pipeline.”
    • Also stated in: § Details of eNRRCrew · plain
      “The pipeline includes sophisticated mechanisms for entity disambiguation, description summarization and efficient incremental updates to ensure the graph remains current and coherent.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    There is also a code execution agent that can interact with the CSV file handler, yield predictor and FE predictor to complete complex tasks by executing code in a Docker image.

    “In addition to the agents described in the main text, there is also a code execution agent. The code execution agent can interact with the CSV file handler, yield predictor and FE predictor to complete complex tasks by executing code in a Docker image.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Findingnone given

    The yield predictor and FE predictor, based on pretrained models, can automatically perform data preprocessing, feature engineering and model prediction.

    “The yield predictor and FE predictor, based on pretrained models in the former section, can automatically perform data preprocessing, feature engineering and model prediction.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Findingnone given

    The GraphRAG retriever provides evidence-based responses by invoking GraphRAG queries as functions, and is capable of recommending novel catalysts.

    “The GraphRAG retriever provides evidence-based responses by invoking GraphRAG queries as functions, and is also capable of recommending novel catalysts.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    The regression model struggles to accurately capture the underlying relationships between the descriptors and the FE, which demonstrates that the factors influencing FE are more complex, involving side reactions such as the hydrogen evolution reaction.

    “This illustrates that the regression model struggles to accurately capture the underlying relationships between the descriptors and the FE, which demonstrates that the factors influencing FE are more complex, involving side reactions such as the hydrogen evolution reaction.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    The text mining workflow is more powerful, effective and concise compared with traditional NLP methods, providing an automated and efficient solution for processing large volumes of electrocatalysis literature data.

    “The text mining workflow is more powerful, effective and concise compared with traditional NLP methods, providing an automated and efficient solution for processing large volumes of electrocatalysis literature data, making it more versatile and universally applicable.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    Samples within Cluster 3 are associated with more favorable outcomes, suggesting that the features characterizing this cluster may be particularly advantageous for higher FE.

    “This indicates that samples within Cluster 3 are associated with more favorable outcomes, suggesting that the features characterizing this cluster may be particularly advantageous for higher FE.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    The few-shot prompt strategy helps mitigate hallucinations within ChatGPT and enhances the accuracy and reliability of the responses.

    “This approach helps mitigate hallucinations within ChatGPT and enhances the accuracy and reliability of the responses.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    The yield is influenced by a complex interplay of multiple factors, and no single descriptor can solely determine yield levels.

    “This indicates that the yield is influenced by a complex interplay of multiple factors, and no single descriptor can solely determine yield levels.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The extracted eNRR data entries from abstracts are comparable across studies despite differences in experimental conditions.

  • May combine two claims.

    Premisecitation

    Elements with small space group numbers typically exhibit simpler symmetry and fewer symmetry operations, contributing to more isotropic physical properties and easier combinations with other components, leading to enhanced stability and performance.

    “Elements with small space group numbers typically exhibit simpler symmetry and fewer symmetry operations, contributing to more isotropic physical properties and easier combinations with other components, leading to enhanced stability and performance [40].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [40] Zhou, 2022, Prediction of formation energies of UCr4C4-type compounds from Magpie feature descriptor-based machine learning approaches, Opt Mater X source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    This finding is consistent with previous studies that identified the electron number of constituent elements as an integral component of the universal descriptor for eNRR.

    “This finding is consistent with previous studies that identified the electron number of constituent elements as an integral component of the universal descriptor for eNRR [48].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [48] Ren, 2022, A universal descriptor for complicated interfacial effects on electrochemical reduction reactions, J Am Chem Soc source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationcitation

    The correlation between greater electronegativity difference and higher eNRR yield may be due to enhanced charge transfer, modulation of the d-band center of the central metal, or generation of Lewis acid sites that promote N2 activation and inhibit H adsorption.

    “This correlation may be due to the enhanced charge transfer [42], the modulation of the d-band center of the central metal [43], or the generation of Lewis acid sites that promote N2 activation and inhibit H adsorption [44].”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [42] Ma, 2019, Computational evaluation of electrocatalytic nitrogen reduction on TM single-, double-, and triple-atom catalysts (TM = Mn, Fe, Co, Ni) based on graphdiyne monolayers, J Phys Chem C source

      Not checked in the free run.

    • cites [43] Das, 2022, Computational screening of first-row transition-metal based alloy catalysts-ligand induced N2 reduction reaction selectivity, ACS Phys Chem Au source

      Not checked in the free run.

    • cites [44] Ying, 2022, Rational design of atomic site catalysts for electrocatalytic nitrogen reduction reaction: one step closer to optimum activity and selectivity, Electrochem Energy Rev source

      Not checked in the free run.

    • Rests on an assumption: The SHAP feature importances from the random forest model reflect causal structure–activity relationships rather than only model-specific associations.

  • May combine two claims.

    Interpretationown data

    The collected abstracts are comprehensive and diverse, serving as an excellent corpus base for LLMs to extract structure-activity relationships in eNRR.

    “These results suggest that the collected abstracts are comprehensive and diverse, serving as an excellent corpus base for LLMs to extract structure–activity relationships in eNRR.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The 2321 abstracts retrieved from Web of Science are representative of the eNRR literature.

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    eNRRCrew empowers researchers to extract, analyze and predict structure-activity relationships with unprecedented efficiency.

    “The framework empowers researchers to extract, analyze and predict structure–activity relationships with unprecedented efficiency.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    Recursive feature elimination with cross-validation (RFECV) was employed to select the most informative subset of features.

    “First, we employed recursive feature elimination with cross-validation (RFECV) to select the most informative subset of features.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Study design and conductnone given

    Abstracts may omit granular experimental details.

    “abstracts may omit granular experimental details.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    This experimental validation underscores the potential of eNRRCrew not only to organize existing knowledge but also to generate novel, experimentally verifiable hypotheses, thereby accelerating the discovery cycle for advanced eNRR electrocatalysts.

    “This validation underscores the potential of eNRRCrew not only to organize existing knowledge but also to generate novel, experimentally verifiable hypotheses, thereby accelerating the discovery cycle for advanced eNRR electrocatalysts.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationreasoning

    eNRRCrew signifies an advancement in applying multi-agent LLM frameworks to scientific research.

    “eNRRCrew signifies an advancement in applying multi-agent LLM frameworks to scientific research.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    The co-clustering of recommended high-yield systems with known high-yield catalysts provides compelling evidence for the validity and effectiveness of the trained performance prediction model in identifying promising new catalyst candidates, as well as the rationale for using eNRRCrew for novel catalyst recommendations.

    “This co-clustering provides compelling evidence for the validity and effectiveness of our trained performance prediction model in identifying promising new catalyst candidates, as well as the rationale for using eNRRCrew for novel catalyst recommendations.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    Optuna was utilized for automatic hyperparameter optimization within a cross-validation loop to prevent overfitting to the training data.

    “For each candidate model, we utilized Optuna [37], an automatic hyperparameter optimization framework, to find the optimal hyperparameters for each model. This optimization was performed within a cross-validation loop to prevent overfitting to the training data.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [37] Akiba, 2019, Optuna: a next-generation hyperparameter optimization framework source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Interpretationreasoning

    eNRRCrew provides a versatile, scalable platform for accelerating eNRR studies and guiding rational catalyst design, with its methodologies holding promise for broader applications across diverse electrocatalysis domains and other scientific fields.

    “It provides a versatile, scalable platform for accelerating eNRR studies and guiding rational catalyst design, with its methodologies holding promise for broader applications across diverse electrocatalysis domains and other scientific fields.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Premisecitation

    An inherent limitation of LLMs is their restricted capability to reason and understand complex scenarios, often struggling with intricate reasoning tasks due to reliance on patterns in training data rather than genuine comprehension, known as hallucination.

    “In addition, an inherent limitation of LLMs is their restricted capability to reason and understand complex scenarios. These models often struggle with intricate reasoning tasks due to their reliance on patterns in the training data rather than genuine comprehension, known as hallucination [18].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [18] Li, Enhancing LLM factual accuracy with RAG to counter hallucinations: a case study on domain-specific queries in private knowledge-bases

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Study design and conductnone given

    The responses provided by eNRRCrew are all evidence-based, and users can access the references cited by the agents.

    “Unlike typical chatbots, the responses provided by the crew are all evidence-based, and users can also access the references cited by the agents for further research.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    The eNRR field faces challenges including low Faradaic efficiency, limited NH3 yield, and an incomplete understanding of structure-activity relationships.

    “Despite significant research efforts to develop efficient electrocatalysts for eNRR, the field faces several critical challenges [3,4], including low Faradaic efficiency (FE), limited NH3 yield, and an incomplete understanding of the structure–activity relationships [5].”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [3] Choi, 2020, Identification and elimination of false positives in electrochemical nitrogen reduction studies, Nat Commun source

      Not checked in the free run.

    • cites [4] Tang, 2019, How to explore ambient electrocatalytic nitrogen reduction reliably and insightfully, Chem Soc Rev source

      Not checked in the free run.

    • cites [5] Shen, 2021, Electrochemical ammonia synthesis: mechanistic understanding and catalyst design, Chem source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

    Premisecitation

    LLM agents serve as intelligent systems that bridge the gap between LLMs and practical applications, making significant impacts in various fields.

    “LLM agents [19] serve as intelligent systems that bridge the gap between LLMs and practical applications, making significant impacts in various fields like human behavior simulation [20], chemistry [21–24], bioinformatics [25] and paper searching [26].”
    Quote in the paperNot a substantial claim0 of 6 sources checked
    • cites [19] Yao, ReAct: synergizing reasoning and acting in language models

      Not checked in the free run.

    • cites [20] Park, 2023, Generative agents: interactive simulacra of human behavior source

      Not checked in the free run.

    • cites [21] Bran, 2024, Augmenting large language models with chemistry tools, Nat Mach Intell source

      Not checked in the free run.

    • cites [22] Kang, 2024, ChatMOF: an artificial intelligence system for predicting and generating metal-organic frameworks using large language models, Nat Commun source

      Not checked in the free run.

    • cites [23] Ruan, 2024, An automatic end-to-end chemical synthesis development platform powered by large language models, Nat Commun source

      Not checked in the free run.

    • cites [24] Song, 2025, A multiagent-driven robotic AI chemist enabling autonomous chemical research on demand, J Am Chem Soc source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Computational methods such as DFT have provided valuable insights into reaction mechanisms but often struggle to capture the intricate complexities of the realistic electrocatalytic environment and experimental conditions.

    “Though computational methods such as density functional theory (DFT) have provided valuable insights into reaction mechanisms [8–11], they often struggle to capture the intricate complexities of the realistic electrocatalytic environment and experimental conditions.”
    Quote in the paperSubstantial claim0 of 4 sources checked
    • cites [8] Singstock, 2022, How the bioinspired Fe2Mo6S8 chevrel breaks electrocatalytic nitrogen reduction scaling relations, J Am Chem Soc source

      Not checked in the free run.

    • cites [9] Sun, 2024, Interpretable machine learning-assisted high-throughput screening for understanding NRR electrocatalyst performance modulation between active center and C-N coordination, Energy Environ Mater source

      Not checked in the free run.

    • cites [10] Zhang, 2018, Double-atom catalysts: transition metal dimer-anchored C2N monolayers as N2 fixation electrocatalysts, J Mater Chem A source

      Not checked in the free run.

    • cites [11] Abed, Open catalyst experiments 2024 (OCx24): bridging experiments and computational models

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Traditional approaches to analyzing structure-activity relationships in eNRR have relied heavily on experimental trial-and-error methods and isolated case studies, which are time-consuming and limited in identifying broader patterns across diverse electrocatalytic systems.

    “Traditional approaches to analyzing structure–activity relationships in eNRR have relied heavily on experimental trial-and-error methods and isolated case studies, which are both time-consuming and limited in their ability to identify broader patterns across diverse electrocatalytic systems [7].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [7] Taylor, 2023, A brief introduction to chemical reaction optimization, Chem Rev source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductown data

    20% of the data was allocated as an independent test set, and 80% constituted the training set.

    “We allocated 20% of the data to serve as an independent test set, while the remaining 80% constituted our training set.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    LLMs have demonstrated remarkable capability in natural language understanding and pattern recognition across various scientific domains.

    “These tools have demonstrated remarkable capability in natural language understanding and pattern recognition across various scientific domains [15–17].”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [15] Zheng, 2023, ChatGPT chemistry assistant for text mining and the prediction of MOF synthesis, J Am Chem Soc source

      Not checked in the free run.

    • cites [16] Xiong, 2025, Bridging chemistry and artificial intelligence by a reaction description language, Nat Mach Intell source

      Not checked in the free run.

    • cites [17] Song, LLM-Feynman: leveraging large language models for universal scientific formula and theory discovery

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Interpretationreasoning

    eNRRCrew addresses the limitations of traditional methods, making it a valuable tool for advancing research in fields like electrocatalysis and beyond.

    “By bridging the gap between extensive data and actionable insights, eNRRCrew addresses the limitations of traditional methods, making it a valuable tool for advancing research in fields like electrocatalysis and beyond.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    A rigorous machine-learning protocol was implemented to mitigate overfitting and ensure selection of a high-performing model.

    “To build a robust and generalizable predictive model, we implemented a rigorous machine-learning protocol. This protocol involved several key steps to mitigate overfitting and ensure the selection of a high-performing model.”
    Quote in the paperSubstantial claim
61 of 61

  • The authors state this as an assumption.

    The electrocatalyst loading is about 1 mg cm−2 when comparing eNRR yields to the practical threshold of 100 nmol s−1 cm−2.

    “assuming an electrocatalyst loading of about 1 mg cm−2 [32,33].”

  • The paper relies on this without support.

    The SHAP feature importances from the random forest model reflect causal structure–activity relationships rather than only model-specific associations.

    “To get the implication of structure–activity relationships, we used Shapley additive explanations (SHAP) analysis to quantify the contribution of each feature to the model's predictions, as shown in Fig. S21.”

  • The paper relies on this without support.

    The extracted eNRR data entries from abstracts are comparable across studies despite differences in experimental conditions.

    “A total of 1837 eNRR-related data entries were identified where the reaction type is electrochemical, the reactant is N2 and the product is NH3.”

  • The paper relies on this without support.

    The 2321 abstracts retrieved from Web of Science are representative of the eNRR literature.

    “In this study, we systematically analyzed a literature corpus comprising 2321 paper abstracts retrieved from the Web of Science database (see METHODS for details).”

  • The paper relies on this without support.

    The effects of the electrocatalyst, reaction conditions such as pH, applied electrode potential and electrolyte on eNRR performance are not fully understood.

    “Specifically, the effects of the electrocatalyst, reaction conditions such as pH, applied electrode potential and electrolyte on eNRR performance are not fully understood.”

  • The paper relies on this without support.

    The eNRR field commonly reports key performance metrics in abstracts.

    “the common practice in the eNRR field of reporting key performance metrics in abstracts”

  • The paper relies on this without support.

    There is a field-wide scarcity of standardized, machine-readable data for quantitative stability metrics.

    “this is currently challenged by a field-wide scarcity of standardized, machine-readable data”

  • The paper relies on this without support.

    There is a publication bias against negative results in the eNRR field.

    “a publication bias against negative results”
8 of 8

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • These papers typically split the workflow into role-specific agents for planning, structure retrieval, simulation execution, analysis, and quality control.

    “These papers typically split the workflow into role-specific agents for planning, structure retrieval, simulation execution, analysis, and quality control [18,19,21,25,26,35-38,41,42].”

    From From Scientific Copilots to Tool-Grounded Autonomy: AI Agents in Simulation-Driven Materials Discovery (2026) · cites this paper as [19]

    The Results and Discussion section describes eNRRCrew as a set of role-specific agents for task allocation, prediction, knowledge retrieval, and CSV handling; the Methods section adds a code-execution agent. But this paper does not describe agents specifically assigned to structure retrieval, simulation execution, or quality control. Its code executor runs code for tasks such as plotting, not a stated simulation workflow. (LLM, conf=0.86)

    Experimental — accuracy not yet measured

  • In several papers, the strongest reproducibility evidence came from public code, tutorials, datasets, or supplementary traces.

    “In several papers, the strongest reproducibility evidence came from public code, tutorials, datasets, or supplementary traces [18-21,23,25,28,30,31,35-42].”

    From From Scientific Copilots to Tool-Grounded Autonomy: AI Agents in Simulation-Driven Materials Discovery (2026) · cites this paper as [19]

    The Methods (page unavailable) point to supplementary data describing the framework’s graph schema, storage architecture, and entity-management strategies; the Results (page unavailable) also refer to supplementary figures documenting validation. These provide some supporting detail, but our copy does not say that public code, tutorials, datasets, or supplementary traces were the strongest reproducibility evidence, or compare this paper with the other cited papers. (LLM, conf=0.72)

    Experimental — accuracy not yet measured

  • Catalysis and adsorption tasks formed the largest application block, represented by six studies on electrocatalyst design, adsorption-configuration search, biomass-valorization catalysis, MOF catalyst optimization, and single-atom-alloy discovery.

    “Catalysis and adsorption tasks formed the largest application block, represented by six studies on electrocatalyst design, adsorption-configuration search, biomass-valorization catalysis, MOF catalyst optimization, and single-atom-alloy discovery [19,20,27,28,30,34].”

    From From Scientific Copilots to Tool-Grounded Autonomy: AI Agents in Simulation-Driven Materials Discovery (2026) · cites this paper as [19]

  • Other studies act more as discovery managers around retrieval, ranking, surrogate models, or protocol navigation.

    “Others act more as discovery managers around retrieval, ranking, surrogate models, or protocol navigation [19,20,26,30,41].”

    From From Scientific Copilots to Tool-Grounded Autonomy: AI Agents in Simulation-Driven Materials Discovery (2026) · cites this paper as [19]

Cites

Flagged first, then by citations

Research works printed but not linked: 1

  1. Lewis, Retrieval-augmented generation for knowledge-intensive NLP tasks, NIPS '20: 34th International Conference on Neural Information Processing Systems Vancouver BC Canada, 6–12 December 2020, 9459

Cited by

Sorted by citations

Related

Authors
Hu, X., Chen, S., Chen, L., Wang, H., Zhang, X., & Zhou, Z.
Affiliations
  • Nankai University (CN)
  • Zhengzhou University (CN)
Corresponding author
  • Zhang, Xu
  • Zhou, Zhen
ORCID
Venue
National Science Review
Publisher
Oxford University Press (OUP)
Published
2025-09-04
Volume
12
Issue
11
Pages
nwaf372-nwaf372
ISSN
  • 2053-714X (ISSN-L)
  • 2095-5138
DOI
10.1093/nsr/nwaf372
PMC
PMC12586994
OpenAlex
W4413986282
Semantic Scholar
0c88517ff3551633ab3d45a965c8f4c7fd337641
Cited by
19
References
53
Citations by year
2026: 19
FWCI
2.57
Type
journal article
Language
en
Open access
Licence
Locations
  • National Science Review (publishedVersion, cc-by)
  • PubMed (publishedVersion)
Funders
  • National Natural Science Foundation of China (U21A20281, 223B2305, 22203077)
  • Natural Science Foundation of Henan Province (242300421129, 232301420051)
  • Frontier Exploration Projects of Longmen Laboratory (LMQYTSKT021)
Topics
  • Physical Sciences › Chemical Engineering › Catalysis › Ammonia Synthesis and Nitrogen Reduction
  • Physical Sciences › Materials Science › Materials Chemistry › Machine Learning in Materials Science
  • Physical Sciences › Energy › Renewable Energy, Sustainability and the Environment › Electrocatalysts for Energy Conversion
Keywords
electrochemical nitrogen reduction, ammonia synthesis, structure-activity relationship, electrocatalyst design, multi-agent systems, large language models, random forest prediction, Faradaic efficiency

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-08.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-08.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review