Ammonia, a key component of fertilizers, is vital for global food security, but its production via the Haber-Bosch process consumes 1%–2% of the world's energy and creates significant carbon emissions. The electrochemical nitrogen reduction reaction (eNRR) offers a cleaner, sustainable alternative by using renewable electricity to convert nitrogen and water into ammonia at ambient conditions. However, finding the right catalyst for this process is a major bottleneck, with researchers facing a "trial-and-error" cycle and a rapidly growing body of literature that is nearly impossible to analyze manually.
To break this deadlock, a research team led by Prof. Zhen Zhou from Nankai University and Prof. Xu Zhang from Zhengzhou University has developed eNRRCrew, a novel multi-agent AI framework. Described in the journal National Science Review , this system integrates large language models (LLMs) and machine learning to automate the entire research cycle—from data extraction to catalyst recommendation.
The eNRRCrew began by analyzing 2,321 scientific papers, constructing a comprehensive database of eNRR catalysts, reaction conditions, and performance metrics. This task, which would take human researchers even months, was completed automatically and efficiently. Using this database, the framework trained a machine learning model to predict a catalyst's ammonia yield. The model revealed that a catalyst's crystal structure symmetry (space group number) and the difference in electronegativity between its elements are critical factors for high performance. This insight moves beyond simple correlations and provides rational design principles for future experiments.
The framework consists of five specialized AI agents that collaborate to assist researchers. Users can interact with eNRRCrew in natural language to ask complex questions, predict the performance of a new catalyst design, or request novel catalyst recommendations. For instance, the system can generate data plots on demand or provide evidence-based summaries of the research landscape, citing its sources. Crucially, eNRRCrew is not just an analysis tool; it is a discovery engine. The system proactively recommended 13 novel catalyst systems, predicting their potential ammonia yield and efficiency. One of the top candidates, a Mo–W dimer on a Ti 2 NO 2 MXene substrate, was confirmed to be highly stable through advanced computational simulations. Another recommendation, MoFeNC, was recently synthesized and experimentally validated by the team, showing promising activity.
"This multi-agent approach overcomes the limitations of traditional research methods and even single LLMs," says Prof. Zhou. "It provides a powerful, scalable platform that can guide rational catalyst design and significantly accelerate the pace of scientific discovery." The success of eNRRCrew demonstrates a new paradigm for AI-driven science, where autonomous agents can collaborate to solve complex scientific problems. The framework's design is flexible and has already been adapted for other electrocatalytic reactions, showcasing its potential to impact a wide range of fields beyond green ammonia production.
National Science Review
Computational simulation/modeling