Semantic metadata extraction from crop model components with large language models to facilitate platform interoperability

Process-based crop models encapsulate agro-ecosystem knowledge within modular components that have been developed across different crop modelling platforms. However, structural heterogeneity and differences in programming languages limit component reuse. The Agricultural Model Exchange Initiative recently proposed Crop2ML, a framework for crop modeling platform interoperability to address this issue. Nevertheless, transforming existing models into Crop2ML specifications still requires substantial manual curation of semantic metadata, which is error-prone and time consuming. Recent advances in large language models (LLMs) have increased their capacity to support automated code analysis through multi-agent system architectures. Here, a workflow was developed leveraging LLMs to automatically create Crop2ML model specifications by extracting and inferring metadata describing input and output variables specifications. The workflow was evaluated on source code written in three programming languages and originating from six different crop modeling platforms. Experiments conducted on eight soil temperature components from these widely used platforms showed a F1-Score exceeding 93% for input/output variables extraction and successful inference of variable specifications. The proposed approach introduces a novel consensus-based multi-agent methodology that improves the reliability of automated extraction results. Overall, this study demonstrated that LLMs-driven multi-agent systems can significantly accelerate the automatic generation of interoperable models with semantically rich model documentation and specification. This ultimately reduces the barriers to crop model reuse.