Nie et al. (2025) Large language models for environmental modeling: Framework, capabilities, constraints

Identification

• Journal: Journal of Environmental Management
• Year: 2025
• Date: 2025-12-30
• Authors: Qiyang Nie, Tong Liu
• DOI: 10.1016/j.jenvman.2025.128417

Research Groups

• Graduate School of Environmental Science, Hokkaido University, Japan
• Faculty of Environmental Earth Science, Hokkaido University, Japan

Short Summary

This study introduces and evaluates two Large Language Model (LLM) integration frameworks, "Copilot" (human-AI collaborative) and "Autopilot" (LLM-driven automation), for environmental modeling workflows like parameter calibration and real-time correction, using the Rainfall–Runoff–Inundation (RRI) model in the Kuzuryu River basin. It finds that Copilot excels in human-supervised tasks, while Autopilot struggles with data-intensive, long-sequence tasks due to attention decay.

Objective

• To introduce and evaluate two Large Language Model (LLM) integration frameworks (Copilot and Autopilot) for environmental modeling, specifically for parameter calibration and real-time correction, using the Rainfall–Runoff–Inundation (RRI) model.

Study Configuration

• Spatial Scale: Kuzuryu River basin, Japan
• Temporal Scale: Processes related to parameter calibration and real-time correction, implying historical periods for calibration and continuous/near-continuous processing for real-time applications.

Methodology and Data

• Models used:
  • Rainfall–Runoff–Inundation (RRI) model
  • Large Language Models (LLMs)
  • Copilot framework (human-AI collaborative LLM integration)
  • Autopilot framework (LLM-driven automation)
• Data sources: Implied observational data for rainfall, runoff, and inundation for model calibration and real-time correction.

Main Results

• The Copilot framework demonstrated robust performance, effectively using prompt engineering for algorithm comprehension and code generation.
• Copilot achieved strong performance in parameter calibration (Nash-Sutcliffe Efficiency of 0.91 for calibration and 0.81 for validation) and delivered stable real-time correction.
• The Autopilot framework showed competence in physics-constrained parameter calibration but failed in long-sequence, data-intensive real-time correction tasks due to "attention decay."
• LLMs are currently most effective as knowledge engines and coding assistants within human-supervised workflows (Copilot), while full automation (Autopilot) is limited by context window size and weaknesses in processing long numerical sequences.

Contributions

• Introduction of two novel, reproducible frameworks (Copilot and Autopilot) for integrating Large Language Models into environmental modeling workflows.
• Empirical evaluation of LLM capabilities and constraints in practical environmental modeling tasks (parameter calibration and real-time correction).
• Identification of key design principles (task decomposition, physics constraints, oversight checkpoints) for generalizable LLM deployment in environmental modeling.
• Highlighting the strengths of LLMs in human-supervised roles and their current limitations in fully automated, data-intensive environmental modeling tasks.

Funding

• Not specified in the provided text.

Citation

@article{Nie2025Large,
  author = {Nie, Qiyang and Liu, Tong},
  title = {Large language models for environmental modeling: Framework, capabilities, constraints},
  journal = {Journal of Environmental Management},
  year = {2025},
  doi = {10.1016/j.jenvman.2025.128417},
  url = {https://doi.org/10.1016/j.jenvman.2025.128417}
}