LUMIR: an LLM-driven unified agent framework for multi-task infrared spectroscopy reasoning.

Area of Science:

• Analytical Chemistry
• Spectroscopy
• Artificial Intelligence

Background:

• Infrared spectroscopy is vital for material characterization but requires complex, expertise-driven analysis workflows.
• Current methods lack robustness and transferability, hindering automated spectral analysis.
• Large language models (LLMs) offer potential for automating and improving these complex processes.

Purpose of the Study:

• To develop a unified, data-efficient framework for automated infrared spectral analysis using LLMs.
• To enhance the accuracy and generalizability of spectral interpretation, especially in low-data scenarios.
• To establish a new paradigm for LLM application in infrared spectroscopy.

Main Methods:

• Introduction of LUMIR (LLM-driven Unified agent framework for Multi-task Infrared spectroscopy Reasoning).
• Integration of a structured literature knowledge base, automated preprocessing, and feature extraction.
• Utilization of few-shot learning with LLMs for classification, regression, and anomaly detection.

Main Results:

• LUMIR achieved performance comparable to established machine learning and deep learning models across diverse datasets.
• The framework demonstrated particular effectiveness in resource-limited settings.
• Successful validation on datasets including near-infrared milk data, medicinal herbs, and industrial wastewater.

Conclusions:

• LLMs, guided by literature and few-shot learning, enable robust and automated spectral interpretation.
• LUMIR offers a new paradigm for applying LLMs to infrared spectroscopy, achieving high accuracy with minimal data.
• The framework shows broad applicability across scientific and industrial domains.