Generative Artificial Intelligence for Synthetic Spectral Data Augmentation in Sensor-Based Plastic Recycling.

Area of Science:

• Spectroscopy
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning for material classification requires extensive labeled spectral data.
• Acquiring large-scale spectral datasets is challenging due to cost and environmental factors.
• Near-infrared (NIR) spectroscopy is vital for applications like plastic sorting.

Purpose of the Study:

• Investigate large language models (LLMs) for synthetic spectral data generation.
• Assess LLMs' capability to create meaningful spectral variations for data augmentation.
• Evaluate if LLM-generated data enhances classification model performance.

Main Methods:

• LLM-guided simulation for generating synthetic spectral data.
• Using minimal empirical mean spectra per class as input.
• Classification accuracy served as the primary metric for augmented spectra plausibility.

Main Results:

• LLM-generated data enabled up to 86% classification accuracy from a single mean spectrum per class.
• The generated spectral variations preserved class-distinguishing information.
• Performance was strongest for spectrally distinct polymers; overlapping classes presented challenges.

Conclusions:

• LLMs show potential for scalable, AI-supported data augmentation in spectroscopy.
• The methodology offers a novel approach to overcome spectral data limitations.
• Applicable to various fields including agriculture and food quality assessment beyond plastic sorting.