EXPRESS: An In-Context Learning Framework for Small-Sample Coal Ash Quantification via Laser-Induced Breakdown...

Area of Science:

• Analytical Chemistry
• Machine Learning
• Spectroscopy

Background:

• Laser-induced breakdown spectroscopy (LIBS) combined with machine learning offers rapid coal ash quantification.
• Practical limitations in coal preparation plants include scarce calibration samples and narrow ash content ranges, hindering reliable quantitative modeling.

Purpose of the Study:

• To develop a structured spectral in-context learning framework for accurate coal ash quantification with limited calibration samples.
• To adapt high-dimensional LIBS spectra for a pre-trained tabular foundation model using novel feature extraction techniques.

Main Methods:

• Proposed a structured spectral in-context learning (SS-ICL) framework utilizing a pre-trained tabular foundation model.
• Developed segmented statistical descriptors and multi-scale Haar-like features to capture subtle spectral variations.
• Applied the framework to small-sample coal ash quantification in a narrow-range scenario.

Main Results:

• The SS-ICL framework demonstrated robust predictive performance under restricted conditions.
• Achieved a coefficient of determination (R²) of 0.8880 and a root-mean-square error (RMSE) of 0.1953 in a narrow-range washed-coal analysis.
• Validated the framework's effectiveness for on-site LIBS-based coal quality analysis.

Conclusions:

• The proposed SS-ICL framework offers a practical and robust solution for coal ash quantification with limited data.
• This approach overcomes challenges posed by restricted industrial environments and narrow ash content ranges.
• Highlights the potential of foundation models and in-context learning for spectral analysis in industrial applications.