An in-Context Learning Framework for Small-Sample Coal Ash Quantification via Laser-Induced Breakdown Spectroscopy.

Abstract

The combination of laser-induced breakdown spectroscopy (LIBS) and machine learning provides a rapid approach for coal ash quantification. However, in practical coal preparation plants, calibration samples are often limited, and ash contents are usually confined to a narrow range, making reliable quantitative modeling difficult. To address this issue, we propose a structured spectral in-context learning framework for small-sample coal ash quantification. Unlike conventional methods that rely on task-specific iterative parameter optimization, the proposed framework performs downstream inference by conditioning on a limited set of calibration samples using a pre-trained tabular foundation model. To adapt continuous high-dimensional LIBS spectra to this framework, segmented statistical descriptors and multi-scale Haar-like features were constructed to capture subtle local spectral variations associated with narrow ash-content changes. The experimental results show that the proposed SS-ICL framework achieves robust predictive performance under restricted conditions, with an R² of 0.8880 and an RMSE of 0.1953 in a narrow-range washed-coal scenario. These results demonstrate that the proposed framework provides a practical and robust strategy for on-site LIBS-based coal quality analysis in restricted industrial environments.