Improving small-sample performance in portable LIBS analysis with transfer learning and synthetic data-driven CNN models

Abstract

To address the challenge of limited and unstable spectral data from portable laser-induced breakdown spectroscopy (LIBS) devices, this study employed a transfer-learning framework augmented with simulated data. A lightweight one-dimensional convolutional neural network (1D-CNN) was developed, allowing knowledge transfer from high-quality simulated or laboratory spectra to portable-device measurements. Under small-sample and noisy conditions, the transfer-learning approach significantly improved the regression accuracy and stability compared to the performance of the baseline CNN model. Experimental evaluation on multiple elements demonstrated that when simulated data serve as the source domain, the regression performance approaches that achieved using high-precision laboratory spectra and, for certain elements, surpasses that of traditional real-data transfer learning. These results suggest that simulated data-based transfer learning constitutes a viable strategy for enhancing the quantitative analytical performance of portable LIBS systems, particularly under field conditions where high-quality spectral data are difficult to obtain.