Many-to-many transfer of LIBS spectra across multiple experimental conditions

Abstract

Laser-Induced Breakdown Spectroscopy (LIBS) is a powerful analytical technique widely used for extraterrestrial remote analysis. However useful, its primary limitation is its sensitivity to measurement conditions, making direct data transfer (DT) between LIBS systems with different analytical systems impractical. Addressing this challenge directly would require costly studies, extensive sample analysis or simulations of plasma formation in different atmospheres, moreover this approach would demand extensive calibration across various LIBS systems. Previous studies have demonstrated that machine learning models can facilitate DT across different instruments and conditions. However, existing approaches either rely on one-to-one spectral pairs or are limited to predefined condition pairs . We propose an alternative solution: a single machine learning model capable of many-to-many transfer across multiple conditions without requiring both one-to-one spectral representations and huge amounts of data. Our model has been trained on regolith LIBS spectra, measured in-house across two simulated atmospheres (Earth, Moon/vacuum) and with two laser energies (30 and 15 mJ). The model evaluation focuses on the Root Mean Square Error (RMSE) of predicted elemental concentrations from transformed spectra, serving as the primary metric for the transfer quality. The proposed model for which task outperforms Piecewise Direct Standardization (PDS) based baseline approaches by around 10% in terms of RMSE.