Transfer learning as a potential tool for improving LIBS analytical performance for space exploration

abstract

English

1. Introduction Laser-induced breakdown spectroscopy (LIBS) is a prominent atomic emission spectroscopic method which is frequently used for extra-terrestrial exploration, e.g., as part of the Curiosity, Perseverance, and Zhurong Mars rovers owing to LIBS’s ability to provide real-time analysis with limited or no sample preparation and in most cases from a distance. However, quantitative LIBS analysis in most cases requires regression models. Recently, there has been a shift from the physics-based univariate regression models to the more versatile multivariate calibration models, often based on machine learning. In addition, LIBS is generally characterized by low repeatability and a considerable susceptibility to various matrix effects. Consequently, constructing calibration models is generally costly. In addition, LIBS calibration datasets are limited to the instrumentation with which they were obtained. That is, the calibration models are tied to the spectral range, resolution, laser source, collection and ablation geometries, detector, and detection settings of the LIBS system that was used for collecting the calibration data. will also have a significant impact on the shape of LIBS spectra. Consequently, changes in the instrumentation and measurement conditions will decrease the calibration model performance leading to a low prediction accuracy and biased results. In this work, we demonstrate the possibility of improving the analytical performance of a LIBS system by using data obtained on a different LIBS instrumentation. Namely, we demonstrate that a regression model built for the SuperCam calibration dataset can be improved by using the ChemCam calibration dataset. This improvement is shown by reducing the test prediction error (in terms of root mean square error) achieved on the SuperCam calibration dataset. Our methodology relies on finding a direct transfer function between the ChemCam and SuperCam datasets parameterized by a fully connected artificial neural network (ANN). 4. Conclusion The aim of the work was to apply transfer learning to improve machine learning-based regression models used in LIBS. This was demonstrated by improving a convolutional neural network regression model applied to the SuperCam dataset by extending the available training dataset with ChemCam spectra adequately transformed using a fully connected ANN. Performance improvements were obtained in the case of the prediction of every major oxide content of homogenized geological targets. Nevertheless, the proposed methodology shows increased risks of overfitting on the training data, despite optimized regularization.

Space exploration; Mars rovers; Laser-induced breakdown spectroscopy; Machine learning; Transfer learning; Regression analysis

KÉPEŠ, E.; VRÁBEL, J.; POŘÍZKA, P.; KAISER, J.

22. 12. 2022

1