Improving laser-induced breakdown spectroscopy regression models via transfer learning

Improving laser-induced breakdown spectroscopy regression models via transfer learning

Článek WoS

Laser-induced breakdown spectroscopy (LIBS) is a well-established analytical tool with relevance in extra-terrestrial exploration. Despite considerable efforts towards the development of calibration-free LIBS approaches, these are currently outperformed by calibration-based approaches to semi-quantitative LIBS analyses. However, the construction of robust calibration models often requires large calibration datasets owing to the extensive matrix effects plaguing the LIBS performance. Moreover, LIBS data are sensitive to changes in the apparatus. Hence, a calibration model constructed for one LIBS system is seldom applicable to a distinct LIBS system. A notable example are the LIBS instruments included in the currently active Mars Rovers' analytical suites, the ChemCam and SuperCam LIBS instruments; while the two instruments exhibit relatively small differences, they required the collection of two separate calibration datasets. Currently, these two datasets are used exclusively for the system they were collected for. In this work, we demonstrate that calibration models constructed for the SuperCam instrument can be improved using data obtained with the ChemCam instrument. Namely, we take advantage of the partial overlap between the targets used to collect the two calibration datasets. Using this overlap, we approximate the function transforming ChemCam spectra into their SuperCam equivalent. Subsequently, the transformed spectra are used to extend the training data available for the regression model constructed for the SuperCam instrument. The proposed approach considerably improves the performance of convolutional neural network regression models.

Laser-induced breakdown spectroscopy (LIBS) is a well-established analytical tool with relevance in extra-terrestrial exploration. Despite considerable efforts towards the development of calibration-free LIBS approaches, these are currently outperformed by calibration-based approaches to semi-quantitative LIBS analyses. However, the construction of robust calibration models often requires large calibration datasets owing to the extensive matrix effects plaguing the LIBS performance. Moreover, LIBS data are sensitive to changes in the apparatus. Hence, a calibration model constructed for one LIBS system is seldom applicable to a distinct LIBS system. A notable example are the LIBS instruments included in the currently active Mars Rovers' analytical suites, the ChemCam and SuperCam LIBS instruments; while the two instruments exhibit relatively small differences, they required the collection of two separate calibration datasets. Currently, these two datasets are used exclusively for the system they were collected for. In this work, we demonstrate that calibration models constructed for the SuperCam instrument can be improved using data obtained with the ChemCam instrument. Namely, we take advantage of the partial overlap between the targets used to collect the two calibration datasets. Using this overlap, we approximate the function transforming ChemCam spectra into their SuperCam equivalent. Subsequently, the transformed spectra are used to extend the training data available for the regression model constructed for the SuperCam instrument. The proposed approach considerably improves the performance of convolutional neural network regression models.

SUPPORT VECTOR MACHINES; CHEMCAM INSTRUMENT; NEURAL-NETWORKS; CCD DETECTORS; ALLOY-STEEL; LIBS; TEMPERATURE; SCIENCE; SYSTEM; SAMPLE

SUPPORT VECTOR MACHINES; CHEMCAM INSTRUMENT; NEURAL-NETWORKS; CCD DETECTORS; ALLOY-STEEL; LIBS; TEMPERATURE; SCIENCE; SYSTEM; SAMPLE

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

2023

31.08.2022

ROYAL SOC CHEMISTRY

CAMBRIDGE

1364-5544

JOURNAL OF ANALYTICAL ATOMIC SPECTROMETRY

37

9

Spojené království Velké Británie a Severního Irska

1883

1893

11

https://pubs.rsc.org/en/content/articlelanding/2022/JA/D2JA00180B