Interpreting convolutional neural network classifiers applied to laser-induced breakdown optical emission spectra.

Area of Science:

• Spectroscopy
• Machine Learning
• Planetary Science

Background:

• Laser-induced breakdown spectroscopy (LIBS) is a key analytical technique.
• Artificial neural networks (ANNs) are increasingly used with LIBS for pattern recognition.
• ANNs are often treated as black-box models, limiting understanding of their predictions.

Purpose of the Study:

• To interpret the decision-making process of convolutional neural networks (CNNs) used in LIBS.
• To identify the spectroscopic features learned by CNNs for elemental analysis.
• To understand how CNNs contribute to LIBS data interpretation.

Main Methods:

• Applied post-hoc interpretation techniques to CNNs trained on LIBS data.
• Generated synthetic spectra (prototype spectra) to understand CNN predictions.
• Investigated a simple CNN architecture trained on the ChemCam LIBS dataset from the Curiosity Mars rover.

Main Results:

• CNNs predominantly learned meaningful spectroscopic features corresponding to elements in major oxides.
• Identified class-specific prototype spectra that perfectly predict classification.
• The convolution operation in the CNN provided a crude baseline correction for spectra.

Conclusions:

• CNNs applied to LIBS data can learn interpretable spectroscopic features.
• This interpretation provides insight into ANN decision-making in spectroscopy.
• The findings enhance the utility of ANNs for LIBS analysis, including planetary exploration.