Interpreting artificial neural networks used in LIBS

Interpreting artificial neural networks used in LIBS

Abstract

The use of multivariate data-based models has become synonymous with modern LIBS analysis, be it qualitative or quantitative [1]. Among the many multivariate models commonly used in spectroscopy, artificial neural networks (ANNs) enjoy a rising popularity mainly fuelled by the various ground-breaking achievements in various fields. Despite the superior performance of ANNs exhibited for both quantitative (regression analysis) and qualitative (classification) analyses, the use of ANNs faces a major limitation. Namely, while classical statistical models—such as partial least squares-based methods—offer an inherent interpretability of their predictions, ANNs are generally used as black-box models. Consequently, we carried out the post-hoc interpretation of the most prevalent ANN architectures found in the LIBS literature, i.e., fully connected ANNs (commonly referred to as multilayer perceptrons, MLPs) [2] and convolutional neural networks (CNNs) [3] trained for classification tasks. Both models were interpreted by finding optimal input spectra that represent the individual classes. Consequently, the spectral features of these optimal spectra were explored and related to the chemical composition characteristic of the classes. We found that both ANN architectures are capable of learning meaningful spectroscopic features.

The use of multivariate data-based models has become synonymous with modern LIBS analysis, be it qualitative or quantitative [1]. Among the many multivariate models commonly used in spectroscopy, artificial neural networks (ANNs) enjoy a rising popularity mainly fuelled by the various ground-breaking achievements in various fields. Despite the superior performance of ANNs exhibited for both quantitative (regression analysis) and qualitative (classification) analyses, the use of ANNs faces a major limitation. Namely, while classical statistical models—such as partial least squares-based methods—offer an inherent interpretability of their predictions, ANNs are generally used as black-box models. Consequently, we carried out the post-hoc interpretation of the most prevalent ANN architectures found in the LIBS literature, i.e., fully connected ANNs (commonly referred to as multilayer perceptrons, MLPs) [2] and convolutional neural networks (CNNs) [3] trained for classification tasks. Both models were interpreted by finding optimal input spectra that represent the individual classes. Consequently, the spectral features of these optimal spectra were explored and related to the chemical composition characteristic of the classes. We found that both ANN architectures are capable of learning meaningful spectroscopic features.

machine learning; interpretability; support vector machines; spectroscopic data; convolutional neural networks

machine learning; interpretability; support vector machines; spectroscopic data; convolutional neural networks

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

05.09.2022