A matrix-insensitive classification of carbon-containing organic compounds via silhouette-optimized feature selection and neural network

The presence of carbon-containing organic matter on Mars is closely linked to the planet's potential habitability and the possibility of past life. Laser-Induced Breakdown Spectroscopy (LIBS) has emerged as a key technology in Martian organic studies due to its advantages, including real-time detection, on-site analysis, and elimination of sample preparation requirements. However, practical implementation of LIBS still faces challenges such as interference from Mars' background CO${}_2$ gas and significant chemical matrix effects caused by the planet's diverse geological environment. In this study, we constructed a laboratory-based simulation chamber that mimics the Martian environment and prepared a sample library consisting of four possible Martian matrices doped with three different organic compounds. Three matrices were used as training sets while another kind of Mars soil simulator JMSS-1 served as an independent test set to simulate the real world exploration on Mars. In order to overcome the matrix effect and influence of CO${}_2$ background, we introduced a novel strategy called SC-SKB-BPNN. Silhouette coefficent(SC) is employed to evaluate the clustering effect after the SelectKBest(SKB) feature selection and t-SNE dimensionality reduction, which enabled us to optimize the number of features and empowered the backward propagation neutral network(BPNN) classification model. Finally, after implementing the SC-SKB-BPNN algorithm, we successfully achieved precise classification of carbon-containing organic compound samples in a simulated Mars environment, attaining a sensitivity and specificity of 95.6 % and 97.8 % for the training set, while obtaining values of 93.3 % and 96.7 % for the test set.