A performance compensation method for miniaturized LIBS instruments in precise rock classification

Abstract

Rocks, as one of the most ubiquitous substances on Earth, bear the historical records of Earth's evolution and traces of geological processes. Rock identification can reveal the geological environment and resource distribution, which are important for geological research. Laser-induced breakdown spectroscopy (LIBS), a real-time, micro-destructive, and multi-element simultaneous analysis technique, has played an important role in rock identification and geological research. However, traditional laboratory-based LIBS instruments are bulky and unsuitable for field site geological exploration. While handheld LIBS instruments offer compactness and convenience, they suffer from reduced laser energy and spectrometer performance, potentially compromising analytical accuracy. There is an urgent demand for a miniaturized LIBS instrument that retains the high analytical capabilities of laboratory-based LIBS while incorporating the portability of handheld LIBS. This study proposes a compensation method for the performance loss of miniaturized LIBS instruments due to their reduced size. We designed six experimental setups of different sizes to compare rock spectra and classification accuracy in detail, aiming to validate the performance loss of a miniaturized LIBS instrument. In our experiments, the miniaturized LIBS instrument, equipped with an MPL-H-1064 laser and an AvaSpec Mini2048 spectrometer, was employed to classify 16 types of rocks using the SVM model, achieving an initial classification accuracy of 77.08%. To compensate for the performance loss inherent in miniaturized LIBS instruments, a range of preprocessing methods and principal component analysis (PCA) were employed to enhance spectral quality and elevate the accuracy of rock classification to 96.25%. Additionally, the Optuna framework was used to automatically search for the optimal hyperparameters of the SVM model, subsequently increasing the accuracy of rock classification to 99.58%. The results demonstrate that this method effectively mitigates the performance loss of miniaturized LIBS instruments and achieves precise rock classification.