RECONSTRUCTING CLIMATE TRENDS IN SYNOROGENIC BASINS IN THE CENTRAL ANDES VIA MACHINE-LEARNING-ENABLED LIBS (LASER-INDUCED BREAKDOWN SPECTROSCOPY)

Recent advances in LIBS (laser-induced breakdown spectroscopy) technology have allowed the development of handheld, battery-operated devices that can analyze chemical compositions of solid materials in remote settings. LIBS collects atomic emission spectra and relates spectral peaks at specific wavelengths to particular elements. Compared to pXRF devices which are common in geoscience, portable LIBS has a wider analytical range including lightweight elements such as C and Na, and lower limits of detection on other elements of particular interest to geologists such as K, Ca and Al. However, use of LIBS for whole-rock geochemistry is difficult due to the chemical heterogeneities, multiple ionization states per element and spectral interference from multiple co-varied elements that occur in rocks. Conventional calibration methods for converting LIBS spectra into compositions fail in providing quantitatively useful accuracy on geological materials.

Here we demonstrate a novel approach to LIBS analysis of major element composition in geological materials. We trained a backwards-propagating artificial neural network (BP-ANN) to calculate percent concentrations of the 10 most abundant elements using a suite of sedimentary and igneous rocks of known composition as well as a large synthetic dataset. Results were compared to the conventional linear calibration-curve method and a partial least squares regression (PLSR)-based calibration. Accuracy was tested using an isolated set of 34 samples. For both trend recognition and total accuracy, the BP-ANN outperformed both alternative methods. For the 10 elements analyzed per sample, average R² increased from 0.73 for PLSR, to 0.95 for the linear model and 0.99 for the BP-ANN. The average sum of squared residuals decreased from 610 for the linear model to 251 for PLSR and 16.5 for the BP-ANN.

We apply this calibration to paleoclimate reconstruction via LIBS analysis of sedimentary rocks in the Central Andes of South America. LIBS data come from more than 400 sedimentary samples from 14 Cretaceous-Pliocene stratigraphic sections spanning the forearc, Altiplano, and retroarc foreland basins in western Bolivia and southern Peru. We use these data to reconstruct climate trends in these basins throughout the period of Andean orogenesis, providing insights into the relationship between mountain building and local climate.