Integrating laser-induced breakdown spectroscopy and non-linear random forest-based algorithms to predict soil unconfined compressive strength

Laser-induced breakdown spectroscopy (LIBS) is a powerful technique for elemental detection across various domains, including engineering, science, and medicine. Concurrently, machine learning's predictive process has garnered substantial attention due to its capability to forecast unknowns through trained algorithms. Crucial to geotechnical engineering, the unconfined compressive strength (UCS) of soil is a fundamental measure guiding environmental and structural designs by reflecting soil compactness and strength. Traditionally, determining UCS entails resource-intensive laboratory-based unconfined compression tests, marked by time and cost factors, as well as sensitivity to equipment quality and operator expertise. In this context, we introduce an innovative approach, leveraging machine learning algorithms to harness emission intensities of constituent elements from LIBS data. Through support vector regression (SVR) and random forest (RF) algorithms, we formulated UCS models. Rigorous evaluation encompassed standard metrics such as mean absolute error (MAE), root mean square error (RMSE), R2 value, and correlation coefficient (CC), gauging predictive performance against observed UCS values. Prominently, our findings underscored SVR's superiority, yielding 97.9% CC and 95.7% R2 during testing. Importantly, validation encompassing lime and cement-stabilized soils, hitherto unconsidered during training, showcased model accuracy and adaptability. This approach merges LIBS and machine learning, redefining UCS estimation. Beyond swift and precise predictions, it introduces cost-effective evaluations, retaining essential accuracy pivotal for geotechnical decision-making.