A highly efficient and interpretable framework for high-precision lithological identification integrating SERS and XGBoost

Rapid and accurate lithological identification is essential for geological exploration and deep earth engineering. However, traditional methods are often limited due to time-consuming procedures, destructive sampling, and subjective experience. In this study, a high-precision lithological classification framework was proposed by integrating Surface-Enhanced Raman Scattering (SERS) technology with artificial intelligence (AI) algorithms. Silver nanoparticles (AgNPs) with an average radius of 32 nm were synthesized as substrates to enhance Raman signals and suppress fluorescence backgrounds. Finite-Difference Time-Domain (FDTD) simulations confirmed the electromagnetic enhancement mechanism, demonstrating substantial localized electric field amplification at the generated hotspots. Six AI models were evaluated using a self-constructed SERS dataset encompassing eight distinct rock types. The Extreme Gradient Boosting (XGBoost) model achieved the best performance, mitigating the high computational latency and training instability commonly observed in deep learning models. Moreover, SHapley Additive exPlanations (SHAP) was utilized to interpret the classification process. The results demonstrated that the model effectively extracts intrinsic physical features, including low-frequency lattice vibrations and SERS-activated trace signals, to accurately distinguish rocks with highly similar spectra. This study provides a reliable, interpretable, and computationally efficient approach for real-time geological monitoring in complex construction environments.