Integrating deep learning and Raman spectroscopy for automated sandstone petrographic analysis

A comprehensive petrographic analysis of rocks and minerals is crucial for addressing global energy crises and water resource challenges. Sandstone, a predominant reservoir for natural gas, petroleum, and groundwater, requires advanced petrographic analysis techniques to overcome existing limitations such as inefficiency, lack of automation, low differential accuracy, and complex data interpretation. Here, we develop an intelligent deep learning−Raman spectroscopy system that integrates Raman spectroscopy and digital image processing with a convolutional neural network to achieve pixel-level fusion of high-resolution microscopic images and Raman spectral signatures. This framework introduces a dual-modal petrographic workflow where instance segmentation-identification and spectral calibration operate cooperatively rather than independently, enabling fully automated mineral identification with a spatial resolution of 1 μm. By coupling this algorithmic framework with a compact and low-cost Raman microscope designed for automated data acquisition, the system achieves a 40-fold improvement in analysis efficiency and reaches 98% identification accuracy, while reducing instrumentation cost to one-tenth of that of conventional systems. This research demonstrates a novel intelligent end-to-end petrographic analysis technique that transforms sandstone characterization from experience-based judgment to scalable, automated, and high-precision mineral screening.