Deep learning-enabled semantic segmentation for high-accuracy fluid inclusions quantification in high-purity quartz

• Developed an automated method for identifying quartz fluid inclusions using four advanced semantic segmentation models. • HRNet achieved the highest segmentation accuracy with an mIoU of 81.75%. • Constructed a high-quality dataset of 9,348 microscopic images to support model training and validation. • Enabled quantitative assessment of fluid inclusions through pixel ratio analysis between quartz and inclusions. Fluid inclusions content is critical factor affecting the suitability of quartz for high-purity processing. While conventional identification methods are typically non-quantitative, time-consuming and heavily reliant on expert knowledge, developing rapid and accurate methods for quantitatively assessing fluid inclusions is of great practical importance. This study proposes the use of semantic segmentation techniques to automate the process of identifying quartz fluid inclusions with four semantic segmentation models—U-Net, DeepLabV3+, PSPNet, and HRNet. Leveraging a comprehensive dataset comprising 17 quartz samples with 9,348 meticulously annotated microscopic images, our comparative analysis demonstrates HRNet’s superior performance in fluid inclusions segmentation, achieving a remarkable mean intersection over union (mIoU) score of 81.75%, and outstanding reliability. Furthermore, this model enables the calculation of the pixel ratio of quartz and its internal fluid inclusions, thus facilitating the automated quantitative detection of fluid inclusions. This work not only expedites the quartz fluid inclusions Segmentation process but also enhances reliability to effectively overcome technical barriers in high-purity quartz resource evaluation, providing a valuable AI-driven solution for the high purity quartz industry.