Geosite evaluation is fundamental to sustainable geotourism planning, geological heritage conservation, and optimal natural resource management. However, traditional linear evaluation methods have limited capacity for analyzing the complex, multi-criteria relationships among indices influencing geotourism potential. This study presents an innovative deep learning approach to evaluate the geotourism potential of selected villages in southern Khalkhal County. Three villages Shal, Gilavan, and Jalalabad were selected as geosites and assessed using the Geosite Assessment Model (GAM) framework. Perspectives from 85 experts across five dimensions scientific/educational, aesthetic, conservation, functional, and tourism values were collected and scored. The resulting data were fed into a Deep Convolutional Neural Network (DCNN) in Python to model nonlinear relationships between sub-indices and the final score, undergoing training, testing, and validation. Results indicated that Shal geosite, due to its high diversity and clarity of geomorphic phenomena, distinctive natural landscapes, and suitable infrastructure, achieved the highest score, despite facing conservation challenges. Jalalabad, featuring karst caves and processes, ranked second, requiring enhanced scientific interpretation and visitor pressure management. Gilavan, limited by services and attraction scale, ranked third. The DCNN model demonstrated high stability and efficiency, with approximately 99% overall accuracy, a final loss of 0.01, and an AUC value of 0.99. This integrated approach provides an effective framework for prioritizing geosites for conservation, investment, and tourism management, recommending future research to employ more advanced architectures and expand the study scope.