GEOLITHAI AS AN AI-ASSISTED WORKFLOW FOR LITHOLOGICAL RECOGNITION AND GRADE CONTROL OPTIMIZATION IN AN UNDERGROUND PERUVIAN MINE M. Vallejo Pablo1 1Department of Geological Engineering, Universidad Nacional Mayor de San Marcos, Peru, (*Presenting author: griselda.archivos@gmail.com*) ABSTRACT This paper presents GeoLithAI as an AI-assisted research workflow for lithological recognition and short-cycle geological mapping in an anonymized underground Peruvian mine. The operational problem addressed is that face mapping is commonly performed under low illumination, shadows, wet surfaces, support mesh, painted markers and fragmented rock, all of which complicate visual interpretation and delay grade-control model updates. The proposed workflow combines standardized RGB image acquisition, optional spatial support from LiDARderived geometry, supervised deep learning and geologist review to delineate lithological domains and flag ambiguous areas for validation. The core method links semantic segmentation and object detection. Segmentation is used to outline visually coherent lithological zones, whereas object detection is used to identify occlusions and operational elements such as mesh, support components, water reflections, paint marks and fragmented material that can bias interpretation. A first pilot cycle is defined with five domain classes relevant to carbonate- and skarn-influenced underground environments: carbonate-dominated, skarn-rich, mixed transition, brecciated, and altered or wet reflective zones. Transfer learning from drill-core imagery is proposed to reduce labelling effort and improve robustness when annotated face images are limited. The workflow is intended to support, rather than replace, the geologist by improving repeatability, traceability and update frequency of geological interpretations. The expected engineering value is safer data acquisition, faster mapping support, and improved integration with 3D geological modelling. The contribution is framed as an anonymized scientific workflow for digital geological mapping in Peru and not as a commercial product case. KEYWORDS Lithological recognition, underground mine mapping, deep learning, semantic segmentation, object detection, grade control 1. INTRODUCTION AND PRACTICAL PROBLEM The World Mining Congress 2026 theme emphasizes the need to deliver minerals faster, smarter and more responsibly. In operating mines, this challenge is directly connected to the quality and update frequency of geological information used for short-term planning and grade control. Geological face mapping remains a high-value activity because it captures lithological contacts, alteration, brecciation, structural discontinuities and visual mineralization features that are not always represented at the resolution required by production decisions. The difficulty is that underground geological mapping is commonly performed in conditions that are not ideal for visual interpretation. Low illumination, shadows, wet reflective surfaces, support mesh, scaling marks, fragmented material and access restrictions may obscure
RkJQdWJsaXNoZXIy MTM0Mzk2