Abstract
Mineral prospectivity mapping increasingly relies on machine learning to integrate geochemical and geological information, yet purely data-driven models often perform unreliably under sparse sampling, strong class imbalance, and complex geological controls, while traditional knowledge-driven approaches remain difficult to formalize and scale. This paper presents a neuro-symbolic prospectivity framework that incorporates geological knowledge extracted from deposit-model literature using large language models and embeds it into a data-driven learning pipeline. The framework integrates multi-element geochemical data, spatial geological features, and embedding-based geological priors derived from 87 ore deposit models, where the priors encode semantic relationships between mapped geological environments and mineral system concepts and act as soft guidance rather than deterministic constraints. Model performance is evaluated using spatial cross-validation and area-based targeting metrics that reflect exploration decision-making under limited spatial footprints. A case study of the CUMO porphyry Cu-Mo district (Boise County, Idaho, USA) shows that the proposed framework improves early-stage prospectivity targeting, particularly for joint Cu-Mo anomalies and under strong area constraints, while producing spatial prospectivity patterns that are consistent with mapped lithological and structural controls. The results demonstrate how LLM-assisted neuro-symbolic modeling can bridge geological reasoning and machine learning, supporting interpretable and decision-relevant mineral exploration workflows.