Abstract:

Mineral resource prediction is a key area of research in mathematical geoscience, involving the integration of diverse, cross-disciplinary geoscientific datasets. This process involves significant computational complexity and workload, along with semantic inconsistencies, which pose significant challenges to researchers. The emergence of next-generation generative artificial intelligence, particularly large language models (LLMs) and intelligent agents, is catalyzing transformative advances across industries, facilitating the transition toward intelligent mineral prediction. This study proposes a super-agent framework for AI-driven mineral resource forecasting, built upon multimodal large language models (MLLMs) (e.g., DeepSeek, Qianwen) and agent orchestration frameworks. The super-agent framework comprises a management agent and multiple specialized agent groups (geological, geophysical, geochemical, remote sensing), each comprising multiple atomic agents or lightweight agent collectives. Individual agents are capable of accessing and invoking specific tools—including those locally deployed, cloud-based, or dynamically generated—as well as datasets and services. Upon receiving external prediction tasks, the management agent dynamically coordinates the workflow. This involves orchestrating specialized agent groups, atomic agents (e.g., for geochemical map generation), analytical tools (e.g., interpolation algorithms), and relevant data sources, executing tasks either sequentially or in parallel as needed. Using geochemical map generation as a case study, we elucidate the internal mechanisms of agent-model collaboration, enabling one-click generation of numerous geochemical element maps (up to 39 different elements), thereby validating the effectiveness of this AI-driven approach. By seamlessly integrating multimodal LLMs, intelligent agents, and domain-specific mineral prediction workflows, this framework enables zero-code operation with multimodal interaction through natural language (text or voice), and presents a promising approach for establishing a new paradigm in intelligent mineral resource prediction.

Key words: mineral resources prediction, intelligent prediction, large models, large language models, agent, super-agents