The GEOQInsight project: leveraging AI for efficient geochemistry decision-making

Geochemical decision-making in exploration and production (E&P) is affected by significant uncertainty, driven by the need to reconcile extensive internal proprietary knowledge bases with a continuously evolving scientific literature, where different and often contradictory interpretations coexist for the same geochemical indicators. To mitigate these bottlenecks and enhance interpretive rigor, the GEOQInsight project develops an intelligent ecosystem based on a multi-agent architecture, knowledge graphs, and language models for specialized decision support. The data architecture relies on a public domain document repository designed to be continuously and incrementally scalable: starting from 4,636 seed terms covering multiple geochemical subdomains (in constant expansion), articles are systematically ingested via OpenAlex, initially comprising around 70,000 processed documents with automated content curation and structured extraction of PDF sections and tables. This enables a hybrid retrieval strategy that couples lexical search with dense vectors for semantic retrieval across nearly 2 million indexed passages. Concurrently, an ontologically grounded knowledge subgraph for Source Rock Evaluation was developed to formally represent both conceptual knowledge and quantitative relationships among fundamental source-rock properties (Lithology, Thermal Maturity, Quality, and Richness) and the analytical parameters used to characterize them, including TOC, S1, S2, S2/S3, HI, PI, Tmax, Ro, and IR.

Beyond a static catalog, this graph encodes value ranges and cut-off thresholds that deterministically validate geological hypotheses (such as kerogen types I to IV, generated product oil vs. gas, maturity stages from immature to postmature, and richness classes), serving as a formal reasoning layer to cross-examine analytical data against established domain rules. As current results, the system features a functional conversational agent capable of querying this indexed public base and dynamically extracting textual and tabular data from user-uploaded documents. The research is advancing toward the secure integration of internal corporate databases, the grounding of agent reasoning on knowledge graph topology to enhance logical consistency, traceability, and auditability, the training and validation of domain-specific language models for geochemistry to improve technical accuracy while reducing inference costs, and the modular extension of the framework to other enterprise workflows. In conclusion, GEOQInsight establishes a robust semantic infrastructure that transforms heterogeneous literature and geochemical data into rapid, explainable, and technically consistent decisions, supporting geochemists in the preparation of a wide range of technical reports and interpretations.