CSTrainer : An Ontology Driven Multiple AIGC Agentic Approach for Oil and Gas Scientific Research Supporting
Abstract
Abstract Scientific research in the energy and oil & gas sector faces significant challenges, including massive heterogeneous data, complex mechanisms, high uncertainty, and substantial reliance on expert experience. While recent advancements in generative AI offer autonomy in multi-step reasoning, generic models offered limitations such as hallucinations, insufficient domain knowledge, and poor alignment with researchers' prior works. To address the need for continuity in research trajectories, this study proposes CSTrainer, an ontology-driven, multi-agent system designed for energy-specific scientific research assistance. CSTrainer integrates 5 foundational models (GPT-4o, DeepSeek, Qwen, Tencent-Hunyuan, and Doubao-Seed (five foundation LLMs)) and employs a hierarchical architecture featuring multi-agent clusters, cross-model voting, and debate mechanisms. A core innovation is its use of a distributed dynamic ontology framework, which constrains reasoning paths to ensure scientific rigour and mitigate hallucinations. The system facilitates end-to-end research support, encompassing scientific problem formulation, technical route planning, ontology construction, and executable workflow generation, augmented by weighted Bayesian updating and lightweight human-in-the-loop (HIL) feedback. The system was validated on 3 datasets: UNISIM-IV (open), tight gas well re-evaluation (proprietary), and transnational pipeline scheduling (proprietary). Results demonstrate that CSTrainer achieves stable, high quality performance for 3–5 scientific questions extraction, outperforming agent-only baselines by 62.5%–87.5%. While 10 rounds of detailed human feedback yielded optimal results, excessive interaction or scaling beyond 10 problems induced instability. The system attained full marks in the tight gas re-evaluation scenario, confirming its efficacy in bridging the gap between conceptual research and engineering implementation. CSTrainer provides a scalable, interpretable, and controllable AI-for-science pathway for industrial research. Keywords: Ontology; Scientific research supporting; Energy engineering research; Multi-Agent; Generative AI
