Scientists have developed an AI-powered "Cyber Academia-Chemical Engineering" system—a digital academic town where specialized AI agents autonomously collaborate to drive scientific discovery in chemical engineering. Led by Professor Zhou Tianhang from China University of Petroleum (Beijing), the team created a multi-agent research environment that mimics a real scientific community, with AI experts in molecular design, pilot-scale process development, engineering validation, experimental research, theoretical modeling, process safety, and quality control. These agents don’t just respond to queries—they interact, debate, propose hypotheses, and autonomously design research pathways from theory to experimentation and process optimization. Unlike traditional AI systems that solve predefined problems, this system enables emergent scientific discovery through continuous interaction and cross-domain collaboration. It operates on a novel principle: instead of waiting for human-defined questions, the AI agents themselves identify gaps in knowledge and explore uncharted scientific territories. This marks a shift from reactive AI to proactive, self-evolving research. The project emerged from the team’s belief that while many AI for Science initiatives remain theoretical, real-world industrial applications—especially in critical sectors like chemical engineering, which contributes 17% to China’s GDP—demand tangible, autonomous innovation. The team, led by Academician Xu Chunming, a veteran in petroleum and petrochemicals, and supported by the NSFC-CAS joint project on “Low-Carbon Intelligent Chemical Engineering,” aimed to bridge the gap between AI and industrial science. Zhou Tianhang, who returned from seven years of study in Germany, and his young, gaming-savvy team were inspired by Stanford’s AI Town, where AI agents simulate social interactions. They asked: what if we built a specialized academic town where AI experts could autonomously collaborate across disciplines, just like real scientists? To make this vision real, the team first identified seven core expert roles, each covering a stage in the chemical R&D pipeline. After deploying these agents, they ran the system for three days, collecting 1,200 rounds of dialogue. However, they quickly noticed a major flaw: the agents suffered from "hallucinations"—they drifted into abstract philosophical discussions instead of staying grounded in technical rigor. To address this, the team implemented a three-tier knowledge enhancement framework: retrieval-augmented generation (RAG), domain-specific fine-tuning, and knowledge graph integration. Each expert was equipped with a curated, domain-specific knowledge base. Evaluation showed that this approach improved response accuracy, relevance, and problem-solving capability by 10%–15% across all agents, ensuring decisions were based on verifiable scientific evidence. Yet a new challenge emerged: poor collaboration between experts from distant domains. For example, molecular designers and safety engineers struggled to communicate due to semantic misunderstandings—what one called “stability,” the other interpreted as “reaction rate.” This created a “semantic gap” that hindered progress. The solution came from ontology engineering—a formal method of defining concepts and relationships within a domain. The team introduced a Collaboration Agent (CA) equipped with a shared domain ontology and coordination strategies. This agent acted as a linguistic and conceptual bridge, aligning terminology and enabling meaningful cross-disciplinary dialogue. As Zhou Tianhang explained, “The breakthrough isn’t just deeper knowledge in individual agents. It’s about building semantic infrastructure so isolated experts can transcend their silos and co-evolve intelligent solutions.” When diverse perspectives operate within a unified conceptual framework, hidden connections—what the team calls “dark knowledge”—emerge. These are subtle, non-obvious relationships buried in data that humans often miss. The system’s ability to uncover them could redefine scientific discovery, enabling AI to pioneer innovations beyond human intuition. The team’s work builds on prior achievements, including the development of FlowBD—a specialized large model for iron-chromium redox flow batteries—and the “Molecule-to-Grid Industrial Agent,” both born from their spin-off company, Zhonghai Energy Storage. This company, focused on long-duration energy storage, will serve as the real-world testbed for the Cyber Academia system, validating its industrial impact. The research has been published on arXiv (arXiv:2510.01293), marking a significant step toward autonomous, AI-driven scientific innovation in industry.