针对熔盐堆用镍基高熵合金研发中面临的高质量数据稀缺、设计空间庞大及传统经验失效等挑战，本研究提出了一种融合大语言模型与多智能体协同系统的新型材料研发范式。我们构建了核电及高熵合金垂类大模型，处理逾9000篇文献并建立包含18000+节点的专业知识图谱与RAG知识库。通过知识图谱特征筛选与多智能体协同工作流，开发了知识嵌入的主动学习框架及小样本级联堆叠机器学习模型，有效克服了神经网络过拟合问题并提升了模型可解释性。实验表明，该模型对合金三大核心力学性能的拟合R²均超0.9，引入TabPFN与物理约束后，预测的最具潜力成分中有3种在850°C下的拉伸性能达到GH3539的两倍，并已通过高温力学性能实验验证。本研究不仅实现了抗腐蚀高温合金的自主设计，也为后续多目标（腐蚀-辐照）性能优化及结合多尺度计算提供了可靠的AI驱动平台。

Addressing challenges in developing nickel-based high-entropy alloys for molten salt reactors, such as scarce high-quality data, vast design spaces, and the failure of traditional empirical methods, this study proposes a novel materials R&D paradigm integrating Large Language Models (LLMs) with multi-agent collaborative systems. We developed a domain-specific LLM for nuclear and high-entropy alloys, processing over 9,000 papers to construct a professional knowledge graph with more than 18,000 nodes and a Retrieval-Augmented Generation (RAG) knowledge base. Through knowledge graph feature screening and multi-agent collaborative workflows, we established a knowledge-embedded active learning framework and a small-sample cascaded stacking machine learning model. This approach effectively mitigates neural network overfitting and enhances model interpretability. Results demonstrate that the model achieves R² values exceeding 0.9 for three core mechanical properties. By incorporating TabPFN and physical constraints, three of the most promising predicted compositions exhibit tensile strengths at 850°C that are double those of GH3539, which have been successfully validated through high-temperature mechanical testing. This research not only enables the autonomous design of corrosion-resistant superalloys but also provides a reliable AI-driven platform for future multi-objective (corrosion-irradiation) optimization and multi-scale computational integration.