基于生成式人工智能的火山喷发科学插画设计——三维可视化科学图像建构与优化
Generative AI-Based Scientific Illustration Design for Volcanic Eruptions—Construction and Optimization of 3D Visualized Scientific Imagery

火山喷发是一种剧烈且高度结构化的地球内部物质快速释放的现象，其多层次的动力过程与快速演化特性，对科学图像的准确再现提出了严峻挑战。传统火山演化模型图多依赖手绘或静态示意，难以兼顾三维结构的物理一致性与视觉表达力。本研究提出一种由语言驱动的图像生成策略，结合GPT-4o的语意理解与Midjourney的跨模态图像生成能力，设计出具备语意链推导与多轮优化的提示词工作流，逐步强化火山喷发图像的结构真实性与动态逻辑。研究透过五组图像版本的时空演化比对，对应提示词语意的层层建构，提出一套可追踪、可控且具教育潜力的科学图像生成流程。结果显示，当语意设计越贴近地质术语与物理因果链时，生成图像的解释力与教学适用性大幅提升。本研究展示了大语言模型在地球科学可视化中的知识转译潜力，并为未来AI辅助科普图像建构开启新的可能性。
Volcanic eruptions are characterized by the violent and highly structured rapid release of Earth’s internal matter. The multi-level dynamic processes and rapid evolutionary nature of these events pose a significant challenge to accurate scientific visualization. Traditional cross-section diagrams often fall short in capturing the three-dimensional complexity and causal dynamics required for educational use. This study introduces a language-to-visualization workflow that integrates the semantic reasoning of GPT-4o with the cross-modal image synthesis capabilities of Midjourney. We designed a prompt workflow incorporating semantic chain deduction and iterative optimization to progressively reinforce the structural authenticity and dynamic logic of volcanic eruption imagery. By comparing the spatiotemporal evolution of five sets of image versions, corresponding to the layered construction of prompt semantics, we establish a traceable, controllable, and pedagogically valuable framework for scientific image generation. Results indicate that when prompt semantics are closely aligned with geological terminology and physical causal chains, the explanatory power and educational applicability of the generated images are significantly enhanced. This study demonstrates the potential of Large Language Models (LLMs) in knowledge translation for geoscience visualization, opening new avenues for future AI-assisted construction of popular science imagery.