📄 中文摘要
合成数据在文本领域的科学推理中展现出显著效果，然而，多模态推理仍受限于合成科学严谨图像的挑战。现有的文本到图像（T2I）模型常生成在视觉上看似合理但科学上不准确的图像，导致持续存在的视觉-逻辑分歧，这严重限制了其在下游推理任务中的价值。为解决这一问题，迫切需要对科学图像合成进行系统性的基准测试、方法论开发及实用性评估。构建能够生成高保真度、科学准确图像的模型，需要深入理解不同科学领域（如生物医学、材料科学、天文学等）的特定数据特征和领域知识。这包括开发新的生成对抗网络（GANs）、扩散模型（Diffusion Models）或变分自编码器（VAEs）架构，使其能够融合符号知识、物理定律或生物机制。此外，评估合成图像的科学有效性需要设计新的度量标准，超越传统的视觉质量指标，涵盖科学准确性、数据分布匹配度、以及对特定科学分析任务的贡献。例如，在生物医学图像合成中，模型不仅要生成形态逼真的细胞图像，还要确保细胞器结构、蛋白质表达水平等生物学特征符合实际。在材料科学中，合成的晶体结构图像需满足晶体学对称性和原子间距的物理约束。这些合成图像的下游应用潜力巨大，例如用于扩充稀缺的科学数据集、辅助新科学现象的预测与探索、或作为教学与可视化工具。通过提供一个全面的框架来评估和改进科学图像合成技术，可以推动多模态AI在科学发现和研究中的应用，弥合视觉表象与科学真理之间的鸿沟。
📄 English Summary
Scientific Image Synthesis: Benchmarking, Methodologies, and Downstream Utility
While synthetic data has proven effective for improving scientific reasoning in the text domain, multimodal reasoning remains constrained by the difficulty of synthesizing scientifically rigorous images. Existing Text-to-Image (T2I) models often produce outputs that are visually plausible yet scientifically incorrect, resulting in a persistent visual-logic divergence that limits their value for downstream reasoning. Addressing this critical gap necessitates a systematic approach to benchmarking, methodology development, and utility assessment for scientific image synthesis. Developing models capable of generating high-fidelity and scientifically accurate images requires a deep understanding of specific data characteristics and domain knowledge across various scientific fields, such as biomedicine, materials science, and astronomy. This involves proposing novel architectures for Generative Adversarial Networks (GANs), Diffusion Models, or Variational Autoencoders (VAEs) that can incorporate symbolic knowledge, physical laws, or biological mechanisms. Furthermore, evaluating the scientific validity of synthetic images demands the design of new metrics that transcend traditional visual quality indicators, encompassing scientific accuracy, data distribution congruence, and contributions to specific scientific analysis tasks. For instance, in biomedical image synthesis, models must not only generate morphologically realistic cell images but also ensure that subcellular structures and protein expression levels align with biological reality. In materials science, synthetic crystal structure images need to satisfy crystallographic symmetries and interatomic distance physical constraints. The downstream utility of these synthetic images is immense, including augmenting scarce scientific datasets, aiding in the prediction and exploration of novel scientific phenomena, or serving as powerful educational and visualization tools. Establishing a comprehensive framework for evaluating and enhancing scientific image synthesis technologies will advance the application of multimodal AI in scientific discovery and research, effectively bridging the chasm between visual plausibility and scientific truth.