Boson20/matter-to-mechanism
收藏资源简介:
Matter to Mechanism 是一个包含2,645个专家整理的问题-假设对的结构化数据集,这些数据对来自材料科学和电化学领域的同行评审文献。每个条目将一个精确陈述的研究问题与其对应的科学假设联系起来,包括逐步推理链、提出的干预措施、机制原理和目标属性。该数据集是BatteryHypoBench的基础,这是一个多维基准,用于评估AI共同科学家系统在科学假设生成方面的表现,不仅测量表面流畅性,还测量机制深度、推理连贯性和物理合理性。
Matter to Mechanism contains 2,645 expert-curated problem–hypothesis pairs extracted from peer-reviewed materials science and electrochemistry literature. Each entry links a precisely stated research problem to its corresponding scientific hypothesis, complete with a step-by-step reasoning chain, proposed intervention, mechanistic rationale, and target property. The dataset is the foundation of BatteryHypoBench, a multi-dimensional benchmark for evaluating AI co-scientist systems on scientific hypothesis generation — measuring not surface fluency but mechanistic depth, reasoning coherence, and physical plausibility.




