AI for ScienceAI for Science
科学数据昂贵又稀缺，是“数据决定上限”最极端的场景。关注化学、材料与蛋白质，以及模型怎样在很少的实验数据上被训练和验证。Scientific data is expensive and scarce — the most extreme case of "data sets the ceiling." My focus is chemistry, materials, and proteins, and how models are trained and validated on very little experimental data.
- 在关注Following
- 当前状态Current status
- V
- 造物之书里的实验Experiments in the book
在科学里，一个配平错误或一个不存在的反应条件，都会让模型带着同样的错。科学 AI 首先是数据质检问题。In science, one unbalanced equation or one impossible reaction condition teaches the model the same mistake. AI for science is, first of all, a data-QA problem.
做过化学合成数据的质检与错误归因，以及化学与学科资料的证据寻源。这是从数据角度进入科学 AI，还没有做过模型层面的科学实验。I have done QA and error attribution on chemical synthesis data, as well as evidence sourcing for chemistry and discipline-specific material. That is entering science AI from the data side; I have not yet run science experiments at the model level.
- 01
化学合成数据里，模型最常犯的是哪一类错？规则与工具能拦下多少？What errors does a model make most often in chemical synthesis data, and how many can rules and tools catch?
- 02
只有几百条实验数据时，预训练模型 + 少量微调能走多远？With only a few hundred experimental records, how far can a pretrained model plus light fine-tuning go?
- 03
科学问答的“正确”怎样被验证：文献、计算，还是实验？How is "correct" verified in scientific Q&A: literature, computation, or experiment?
- 04
蛋白质与分子生成模型的评测，和语言模型的评测有什么根本不同？How does evaluating protein and molecule generative models fundamentally differ from evaluating language models?
现在还不能证明什么What cannot be proven yet
- 化学方向的实践集中在数据质检，不是领域专家Chemistry practice is concentrated on data QA; I am not a domain expert
- 尚无公开的科学 AI 实验No public science-AI experiment yet