EXPERIMENT V · AI FOR SCIENCE

化学数据里的错Errors in Chemical Data

化学合成数据里，模型最常犯的是哪一类错？In chemical synthesis data, which kind of error does a model make most often?

待开始Not started化学数据Chemical data数据质检Data QARDKitRDKit
  1. Q提出可证伪的问题State a falsifiable question当前阶段Current stage
  2. R做最小运行路径Build a minimal runnable path未开始Not started
  3. E连接真实验证Validate against reality未开始Not started
  4. P公开证据与限制Publish evidence and limits未开始Not started

为什么值得做Why it is worth doing

科学数据昂贵又稀缺，是“数据决定上限”最极端的场景。合成数据能补数量，但一个配平错误或不存在的反应条件，会让训练出来的模型带着同样的错。Scientific data is expensive and scarce — the most extreme case of "data sets the ceiling." Synthetic data adds volume, but one unbalanced equation or one impossible reaction condition makes the trained model carry the same mistake.

Q

假设Hypothesis

HYPOTHESIS

化学合成数据中的错误集中在少数几类（配平、单位、反应条件、不存在的物质），用规则 + 工具校验能拦下其中大部分。Errors in chemical synthesis data cluster in a few types (balancing, units, reaction conditions, nonexistent substances), and rules plus tool checks catch most of them.

R

最小装置Minimal setup

MINIMAL SETUP
  1. 01从公开化学问答与合成数据中抽样，人工标注错误类型Sample from public chemistry Q&A and synthesis data and label error types by hand
  2. 02为每类错误写校验：RDKit 结构检查、配平、单位换算Write a check for each error type: RDKit structure checks, balancing, unit conversion
  3. 03比较只用模型评审 / 只用工具校验 / 两者结合的检出率Compare detection rates for model judge only / tool checks only / the two combined
E

什么会证明我错了What would prove me wrong

FALSIFIED IF

错误类型高度分散，工具校验的检出率不高于模型评审。If error types are widely dispersed and tool checks do not beat the model judge, the premise fails.

P

将会公开What will be published

WILL PUBLISH

错误分类表、标注样本、校验脚本和检出率对照。The error taxonomy, labeled samples, validation scripts, and detection-rate comparisons.

前期线索Early leads相关的已有工作，不是这个实验的结论Related earlier work, not a result of this experiment

  • 化学数据质检与资料寻源Chemical data QA and source tracing做过化学合成数据质检，以及化学与学科资料的证据寻源。Ran QA on chemical synthesis data and sourced evidence for chemistry and discipline material.

实验日志Run logRUN LOG

还没有记录。实验开始后，每一次运行、失败和修正都会记在实验室里。Nothing logged yet. Once the experiment starts, every run, failure and fix will be recorded in the lab.

去实验室Go to the lab ↗