Scientific data quality: why it decides your model
Scientific data is less trustworthy than the models built on it assume. Why quality, leakage, and provenance decide your result, and how to check them.
Here is the uncomfortable premise this whole publication is built on: the scientific data underneath a model is usually far less trustworthy than the model assumes. We spend enormous effort tuning architectures and almost none checking whether the numbers we trained on are correct, comparable, and honestly split. In scientific machine learning, the data is the model’s ceiling, and most of the time nobody measures where that ceiling is.
This is the pillar on scientific data quality. It ties together the articles on auditing, leakage, and representation, and it is the reason “scientific AI” is not just ordinary AI with a lab coat on.
Why scientific data is shakier than it looks
Scientific data feels authoritative. It comes from instruments, simulations, and peer-reviewed databases, so it is easy to treat every value as ground truth. In practice it carries the same flaws as any other data, plus a few of its own:
- Method dependence. A computed property depends on the method and settings used. The same material can have different “correct” values from different calculations, and they are not always comparable.
- Silent errors. Typos, wrong units, and mislabelled structures sit in real databases. They are rare per record and devastating in aggregate.
- Selection bias. What gets measured and published is not a random sample of materials, so a model trained on it inherits that bias.
None of this means the data is useless. It means the data has to be checked, not assumed.
Data leakage: the error that flatters you
The most dangerous quality problem is not a wrong value, it is a leaky split. If near-duplicate samples end up in both training and test sets, the reported score is inflated and meaningless, and you will not notice, because the number looks excellent.
The subtle part, and the reason this is a quality issue and not just a coding one, is that whether a split leaks depends on what the model can see. The same split can be fair for one model and cheating for another. That idea is worked through in graph neural networks for materials and underlies the whole band-gap prediction series.
Representation is a data-quality decision
How you describe a material to a model is usually treated as a modelling choice. It is really a data choice, and it caps your result before training starts. Describe a material by composition alone and you have thrown away structure, which is often the very thing that determines the property. That argument, in full: properties belong to structures, not formulas.
What good practice looks like
The habits that keep scientific data honest are unglamorous and non-negotiable:
- Audit before you trust. Learn to look at a single suspect record and decide whether it holds up: how to audit a database entry.
- Test the data, not just the code. Every physically meaningful field has constraints. Assert them and fail loudly, the way materials data engineering describes.
- Report the honest number. When fixing a leak lowers your score, the lower score is the true one. Publish it. That is the house rule of this site.
Why “scientific AI” is different
Ordinary AI can often absorb messy data because it has oceans of it. Scientific AI usually cannot: datasets are small, expensive, and physically constrained, so every bad record weighs more and every leak distorts more. This is exactly why methods that respect physics tend to win, and why data quality is not a side task but the main one. The overview of those methods is here: six ways to put physics into machine learning.
Start here
- How to audit a database entry, trust one record.
- Properties belong to structures, representation as data quality.
- Materials data engineering, test the data.
- Why predicted materials rarely get made, where confident numbers go to die.
A model can only be as good as what it learned from. In science, that is a higher bar than it sounds, and clearing it is most of the real work.
Found a mistake? Good, tell me. This publication flags its own suspect values. Reach me on LinkedIn.