Most headlines about machine learning celebrate scale: more parameters, more data, more compute. In materials science, that framing quietly misleads us. Our datasets are not big — they are expensive. A single high-fidelity calculation or a careful synthesis can cost days of work, so the binding constraint is rarely the model. It is how much we can learn per data point.
The problem with borrowing the large-data playbook
Methods tuned for millions of examples often assume that more data will paper over weak inductive biases. When you have a few hundred structures, that assumption collapses. The question stops being “which architecture wins a benchmark?” and becomes “which approach extracts the most signal from a small, noisy, multi-fidelity dataset?”
Three levers that actually move the needle
- Physically grounded representations. Encoding the chemistry and physics we already understand reduces the amount the model has to learn from scratch.
- Transfer and active learning. Reusing knowledge across related tasks, and choosing the next experiment intelligently, beats collecting data blindly.
- Honest uncertainty. A model that knows what it does not know is worth more to an experimentalist than one extra point of accuracy.
None of these are glamorous. All of them compound.
Why it matters
If we optimize for data efficiency, AI stops being a spectator that summarizes past results and becomes a collaborator that guides the next experiment. That, to me, is the version of the field worth building.
Have a different take? I’d genuinely like to hear it — reach me by email.