Finding a material with a target set of properties means searching a combinatorial space of compositions, structures and processing conditions. Exhaustive experiment is impossible; the space is far larger than any lab's throughput.
The useful role for models
Not to replace experiment, but to rank what to try next. A model trained on known structure-property relationships proposes candidates most likely to hit the target. The lab tests those. Results feed back. The loop narrows the search from millions of possibilities to hundreds worth attempting.
Why this works better than it sounds
- The physics provides strong structure, so models are not learning from nothing.
- Negative results are informative and cheap to record, which is unusual in machine learning.
- Being approximately right is enough, because the experiment is the arbiter.
The honest limits
Models trained on published data inherit publication bias, and failed syntheses are rarely published. Simulated properties diverge from measured ones, sometimes badly, particularly at interfaces and defects, which is where nanoscale behaviour is most interesting. Treat model output as a ranked hypothesis list, never as a result.