Joining 55 databases shows what materials science missed
One merged crystal library suggests many good materials sat unmeasured in plain view.
Scientists often believe they are inventing new materials. A new preprint suggests many of the answers were already written down, just in different places. In the paper on arXiv, Zhilong Song, Zijun Pang and Lixue Cheng matched records across 55 databases into one library of 14.6 million crystal structures and more than one billion property values.
Many entries had gaps, so the authors trained a model to fill in missing properties. The abstract says this points to tens of thousands of strong candidates among known materials that never had their properties recorded. It also reports that between 55% and 100% of structures offered as new by inverse design were already known.
The authors are careful about the limits. Their figures are model predictions, and the page links no dataset or code, so others cannot yet check the library for themselves.
For people who own or use data, the point is simple. Value was hiding in the joins. Records held apart looked ordinary, and held together they showed what was already known and what was still missing. A well-matched dataset can be worth more than the sum of its sources.

