Probing the limits of forward screening and inverse design of crystal materials
By unifying 55 disparate crystal databases into a comprehensive resource and employing the CGTNet model to predict missing properties, this study reveals that forward screening of known materials can uncover thousands of ultra-incompressible candidates like ReN2, while demonstrating that current inverse design methods often rediscover existing compositions, thereby establishing a more rigorous foundation for evaluating material novelty.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Materials scientists are constantly on the hunt for new crystals, the rigid, repeating structures that make up everything from the silicon in our phones to the diamonds in jewelry. To find a material with a specific superpower—like the ability to withstand immense pressure or conduct electricity perfectly—researchers usually take one of two paths. The first is a treasure hunt: they sift through millions of known crystals, looking for one that already possesses the desired trait. The second is a creative act: they use powerful computer programs to imagine and generate entirely new crystal structures that should have that trait. Both methods rely on massive public libraries of data, where scientists store the blueprints and properties of crystals they have discovered or simulated. But for years, these libraries have spoken different languages. The same crystal might be listed in one database with one name and in another with a completely different name, or it might be missing key information entirely. This confusion has made it difficult to know how far these search methods can really go, or how truly "new" the materials they find actually are.
A team of researchers at the Hong Kong University of Science and Technology has now cleared away this fog by building a single, unified library that brings together 55 different databases. They did not just copy and paste the data; they carefully matched every crystal in these collections by its physical shape, ignoring the confusing names and chemical formulas that often fail to distinguish between different versions of the same material. This massive effort resulted in a single, self-updating catalog containing nearly 14.6 million unique crystal structures and over 131 million property measurements. Crucially, they kept a record of where every single piece of information came from, allowing them to trace the history of every value. They found that while this library is enormous, it is also incredibly sparse. For the vast majority of crystals, the specific properties researchers care about are simply missing from the records. To fill these gaps, the team developed a new artificial intelligence model called CGTNet. This model acts like a highly skilled guesser, trained to predict missing properties based on the crystal's structure with an accuracy that surpasses previous methods. By using this tool, they were able to complete the library, predicting values for millions of crystals that had never been measured or calculated before.
With this completed library in hand, the researchers tested the limits of both the treasure hunt and the creative generation approaches. They first looked at the treasure hunt, or "forward screening," asking how many ultra-hard materials might be hiding in plain sight among the known crystals. Before this study, scientists believed they had already found most of the candidates. However, by applying their new AI model to the completed library, they discovered tens of thousands of known crystals that were likely ultra-incompressible but had never been tested for this specific property. Among these hidden gems, they found a material called ReN2. Although this crystal had been synthesized in a lab under high pressure and its hardness had been measured at 428 gigapascals, no database had ever recorded its bulk modulus, a measure of its resistance to compression. The researchers' screen successfully identified it as a top candidate, proving that the treasure hunt can still find valuable, known materials that were simply overlooked because the data was scattered and incomplete.
The team then turned their attention to the creative approach, or "inverse design," where computers generate new crystal ideas. For years, these computer programs have claimed to be discovering vast numbers of brand-new materials. The researchers tested these claims by checking the generated crystals against their new, comprehensive library. They found that the sense of novelty was largely an illusion caused by using incomplete reference lists. When they checked the generated crystals against the full library of 55 sources, they discovered that a huge number of them were not new at all. In one specific test, a computer program generated 100 crystals that it claimed were new and extremely stiff. When the researchers checked these against their complete database, they found that 82 of them were actually variations of compositions that were already known and recorded. Furthermore, every single one of these "new" crystals contained a specific element, osmium, suggesting the generator was just remixing a very narrow set of known ingredients rather than exploring the full universe of possibilities.
The study concludes that the way we search for materials needs to change. The treasure hunt is far more powerful than previously thought; by integrating all known data and filling in the missing pieces, we can find thousands of useful materials that already exist but were hidden in the gaps of our records. Conversely, the creative generation approach has been overestimating its own success. The "new" materials it produces are often just rehashes of what we already know, simply because the reference lists used to judge them were too small. By using a single, complete, and verified database, scientists can now tell the difference between a genuine discovery and a duplicate, ensuring that the search for the next great material is built on a foundation of truth rather than confusion.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.