Probing the limits of materials design via a comprehensive crystal database of 14.6 million structures with 140 properties

summary

The gist

Crystal materials are designed either by screening known crystals for a target property or by generating new crystals with that property The limits of both routes remain unmeasured Here we probe them…

In short

Researchers integrated data from 55 crystal databases into one comprehensive database of 14.6 million structures and 131.8 million property values, filling missing data with an AI model called CGTNet. This integration showed that forward screening can find more known materials than previously thought, while inverse design is limited by its narrow reference set.

Key concepts

Forward Screening
This method involves searching existing crystal databases for materials possessing a specific target property, like being ultra-incompressible. The study found this route reaches tens of thousands of candidates that were previously unrecorded in smaller databases.
Inverse Design
This approach uses an AI generator to create new crystals predicted to have a desired property. The research suggests its novelty is overestimated because it is judged against a limited set of known materials, and generated hits often contain common elements like osmium.
CGTNet
CGTNet is an equivariant graph transformer trained to predict crystal properties by learning from existing data. It creates a library of property models, allowing researchers to predict properties for millions of new crystals that have never been seen before.
Database Integration
The core effort involved matching crystals across 55 different databases using structural fingerprints and cross-database identifiers. This process created a unified, self-extending database that resolves identity issues across sources.

Terminology used across episodes

This episode discusses

The paper

Probing the limits of materials design via a comprehensive crystal database of 14.6 million structures with 140 properties · Read on arXiv

Zhilong Song, Zijun Pang, Lixue Cheng

Hong Kong University of Science and Technology

Crystal materials are designed either by screening known crystals for a target property or by generating new crystals with that property. Both routes are judged against public databases of known crystals, but these databases name the same crystal differently. Several materials announced as new by recent generative and autonomous discovery campaigns have since proved to be known phases or close variants of them. The limits of both routes remain unmeasured. Here we probe them by matching the crystals of 55 databases by structure into one self-extending database of 14.6 million crystals and 131.8 million property values, each traced to its source. Only 4.5% of its crystal-property pairs carry a value. To fill the empty pairs, we developed CGTNet, an equivariant graph transformer with state-of-the-art accuracy on several tasks of two public benchmarks. Trained on each task, it forms a library of property models whose predictions we release. The reach of forward screening has been underestimated. Known crystals hold tens of thousands of ultra-incompressible candidates without a recorded bulk modulus, and a post hoc audit of synthesis evidence finds among them ReN2, whose bulk modulus was measured at 428 GPa. The novelty claimed for inverse design has been overestimated. At an equal verification budget, a stiffness-conditioned generator returned more stiff hits than screening (100 versus 55), but all contained osmium and 82 repeated recorded compositions. Integrating and completing the databases of known crystals therefore moves the limits of both routes: it widens what screening can reach and sets the reference against which generated crystals are judged new.

Transcript

Introduction to the show: ident: Condensed Matter. The day in condensed matter: materials, phases and the measurements under them.

Kai: We're looking at the paper: "Probing the limits of materials design via a comprehensive crystal database of 14.6 million structures with 140 properties".

Paper summary: Kai: They build this massive database to resolve that identity across all sources by structure >

Mira: And they fill in those missing property values using something called CGTNet, which is a rotation equivariant graph transformer >

Lev: So what the paper claims is that by doing this matching and filling in the gaps, forward screening actually reaches further than it did when you only looked at the original databases >

Kai: They found fifty-three point eight hundred thirteen candidates without a recorded bulk modulus for ultra-incompressible solids, and after checking synthesis evidence, they kept ReN2 as an example >

Mira: But they also look at inverse design, and the novelty claimed for it seems overestimated because when you compare an equal verification budget, the stiffness-conditioned generator returned one hundred hits compared to only fifty-five from screening > <ref:2610.32678#pg1>

Conclusion: Kai: So, this paper is about building one definitive map for all known crystal materials and their properties so we can actually measure where the limits of searching are >

Mira: It means that when you look at forward screening, you can now see candidates that weren't previously visible because they were just named differently in another database >

Lev: And for inverse design, it gives us a better yardstick. Now we know what the generated crystals are competing against when we decide if something new is actually novel >

More episodes

← Home