The 25 Materials That Should Actually Exist
Every year, computational chemists publish lists of new materials that, according to their calculations, should be stable. The structures are plausible. The energies are negative. The phonon spectra show no imaginary frequencies. The papers get published. Then nothing happens.
There is a reason for that silence. The gap between a theoretically stable crystal and something you can actually synthesize in a lab — something that survives on a benchtop, in air, at room temperature — is wide enough to swallow entire classes of predicted materials. A material predicted by density functional theory may be stable at zero Kelvin in vacuum, but fall apart the moment a postdoc tries to grow it on a substrate.
Now a collaboration between Toyota Research Institute, Toyota Central R&D Labs, and the University of Tokyo has built a framework that might finally close that gap. Their approach, described in a preprint posted to arXiv on July 2, takes a list of 894 computationally stable materials and winnows it down to 25 that are actually worth trying to make in a lab.
The gap between prediction and synthesis is not a minor inconvenience. It is the central bottleneck in computational materials science. Researchers can now screen thousands of hypothetical compounds in silico, ranking them by thermodynamic stability, elastic constants, band gaps, and dozens of other properties. But when experimentalists try to make the top-ranked candidates, they often fail. The compound turns out to be metastable in a way that the initial calculation missed. Or it decomposes into competing phases that the original screening simply ignored.
“Many remain experimentally unrealized,” the authors write, “underscoring a persistent gap between prediction and synthesis.” The understatement is characteristic of a field that has produced far more candidate materials than successful syntheses.
What the Researchers Did
The team, led by Yuqi An, Sihong Zhu, and Joseph Montoya, designed a hierarchical screening protocol that applies increasingly stringent filters in sequence. Each filter is designed to catch a different reason why predicted materials fail in practice.
They started with the 894 stable materials previously reported in a well-known 2022 dataset published in Scientific Data. These were materials that had passed an initial PBE-level DFT stability check: their formation energies were negative relative to the elemental phases. That sounds rigorous, but PBE is a functional known to overestimate stability. It can predict that a material is stable when, under a more accurate functional, it is not.
The first pass was structural curation. Many of the 894 entries were duplicates, incomplete structures, or materials whose reported compositions didn’t match their crystal structures. The team cleaned the set down to 603 unique structures with verified stoichiometries.
Then came the real tests.
Stage One: Thermodynamic Stability on Complete Phase Diagrams
The first major filter recalculated stability against all competing phases on the full PBE phase diagram, not just the elemental references. This is a crucial distinction. A material can have a negative formation energy relative to its elements but still be unstable because a different combination of compounds (the competing phases) is energetically more favorable. The team found that only 298 of the 603 curated structures survived this check. More than half of the supposedly stable materials decomposed on paper when all their neighbors were considered.
Stage Two: Dynamical Stability
Surviving thermodynamic stability does not mean a material will hold together. It could be dynamically unstable, meaning that at finite temperature, the atoms would vibrate their way into a different structure. The team screened for dynamical stability using two complementary methods: harmonic phonon calculations (which check for imaginary frequencies in the phonon spectrum) and finite-temperature molecular dynamics simulations (which check whether the structure holds up at 300 K and above).
They used universal machine-learned interatomic potentials for the phonon and MD calculations, which was itself a notable methodological choice. Instead of running expensive DFT phonon calculations for every candidate (which would have required thousands of hours of compute), they used a pre-trained universal potential. This is the same class of models that powers much of the recent progress in computational materials science.
After dynamical screening, 166 materials remained.
Stage Three: SCAN-Level Thermodynamic Refinement
PBE is fast but not especially accurate. The meta-GGA functional SCAN is significantly more accurate for lattice constants and formation energies, but too expensive to run on thousands of candidates. By stage three, however, the team had only 166 candidates left, making SCAN calculations feasible.
SCAN narrowed the set further to 109 materials that remained thermodynamically stable under the more accurate functional.
Stage Four: Prioritization for Synthesis
Having 109 candidates was still too many to hand to an experimentalist. The team applied two final criteria. First, they ranked materials by decomposition enthalpy: how much energy would be released if the material decomposed into its competing phases. Materials with large decomposition enthalpies are more likely to actually form, because the thermodynamic driving force against decomposition is stronger. Second, they measured chemical-space completeness: how well each material filled a gap in the known chemical landscape. Materials that represented entirely new combinations of elements, or new structure types for known compositions, were prioritized.
The result: 25 high-confidence targets for experimental synthesis.
| Stage | Screening Criterion | Number Remaining | Filtering Method |
|---|---|---|---|
| Start | Materials from Sci. Data dataset | 894 | Literature collection |
| Curation | Unique, verifyable structures | 603 | Structural deduplication and stoichiometry check |
| Stage 1 | Full PBE phase-diagram stability | 298 | Grand-canonical phase diagram analysis against all competing phases |
| Stage 2 | Dynamical stability (phonon + MD) | 166 | Universal MLIP phonon spectra + 300K MD trajectories |
| Stage 3 | SCAN thermodynamic refinement | 109 | SCAN-functional formation energy on phase diagram |
| Stage 4 | Decomposition enthalpy + chemical novelty | 25 | Prioritization by decomposition driving force and chemical-space gap-filling |
Why It Matters
The numbers tell a story that every computational materials scientist knows intuitively but rarely quantifies so starkly. Of 894 materials that looked stable in the simplest calculation, only 109 survived a moderately rigorous set of checks. That is a false-positive rate of nearly 88 percent. If you had picked a random material from the original list and tried to make it, you would have had roughly one chance in eight of success.
More importantly, the 25 finalist materials are not just theoretically stable: they are, in the authors’ phrase, “experimentally actionable.” Each comes with a specific predicted synthesis pathway, a known set of competing phases to avoid, and a clear signature that experimentalists can check (XRD pattern, lattice parameters, elastic moduli) to confirm that they made the right thing.
The framework also reveals something useful about why predicted materials fail. The single biggest filter was the full phase diagram analysis. More than half the candidates were eliminated not because they were unstable in an absolute sense, but because competing phases offered a more stable arrangement of the same atoms. This suggests that the bottleneck in materials discovery is not predicting stable compounds, but predicting the most stable arrangement among many similar possibilities. It is a ranking problem, not a classification one.
How It Could Change Our Lives
Better synthesis prediction has direct, measurable consequences. Battery researchers need electrolytes that don’t decompose at high voltage. Thermoelectric developers need materials with the right trade-off between electrical and thermal conductivity. Catalysis groups need surfaces that bind reactants strongly but products weakly. All of these depend on experimental synthesis of predicted materials.
Right now, the typical workflow in computational materials discovery goes like this: screen thousands of candidates computationally, pick the top 20, send them to a collaborator who makes three or four over the next year, and hope one works. The 88 percent failure rate is built into the system. Screening protocols that cut that failure rate in half would roughly double the effective throughput of the entire discovery pipeline.
There is also a more subtle implication. The framework is hierarchical by design: easy filters first, expensive filters last. The cheap PBE screen eliminated the bulk of candidates. The more expensive SCAN screen was only needed for the survivors. This kind of tiered approach makes large-scale screening economically feasible. As universal MLIPs improve, the dynamical screening step (currently done with phonon calculations and short MD runs) could become even cheaper, allowing researchers to screen tens of thousands of candidates instead of hundreds.
The Bigger Picture
This paper is part of a larger reckoning in computational materials science. For a decade, the field has been generating predicted materials faster than experimentalists can verify them. The MatBench leaderboard, the Materials Project, AFLOW, the OQMD: each database contains hundreds of thousands of theoretical compounds. Many of them are almost certainly correct. Many are probably wrong. Nobody knows which is which.
The Toyota-TRI-Toyota CRDL approach offers a practical path forward. Instead of chasing perfect accuracy in a single calculation (which would require expensive correlated wavefunction methods that cannot scale to thousands of materials), the framework layers multiple approximate methods in sequence, each catching a different failure mode. It is a strategy that reflects how materials science actually works: an iterative process of calculation, measurement, and refinement, not a single shot from a black-box predictor.
Limitations & What’s Next
The framework is not perfect. The universal MLIP used for dynamical screening is itself an approximation, and the authors acknowledge that some materials eliminated at stage two may be stable under more accurate (but more expensive) phonon calculations. The SCAN functional, while better than PBE, is still not exact. And the final prioritization into 25 candidates relies on heuristics (decomposition enthalpy, novelty) that may not capture all the practical factors that determine whether a material can be synthesized, such as precursor availability, reaction kinetics, or surface energy effects.
The real test will come when experimental groups attempt to synthesize the 25 candidates. If even a fraction of them succeed, the framework will have proved its value. If they fail, the screening protocol will need another layer, perhaps incorporating kinetic barriers or nucleation considerations.
The team has made their protocol and results available. The 25 recommended materials are listed by composition and space group, with predicted XRD patterns and competing-phase lists. For computational materials scientists, it is a target list. For experimentalists, it is a challenge.
📄 Source: An, Y., Zhu, S., Montoya, J., et al. “Predicting Novel Stable Materials for Experimental Synthesis.” arXiv:2607.01713 (2026). https://arxiv.org/abs/2607.01713