Materials scientists hunting new materials usually look forward: compute new structures, try new formulas. A team at Linköping University in Sweden looked backward instead—using machine learning to audit existing databases, they found 33 MXenes that had already been synthesized but misidentified by peers, and followed the trail to create 11 entirely new rare-earth MXenes. The paper was published in *Science* on October 29.
What MXenes Are, and Why They Matter
MXenes are a family of materials named only in 2011: two-dimensional layered structures of metals with carbon or nitrogen, a few atoms thick, with enormous surface area—millions of layers can stack in a film a few millimeters thick, with composition-dependent interlayer chemistry. Applications range from water purification and energy storage to telecommunications and biomedicine. Team leader Johanna Rosén put it plainly: "The literature and databases are treasure troves—you just need to lift the lid and see what's inside." The project has an informal name: the Treasure Chest Project.
The screening pipeline deserves precise accounting. The actual chain per primary sources: merging the Materials Project and OQMD databases yielded 541 candidate structures. A random forest classifier (trained on 1,050 balanced labeled samples) plus literature verification identified 56 experimentally reported multilayer MXenes, of which 33 had never previously been identified as MXenes—many were synthesized before the word "MXene" even existed, including German work from the 1970s. Confirming these 33 "known" materials required zero new experiments.
The Synthesis Route Got a Rework Too
The new synthesis route uses self-propagating high-temperature synthesis (SHS): a brief ignition triggers the reaction, which sustains itself with its own heat—no continuous external heating needed, completing within minutes (the university press release says seconds; the paper says minutes—the discrepancy is noted as-is). Hazardous chemical and water usage dropped significantly compared to older methods. The team made 18 phases in total, including 11 brand-new rare-earth MXenes (Pr, Nd, Sm, Gd, Tb, Ho, Tm, plus known yttrium compounds as controls). Introducing rare earths gave MXenes magnetism for the first time: holmium-based samples showed measured antiferromagnetic ordering, while calculations also yield ferromagnetic and antiferromagnetic ground states. Most predicted phases are semiconductors with 1–2 eV band gaps; some cerium and ytterbium phases are semimetals.
Rosén puts the key to industrialization on scale-up: "I believe scale-up is what will push MXenes over the tipping point." Room-temperature synthesis hits the scaling pain point directly: low energy use, fewer hazardous chemicals, lower retrofit costs for production lines. *Science*'s editors' summary likewise flagged the old route's pain: wet chemistry involves many dangerous reagents and time-consuming steps.
The Methodology May Be Worth Copying More Than the Materials
This structure is increasingly common in AI for Science: AI as a classifier, not a prophet—it shrinks the search space, while literature, theory, and experiment make the final call. The cost comparison is brutally concrete: confirming 33 "known" materials cost roughly nothing, whereas blind trial-and-error could mean waiting another thirty years for a missed candidate. The same division of labor appears in AI-designed proteins and high-energy physics. Novoselov's late-October reminder at a leading scientists' forum applies equally: AI predictions use idealized materials like pure iron and pure silicon, while real materials are more complex—blueprints from the "AI designer" must still pass the reality gate of materials synthesis.
> Tip: MXene is pronounced "maxene," with formula M(n+1)X(n)T(x): M is a transition metal, X is carbon or nitrogen, T is a surface termination group. The space between layers, open to chemical modification, is the family's ticket to multiple industries.
References
- *Science*: Data-driven discovery and rapid, direct synthesis of MXenes (Saffar Shamshirgar & Rosén, Oct 29, DOI 10.1126/science.aeh5957)
- Linköping University press release: Hidden and forgotten 2D materials are discovered with AI (Oct 29)
- TSNMedia: notes on the screening pipeline and the two inconsistent duration claims
- Interesting Engineering: background interview on the 1970s German precursor work