MatterSpace Blindly Rediscovered Re₁@Ni and Ir₁@Ni Single-Atom Alloy Catalysts

This is not an approximate match or a vague similarity. MatterSpace reproduced the full three-dimensional atomic arrangements of these known single-atom alloy catalysts to within half an angstrom—a precision that places the generated structures within the range of experimental measurement uncertainty for crystallographic determination. The engine found what nature already confirmed, without being told what nature had found.

What was rediscovered

The sealed targets for this blind rediscovery benchmark were Re₁@Ni and Ir₁@Ni single-atom alloy (SAA) catalysts—materials where a single atom of rhenium or iridium is embedded in a nickel host lattice. These SAA catalysts have been experimentally validated for methane cracking, one of the most important reactions in industrial chemistry and hydrogen production. They represent a specific, well-characterized class of materials whose catalytic properties arise from the precise geometric and electronic environment created by the single dopant atom in the nickel surface.

MatterSpace was given no information about these targets. The generation pipeline was completely firewalled from the target structures. The engine explored the landscape of possible dopant-host combinations purely from physics-first principles, evaluating candidates based on adsorption energetics and structural stability without any signal indicating which dopant elements or geometric configurations were correct. Both rhenium and iridium were independently identified as promising dopants for nickel-hosted single-atom alloy catalysts for methane activation—across a search space spanning 23 different dopant elements.

The three-level validation protocol

MatterSpace evaluates blind rediscovery through a three-level validation protocol, each level answering a progressively more demanding question about the quality and fidelity of the generated candidates.

Level A — Adsorption energy threshold

Level A asks whether the generated candidates meet the performance threshold relevant to the target application. For methane cracking catalysis, the critical metric is methane adsorption energy: candidates must exhibit adsorption energies below −1.3 eV, indicating sufficient catalytic activity for C–H bond activation. Of the 600 candidates generated across 23 dopant elements, 581 passed Level A—a 96.8% pass rate. This confirms that MatterSpace was generating candidates in the correct performance envelope for the target application, not merely producing physically valid but catalytically irrelevant structures.

Level B — Structural fingerprint matching

Level B asks whether the structural fingerprints of generated candidates match known target materials. Structural fingerprints encode the local chemical environment around each atomic site—coordination geometry, bond angles, nearest-neighbor distances—into compact numerical representations that enable rapid comparison against known structures. Of the 600 candidates, 75 achieved fingerprint matches, with the best fingerprint similarity reaching 0.814. Critically, both rhenium and iridium were independently identified as matching dopant elements. The engine did not merely find one of the two targets by chance—it found both, through independent physics-driven exploration of the dopant landscape.

Level C — Full structural RMSD

Level C is the strictest test: full atomic-level structural comparison using root-mean-square deviation (RMSD) against the known crystal structures of the target catalysts. This measures whether the engine reproduced the complete three-dimensional arrangement of atoms—not just the right performance properties or local environment—but the actual structure at crystallographic resolution. MatterSpace achieved a full RMSD of 0.408 Å for Ir₁@Ni and 0.466 Å for Re₁@Ni. Both values fall below the 0.5 Å threshold that defines a successful structural rediscovery. At this precision, the generated structures are crystallographically equivalent to the experimentally known materials.

600 candidates. 23 dopant elements. Zero target knowledge. Both Re₁@Ni and Ir₁@Ni independently rediscovered with sub-half-angstrom structural precision. That is blind generative rediscovery.

Computational cost and efficiency

The entire blind rediscovery campaign—all 600 candidates across 23 dopant elements—ran on a single NVIDIA A100 GPU in approximately 4.7 hours at an estimated cloud compute cost of roughly $15. This represents a 130–270× cost reduction compared to density functional theory (DFT) screening, which would require hundreds to thousands of individual DFT calculations to evaluate the same candidate space, each taking minutes to hours on high-performance computing clusters. MatterSpace compressed what would traditionally be weeks of supercomputer time and thousands of dollars in compute cost into a single afternoon on a single GPU.

The structural validity of the generated candidates was 97.5–99%, meaning that virtually all output structures were physically valid without post-hoc filtering or repair. This is what valid by construction means in practice: the physics enforcement during generation ensures that candidates satisfy crystallographic constraints, charge neutrality, and coordination requirements as they are being created, not as a cleanup step applied afterward. The near-total validity rate means the engine wastes almost no compute on physically impossible structures—a sharp contrast to autoregressive generative models where the majority of raw outputs typically require filtering or rejection.

First demonstration of its kind

This is the first demonstration of blind generative material rediscovery that achieves all three validation levels—performance threshold, fingerprint matching, and sub-angstrom structural reproduction—in a single campaign. The distinction matters because existing generative approaches in materials science have achieved partial validation at best. GNoME has demonstrated large-scale stability prediction but does not perform generative blind rediscovery. MatterGen generates novel crystal structures but has not demonstrated three-level blind rediscovery of specific known catalysts. Open Catalyst provides energy predictions for catalytic surfaces but is a prediction tool, not a generative discovery engine. CDVAE and DiffCSP generate crystal structures through diffusion-based approaches but have not been evaluated against a sealed three-level blind rediscovery protocol.

None of these systems—each significant in its own right—has demonstrated the ability to blindly generate candidates with zero target knowledge and then pass performance, fingerprint, and structural validation against sealed reference materials. MatterSpace is the first to achieve all three levels, and it did so for two independent targets simultaneously, confirming that the result is not a statistical anomaly for a single fortuitous case.

What blind rediscovery proves

Blind rediscovery is not an end goal. Nobody needs an engine to rediscover materials that already exist. The value is epistemic: it establishes that the physics encoded in the engine is doing real scientific work. If MatterSpace can independently arrive at the same structures that human researchers validated through years of experimental synthesis—without any hint that those structures exist—then the novel candidates the engine produces in open discovery campaigns deserve serious scientific attention. They emerged from the same physics-first process that independently confirmed known science.

This is a fundamentally different evidentiary standard than what generative models typically offer. Most generative systems are evaluated by how plausible their outputs appear, measured by proxy metrics like validity rates, novelty scores, or energy predictions. Plausibility is not evidence. Blind rediscovery is evidence. It demonstrates that the engine's internal representation of chemical physics is faithful enough to reproduce nature's actual solutions to a specific catalytic problem, starting from nothing but the physics and the constraints.

Implications for catalysis and beyond

Single-atom alloy catalysts represent one of the most active frontiers in catalysis research. By isolating individual dopant atoms in a host metal lattice, SAA catalysts can achieve unique selectivity and activity profiles that neither the pure host metal nor the pure dopant can provide. The blind rediscovery of Re₁@Ni and Ir₁@Ni for methane cracking demonstrates that MatterSpace can navigate this design space effectively—identifying not just plausible compositions but the specific atomic configurations that produce catalytic activity.

Methane cracking is itself a reaction of enormous practical significance. It is the primary industrial route to hydrogen production and a critical step in natural gas processing. Catalysts that can activate the C–H bond in methane efficiently and selectively are among the most sought-after targets in catalysis research. The fact that MatterSpace independently identified two experimentally validated catalysts for this reaction—from a pool of 23 possible dopant elements—without any guidance toward methane cracking or C–H activation specifically, underscores the depth of the physics driving the generation process.

The blind rediscovery protocol generalizes beyond catalysis. MatterSpace applies the same three-level validation methodology across every domain pack—batteries, superconductors, magnets, thermoelectrics, and others. In each domain, the performance threshold, fingerprint criteria, and structural RMSD standards are adapted to the relevant physics, but the protocol structure remains the same. Blind rediscovery is not a one-off demonstration for a single application. It is a systematic validation methodology that MatterSpace applies to every domain it enters.

The research paper

The full research paper documenting this blind rediscovery—including complete validation results, comparative analysis against GNoME, MatterGen, Open Catalyst, CDVAE, and DiffCSP, and full methodological details—was authored by Faruk Guney at Vareon, Inc. in February 2026. The paper is forthcoming and will be available at vareon.com/research.

MatterSpace does not need to be told what exists. It discovers what the physics demands. When that happens to match what nature already produced—to within half an angstrom, for two independent targets, across 23 dopant elements, in under five hours on a single GPU—it is not a coincidence. It is confirmation that the engine is doing science, not statistics.