A new materials AI model generated candidate crystal structures roughly ten times faster in a benchmark comparison with previously reported results for other models, though its structures could miss fine detail. Called OMatG-flash, it is designed to supply more candidates for computational screening, not to produce a finished material.
Correction — October 6, 2026: LeMat-GenBench maintainers say earlier MACE-based hull calculations sometimes used a changing default model instead of the fixed MACE version used for reference energies. That mixed energy scales, so historical stability, SUN and MSUN results are provisional while the team recomputes them. The 21.5% versus 24.1% combined-score comparison below includes these measures and is provisional too; corrected figures were not posted when this update was prepared.
Crystal-design systems must specify which atoms belong in a repeating unit and where they sit. Many current generators move from a random starting point toward a plausible structure through numerous model evaluations. OMatG-flash learns a flow map that makes that move in fewer calls. Its authors tested both prediction from a given set of elements and generation of new compositions and structures.
Faster generation, mixed benchmark results
For the generation test, the researchers assessed 2,500 proposed structures with LeMat-GenBench, which estimates whether candidates are stable or metastable, unique and novel relative to known materials. With eight model calls per candidate, OMatG-flash scored 21.5% on the benchmark's combined measure, compared with 24.1% for Crystalite. In a speed test on one Nvidia H100 GPU, OMatG-flash took 0.535 seconds to generate 1,000 structures using 16 calls, compared with Crystalite's previously reported optimized time of 5.14 seconds. Results for rival systems were taken from previously reported leaderboard entries rather than rerun by this team.
In a separate structure-prediction test, the model showed a further accuracy trade-off. This test accounts for multiple possible structures of the same composition. Its base version matched 68.56% of reference structures by the researchers' coverage measure. Reward-based fine-tuning raised that figure to 71.10%, just above Crystalite's 70.87%, but its corrected structural error remained higher: 0.221 against 0.174, where lower is better. The team applied that fine-tuning only to structure prediction, not to generation of new compositions.
The tests used datasets of inorganic crystals with no more than 20 atoms in their repeating unit. The authors say their base model often finds the broad structural pattern without resolving fine detail. Its candidates still need downstream relaxation and validation, and the paper reports benchmark calculations rather than laboratory synthesis. Code for OMatG-flash is planned for the OMatG package, but the paper does not claim that this version is already released.
Sources
- OMatG-flash research paper (22 September 2026; full HTML text).
- LeMat-GenBench evaluation framework.
- Lead image: Figure 1 from the OMatG-flash paper, by Thomas Egg, Harry Winston Sullivan, Ellad B. Tadmor and Stefano Martiniani, licensed CC BY 4.0. The chart plots benchmark quality against crystal candidates generated per second.