FluxMateria predicts surface-reaction activation barriers from composition and bond topology — matching single-method DFT accuracy at analytical speed, with no DFT inputs.
MAE by reaction family across 29 published literature cases.
| Family | n | MAE (eV) | RMSE (eV) | ≤0.3 eV | ≤0.5 eV |
|---|---|---|---|---|---|
| Combined benchmark | 29 | 0.236 | 0.309 | 66% | 93% |
| N₂ dissociation (ammonia volcano) | 7 | 0.223 | 0.252 | 57% | 100% |
| H₂ dissociation | 6 | 0.188 | 0.218 | 67% | 100% |
| O₂ dissociation | 4 | 0.180 | 0.232 | 75% | 100% |
| C-H activation (CH₄) | 8 | 0.220 | 0.291 | 75% | 88% |
| CO dissociation | 4 | 0.417 | 0.535 | 50% | 75% |
The 4d series hits 0.147 eV MAE — comparable to the error of individual DFT functionals on the same reactions.
Fe, Co, Ni, Cu
Ru, Rh, Pd, Ag, Mo
Re, Pt, Ir, Au
Where FluxMateria lands against direct DFT, analytical BEP/Hammer–Nørskov relations, and ML foundation models trained on DFT.
| Method | Type | MAE (eV) | Training data |
|---|---|---|---|
| FluxMateria (this work) | First-principles analytical | 0.236 | none |
| PBE DFT (single method) | Direct DFT | 0.20 – 0.30 | none (full DFT per case) |
| BEEF-vdW DFT | Direct DFT | 0.15 – 0.25 | none (full DFT per case) |
| BEP scaling relations (Nørskov) | Analytical on DFT fit | 0.20 – 0.30 | DFT database |
| Hammer–Nørskov analytical | Analytical | 0.30 – 0.50 | DFT database |
| UBI-QEP (Shustorovich) | Analytical | 0.30 – 0.60 | experimental fit |
| SISSO surrogate | ML on DFT features | 0.20 – 0.30 | ~10³ DFT points |
| CGCNN / GemNet-OC / Equiformer | Graph neural networks | 0.15 – 0.25 | ~10⁵–10⁶ DFT points |
| MACE / M3GNet foundation MLFF | Universal ML force field | 0.10 – 0.20 | ~10⁶–10⁷ DFT points |
Large ML foundation models (MACE, GemNet-OC) reach lower MAE, but only after consuming millions of DFT calculations and often failing out-of-distribution on unseen metals or adsorbates. FluxMateria delivers the same accuracy class as single-method DFT without consuming a single DFT result.
| Use case | Status | Notes |
|---|---|---|
| Catalyst screening & ranking | ✓ Production-ready | MAE < 0.3 eV clears the bar for ordering candidate catalysts. |
| Inverse search & discovery | ✓ Production-ready | Feeds FluxMateria's inverse-search layer to find novel active compositions. |
| Qualitative activity classification | ✓ Production-ready | 93% within 0.5 eV — well above the typical production threshold. |
| Reaction-family ranking | ✓ Production-ready | N₂, H₂, O₂ all at 100% within 0.5 eV. |
| Trend prediction across metal series | ✓ Production-ready | Per-period MAE tracks DFT method-to-method spread. |
| Quantitative turnover-frequency prediction | ⚠ Edge of usefulness | Rate ∝ exp(−E_a / kT). Wants MAE < 0.1 eV for quantitative TOFs. |
| Selectivity between close competing mechanisms | ⚠ Edge of usefulness | Two paths differing by < 0.3 eV are hard for any analytical method. |
| Transition-state geometry | Not applicable | Analytical barrier only; TS geometry is a separate DFT task. |
Activation barriers are the single most important quantity in heterogeneous catalysis. They determine turnover frequencies, selectivity between competing paths, and whether a candidate catalyst is worth making at all.
0.236 eV MAE puts FluxMateria inside the method-to-method spread between PBE, RPBE, and BEEF-vdW DFT functionals on the same benchmark set.
Microseconds per prediction, no DFT calculation, no geometry optimisation. Built for high-throughput screening and inverse search.
Unlike graph-neural-network catalysis models, FluxMateria needs zero DFT training data. Works on novel metals and adsorbates the benchmark never saw.
Ready for catalyst screening, ranking, and inverse discovery. Feeds the FluxMateria catalyst-scoring and microkinetics layers.
Every predicted barrier was compared against one or more of these independent studies. Twelve independent publications contribute to the benchmark.
Logadottir A., Rod T.H., Nørskov J.K. et al., J. Catal. 197, 229 (2001)
Bligaard T., Nørskov J.K., Dahl S. et al., J. Catal. 224, 206 (2004)
Vojvodic A., Nørskov J.K., Science 334, 1355 (2011)
Bengaard H.S., Nørskov J.K. et al., J. Catal. 209, 365 (2002)
Chin Y.H., Buda C., Neurock M., Iglesia E., J. Am. Chem. Soc. 133, 15958 (2011)
Andersson M.P., Abild-Pedersen F., Nørskov J.K., J. Catal. 255, 6 (2008)
Eichler A., Hafner J., Surf. Sci. 433-435, 58 (1999)
Michaelides A., Scheffler M. et al., J. Am. Chem. Soc. 127, 6289 (2005)
Tkatchenko A. et al., Phys. Rev. Lett. 101, 073005 (2008)
Greeley J., Stephens I.E.L., Bondarenko A.S. et al., Nat. Chem. 1, 552 (2009)
Campbell C.T., Surf. Sci. 157, 43 (1985)
Nørskov J.K. et al., Fundamental Concepts in Heterogeneous Catalysis, Wiley (2014)
Honest bounds, published up front. The residuals trace to physics beyond the current descriptor set — the same boundary that challenges DFT and ML models.
Under-predicted by ~0.9 eV. Ni's magnetic coupling to CO fragments is beyond the d-band-center framework — DFT without explicit spin treatment shows similar residuals.
Under-predicted. The shallow-d-band modulation on the C-H activation floor is conservative; the true volcano shape is multi-descriptor.
Over-predicted. Hits the universal bond-breaking floor that doesn't yet distinguish spin-triplet O₂ physics.
Transition-metal (111)/(110)/(0001) surfaces. Step, kink, and alloy-site corrections are handled by separate FluxMateria layers.
The barrier predictor is the activation-kinetics engine for FluxMateria's catalyst module. It drives the catalyst scoring, inverse search, and microkinetics layers that power end-to-end catalyst discovery.
Activation barriers are computed from Flux energy and barrier terms. Experimental references are used to score the benchmark.