⚗️ FLUXMATERIA — CHEMISTRY

Catalyst-native screening
at 35 candidates per second

From a composition alone, rank candidates across eight industrial families, simulate the full catalytic cycle with activation barriers and turnover frequencies, solve steady-state microkinetics, and run inverse discovery under constraints. DFT-competitive accuracy with no machine-learning training, microseconds per prediction.

8 industrial families 9-phase readiness scoring Full catalytic cycle + microkinetics Electrochemistry: potential + pH Mini-MKM composition-dependent rates Selectivity-aware operating window ASF product distribution (FT) Inverse discovery with constraint DSL
93.4%
Pairwise ranking accuracy on the catalyst benchmark
12/12
Top-1 on the same benchmark (run 17 Apr 2026)
0.667 eV
vs 28 primary-lit experimental points
0.236 eV
Activation-barrier MAE (29 cases)
35 / s
Full-stack throughput per candidate
0
Training data · no ML training
The breakthrough

Catalyst screening at DFT accuracy, at ML speed.

DFT is honest but too slow to screen thousands of candidates. ML force fields are fast but degrade out-of-distribution and need millions of DFT points to train. Catalyst runs the full cycle — barriers, turnover, selectivity, microkinetics, electrochemistry — in microseconds per candidate with no machine-learning training. Same accuracy band as PBE-DFT on published benchmarks. Unlimited out-of-distribution reach.

What the Catalyst module gives you

From a single composition to an auditable catalyst decision package.

🏭

Family classification + readiness

Assigned to one of 8 industrial families (ammonia synthesis, Fischer-Tropsch, selective oxidation, WGS, HDS, reforming, electrocatalysis, hydrogenation) with a 0–100 readiness score.

📏

9-phase readiness breakdown

Activity, selectivity, thermal stability, poison resistance, regenerability, cost, supply-chain risk, toxicity, novelty — each scored independently so trade-offs are explicit.

🔁

Full catalytic cycle

Elementary steps, activation barriers, transition-state positions, Eyring rate constants, ΔH / ΔG / Keq, predicted turnover frequency, rate-limiting step flagged.

🌋

Sabatier volcano position

Descriptor binding energy placed on a family-specific volcano with weak / optimal / strong classification. Descriptor and peak resolved in the same engine pass for scale consistency.

⚙️

Steady-state microkinetics

Surface coverages, net elementary-step rates, gas-phase TOF, and Degree-of-Rate-Control sensitivity for every step in the cycle.

⚡

Electrocatalysis

Applied potential (vs SHE) and electrolyte pH propagate into proton-coupled electron-transfer barriers. HER / OER / ORR ready with PCET coverage physics.

📊

Operating window (T, P)

2D heatmap of TOF × selectivity over (T, P). Feasibility contour identifies zones where both activity and selectivity clear thresholds. ASF chain-growth α included for Fischer-Tropsch.

🔎

Inverse discovery

Constraint-driven search with 9 presets (from “cheapest ammonia catalyst” to “non-silver EO”) over 10 stoichiometry templates, Pareto visualisation, and novelty flagging.

The catalyst-native pipeline

Six passes from composition to audit-ready decision package.

1

Classify family

Composition in → industrial family out. Readiness score and family-specific descriptor selection locked in.

2

Simulate the cycle

Elementary steps, barriers, Eyring TOF, rate-limiting step. Three internally-consistent physics paths reported.

3

Solve microkinetics

Steady-state coverages, composition-dependent rates, DRC sensitivity. Captures Haber-Bosch pressure order and FT Langmuir inhibition.

4

Map the window

TOF × selectivity heatmap across (T, P). Feasibility contour. ASF distribution for FT candidates.

5

Band the uncertainty

Three TOF estimates (cycle-native / BEP / full MKM) report a method-spread band. Outliers flagged.

6

Package decision

Readiness + cycle + window + uncertainty + inverse-discovery suggestions exported as one audit-ready packet.

Why you can trust it

Validated against published benchmarks and primary literature, not internal splits.

93.4%
Pairwise ranking accuracy on the catalyst benchmark. The number that actually drives portfolio triage.
12/12
Top-1 on the same benchmark — the single-best-candidate call across families (run 17 Apr 2026).
0.667 eV
vs 28 primary-literature experimental points. Calibrated directly against measured activation energies, not model-to-model.
0.236 eV
Activation-barrier MAE across 29 cases — in the same band as PBE-DFT, at microsecond-per-call throughput.
35 / s
Full-stack throughput: family + cycle + microkinetics + operating window + uncertainty per candidate.
0
Training data. No ML training. Unlimited out-of-distribution reach.

How FluxMateria compares

Head-to-head against the two dominant catalyst-modelling approaches.

CapabilityFluxMateriaPBE-DFTML force fields (MACE, GemNet)
Activation-barrier MAE0.236 eV0.20–0.30 eV0.10–0.20 eV
d-band descriptor MAE0.197 eV0.15–0.30 eV0.15–0.25 eV
Training dataNoneNone (full DFT per case)106–107 DFT points
Fitted parameters00millions
Time per predictionmicrosecondsCPU-hoursseconds (GPU)
Out-of-distributionUnlimitedUnlimitedDegrades
Full cycle + microkineticsBuilt-inManualNot provided
Inverse discovery9 presets + DSLNot providedNot provided

The key insight: DFT is the reference — but it can’t screen thousands of candidates in a week. ML force fields can, but they degrade the moment you leave the training distribution. Catalyst matches PBE-DFT on barriers and d-band descriptors at microsecond throughput, with unlimited out-of-distribution reach and no training data required. See the full benchmark →

Where Catalyst wins

Catalyst-discovery workflows where DFT is too slow and ML force fields are out-of-distribution.

Use case 1

Inverse discovery

“Cheapest ammonia catalyst”, “non-silver ethylene oxide”, “novel compositions” — 9 constraint presets, Pareto-front visualisation, literature-novelty flagging.

Use case 2

Fischer-Tropsch scouting

Predicted ASF α, full C1–C19+ distribution, bucketed to gasoline / diesel / wax. Co α ≈ 0.80, Fe α ≈ 0.84 — in the industrial range.

Use case 3

Electrocatalysis design

Applied potential + pH propagate into PCET barriers. HER / OER / ORR ready with metal-dependent Heyrovsky barriers and proper coverage physics.

Use case 4

Operating-window sizing

2D (T, P) heatmap of TOF × selectivity. Find zones where both activity and selectivity clear thresholds — before committing reactor time.

Use case 5

Supply-chain triage

Cost, supply-chain risk, toxicity, regenerability scored per candidate. Ship only the catalysts that survive the full 9-phase readiness filter.

Use case 6

Portfolio rebalancing

Rank an internal library against a new spec in seconds. Keep the full audit trail for the committee review.

Scope & Limitations

Strengths

  • Same-band accuracy as PBE-DFT on barriers and d-band descriptors, at microseconds per call.
  • Full cycle + microkinetics + electrochemistry + inverse discovery — not just a surrogate.
  • 93.4% pairwise ranking accuracy on the catalyst benchmark — the number that drives portfolio triage.
  • No machine-learning training; the scoring layer combines physics outputs with stated engineering criteria.
  • Method-spread uncertainty band from three internally-consistent physics paths per TOF prediction.

Known limitations

  • Scope is heterogeneous / electro- catalysis; homogeneous single-site catalysis is on the roadmap but not in v1.
  • Operating-window heatmaps use family-default selectivity models; bespoke selectivity physics requires a pilot workshop.
  • Inverse discovery over 10 stoichiometry templates covers the common industrial families; exotic scaffolds need explicit seeding.
  • ASF product distribution shipped for Fischer-Tropsch; other chain-growth regimes (methanol-to-hydrocarbons) are on the roadmap.

Catalyst decisions deserve catalyst-native reasoning.

Pilot access includes Catalyst, the full Chemistry stack, MechanismOS, and a Workspace seat for audit-ready runs.

Request Pilot Access →