Accuracy across materials, chemistry and life sciences.

FluxMateria benchmarks span chemistry, materials, ADMET, reaction mechanisms, spectroscopy, solvation, catalysis, battery, and synthesis-planning modules.

Find the performance most relevant to your R&D, then explore what FluxMateria can do on your own materials or molecules.

Strong results across FluxMateria.

0.989

DILI AUROC

96 public test compounds · identity, fingerprint and scaffold references excluded.
0.952 with identity/fingerprint exclusions only; same 96 cases.

Up to 16.8×

Lower torsion MAE than force fields

Benchmark descriptors · 98 matched cases.
Automatic SMILES: 4.7–7.0× lower MAE on 92 matched cases.

0.228 eV

Band-gap MAE

1,048-material reference cohort.

3.66×

Lower band-gap MAPE than PBE

10.3% vs 37.5% · same 10 gapped materials; specified PBE/GPAW setup.

0.281

Caco-2 MAE, log Papp

Current alternative mode · 182-case known development panel. Production-profile MAE 0.914 under its separate reference exclusions.

0.839

CYP3A4 inhibition AUROC

CYP3A4 · 2,467 reference-excluded test cases; AUPRC 0.814.
CYP2C9 AUROC 0.835; CYP2D6 0.821. Full endpoint results in the benchmark.

Key metrics at a glance

Summary of performance across core capabilities.

Bond Lengths

0.079%error
453 bonds, 64 elements

Single + multiple bonds across p, d, and s-block with row-level validation evidence.

Bond Energies

0.289%error
906 bonds, 64 elements

Singles, doubles, and triples. 870/906 within 1.0% in the public export.

Raw-Output Reference

0.176%MAPE
1,483 atomic and molecular reference targets

Raw Flux formula outputs only: no training, no calibration, no DFT/computed-only targets.

Throughput

10,000+mol/hr
full property panel

Single-threaded, no GPU. Scales linearly with cores for batch jobs.

ADMET Panel

0.989AUROC
DILI: 96 cases, scaffold exclusions

Eight endpoint families in one workflow. DILI AUROC 0.989 with test identities, fingerprints and scaffolds excluded; CYP3A4 AUROC 0.839 / AUPRC 0.814 on 2,467 reference-excluded cases. Explore the endpoint results and input modes.

Materials

0.228eV MAE
Band gap, 1,048 materials

Universal 16: in-sample consistency suite of the February 2026 engine, with inputs and calibration provenance declared. Gemstone color 19/19.

Solar Devices

4 / 4gates
Si Al-BSF terminal response

384-device internal qualification plus one independent same-device published-profile transfer check. External laboratory qualification remains pending.

Solvation

0.3295kcal/mol
FreeSolv hydration MAE, 642 cases

Sub-kcal/mol hydration accuracy: 642 experimental FreeSolv references, evaluated with the production explicit-water screening configuration.

MechanismOS

Real-timesteering
control surfaces + constraint optimizer

Category-defining reaction steering: live control surfaces, pathway boundaries, optimizer output, and evidence-pack export.

Catalyst Scoring

93.4%ranking
12 / 12 ranking tests passed

Production-stack catalyst benchmark with pairwise ranking fidelity, scenario alignment, inverse-search convergence, and experimental chemisorption calibration.

Benchmarks by module

Detailed performance data for each capability.

ADMET: measured performance across endpoints

Strong permeability, metabolism and safety-screening results across named product modes and evaluation cohorts.

0.989
DILI AUROC · scaffold-reference exclusions
0.952: identity/fingerprint only; same 96 cases
0.281
Caco-2 MAE · 182 cases
Alternative-mode development panel
0.821–0.839
CYP AUROC · 3 fixed tests
Reference identities and scaffolds excluded
CLint
Intrinsic-clearance and metabolic-stability estimates
  • DILI: AUROC 0.952 with identity/fingerprint exclusions and 0.989 with additional scaffold exclusions on the same 96 public test cases.
  • Caco-2: current alternative-mode MAE 0.281 on 182 development-panel cases; production-profile MAE 0.914 with stated reference exclusions.
  • CYP inhibition: AUROC 0.835 / 0.821 / 0.839 and AUPRC 0.692 / 0.562 / 0.814 for 2C9 / 2D6 / 3A4 on their separate reference-excluded fixed tests.
  • BBB: current reference-panel evaluation accuracy 86.03% on 7,803 scored of 7,807 requested, with four failures. Fixed reference-excluded test: 82.02% accuracy / 83.01% balanced accuracy on 406.
  • hERG: current reference-panel evaluation AUROC 0.804 and MAE 0.483 pIC50 on 8,879. Fixed reference-excluded test: AUROC 0.775 on 132.
  • Solubility: MAE 1.363 logS on 1,997 fixed test cases, including two fallback results; better than the training-label mean and median controls.
  • PPB: MAE 10.517 percentage points on 559 fixed test cases with test identities, fingerprints and scaffolds excluded.
  • CYP3A4 inducer: 0.9350 balanced accuracy on the primary external holdout.

These measurements describe currently evaluated product modes. Each result retains its own cohort, metric and reference policy. The current ADMET workflow includes clearance and metabolic-stability estimates in addition to the endpoint scores shown here.

Full Results → DILI Benchmark Caco-2 Benchmark CYP3A4 Inducer Methodology

🔬 Materials

0.228 eV
Band Gap MAE
1.1668%
Core S2 MAPE (not refit per fold)
<1%
Universal 16, in-sample
2.741 ms
Universal strict runtime

Band Gap Benchmark

1,048 materials
Overall MAE0.228 eV Metals (exp = 0)0.088 eV Non-metals (exp > 0)0.338 eV

Core Holdout (5 properties)

Five-property core: family (S2) and interaction (S3) partitions; fixed calibration retained
S2 (family holdout)1.17% S3 (interaction holdout)1.38%

Universal benchmark (16 properties)

In-sample consistency suite (Feb 2026); calibration terms set on the same reference set
All 16 strict<1% Runtime mean2.7 ms Cohort856 compounds Gemstone color match19/19
What this means: FLUX now has two primary validated tracks: core thermo-mechanics partitions with fixed calibration retained, and the Universal consistency suite with 16 rows covering 14 quantities at 2.7 ms per material, with reference overlap and evaluation scope declared. It also includes a curated mini-benchmark showing defect-context color flexibility for real-time UI exploration.
Universal Benchmark → Solar Device Studio Band Gap Benchmark Crystal Bond Lengths Module Page

Battery Electrochemistry

1.0
Family Accuracy
0.149 V
Holdout Voltage MAE
5 / 5
Scenario Alignment
26.8 s
End-to-End Workflow
  • Calibrated holdout benchmark tracks capacity, voltage, transport, cycle, electrolyte, interface, cost, and manufacturing together ✓
  • Energy-dense cobalt-free screen lifts LiMnO2 to the top ✓
  • High-voltage frontier screen lifts LiNiPO4 to the top ✓
  • Fast-charge and cycle-life screens surface transport- and stability-led families instead of a single default winner ✓
  • The same pipeline yields different leaders for bulk, interface, battery-native, and build questions ✓

This benchmark validates the battery-native decision layer as a screening and prototype-handoff engine, not as a replacement for electrochemical lab validation.

Full Results → Case Study Module Page

Catalyst Scoring

12 / 12
Ranking Tests
12 / 12
Top-1 (run 17 Apr 2026)
93.4%
Pairwise Accuracy
35 / s
Full-Stack Throughput
  • Measured through the full production scoring path, not a simplified shortcut ✓
  • All 96 benchmark references were FLUX-enriched in the published public run ✓
  • The corrected API path now clears FT activity, FT support, ammonia support, and WGS ordering together ✓
  • Inverse search converges to real industrial catalyst families and chemically serious exclusion lanes ✓
  • The public benchmark and catalyst case study now share the same API-only narrative ✓

This benchmark validates FluxMateria as a catalyst ranking and inverse-discovery engine. Physical synthesis, reactor testing, and long-run deactivation work still remain the next laboratory step.

Full Results → Case Study Module Page

Activation Barrier Prediction

0.236 eV
Combined MAE
0.147 eV
4d-series MAE
93%
Within 0.5 eV
0
DFT inputs / training
  • Predicts surface-reaction activation barriers from Flux energy and topology terms ✓
  • 29 published literature reactions: N₂, H₂, O₂, CO dissociation and C-H activation across 13 transition metals ✓
  • Matches single-method DFT accuracy (PBE ~0.20-0.30 eV) at analytical speed — microseconds per prediction ✓
  • 100% within 0.5 eV for N₂, H₂, and O₂ dissociation families ✓
  • Feeds the catalyst-scoring and microkinetics layers for end-to-end catalyst discovery ✓

Production-ready for catalyst screening, ranking, and inverse discovery. Quantitative turnover-frequency prediction is at the edge of usefulness at this MAE; same is true for single-method DFT.

Full Results → d-Band Benchmark Catalyst Benchmark

Metal / Insulator Classification

0.879
ROC-AUC
4,921
Compounds
0
ML Models Trained
100%
Deterministic
  • Classifies metal vs. insulator from composition alone on the matbench experimental is-metal task (4,921 compounds, 5-fold CV) ✓
  • 0.879 ROC-AUC with no machine-learning training — above standard featurized-ML baselines ✓
  • Resolves band metals, Mott / charge-transfer insulators, mixed-valence conductors, and rare-earth f-electron systems ✓
  • Fully deterministic — no random seeds; millisecond runtime for full-library screening ✓

Metallicity is the front-end gate of the materials band-structure engine. Best-in-class among physics-based methods without machine-learning training; heavy 4d/5d narrow-gap semiconductors remain the hardest cases.

Full Results → Band Gap Benchmark Module Page

d-Band Center Descriptor

0.197 eV
Combined MAE
100
Multi-source Cases
0
ML Models Trained
beats ML
vs linear / kNN / RF
  • Central descriptor in transition-metal catalysis — predicted from Flux atomic and surface descriptors ✓
  • 100-case benchmark: 27 pure TMs, 41 facets, 32 binary alloys — all with published literature targets ✓
  • Facet-specific MAE of 0.154 eV across (111), (100), (110), (211), (0001) and stepped surfaces ✓
  • Outperforms linear regression, k-NN, and random-forest baselines fitted on the same atomic descriptors ✓
  • Cross-validated against five independent literature sources (HN14, K04, GN09, N95, CM20) ✓

The d-band descriptor feeds downstream into the catalyst scoring and inverse-discovery layers. Production-ready for transition-metal catalysis workflows; rare-earth and Pt-3d skin alloys remain known weak spots.

Full Results → Catalyst Benchmark Module Page

🧲 Curie Temperature

4.6%
Overall MAPE
−0.03%
Mean Bias
107
Magnetic Materials
17
Material Families
  • 4.6% MAPE across 107 materials from composition with magnetic closure, branch overrides, and calibration notes ✓
  • 17 families: ferrites, rare-earth intermetallics, double perovskites, manganites… ✓
  • 84.1% within 5%, 90.7% within 10% of experimental Tc ✓
  • Near-zero bias (−0.03%) — no systematic over- or under-prediction ✓
Full Results → Module Page

🧪 DFT Cross-Check

10.3%
Band gap MAPE (PBE 37.5%)
3.6%
Magnetic moment MAPE; Fe / Ni (N = 2)
0.64%
Lattice median (all 15)
~104×
Per-call timing vs GPAW-PBE EOS; see workload scope
  • Head-to-head with GPAW PBE on 15 canonical materials run locally on identical inputs ✓
  • Three-layer comparison covering lattice, band gap, magnetic moment, and bulk modulus ✓
  • Engine band gap median 3.0% vs PBE 36.5%; bulk modulus MAPE 10.3% on the 12 in-scope cells. FluxMateria production configuration evaluated 23 Oct 2026 ✓
  • Reproducible: manifest, DFT settings, and per-material results downloadable as JSON / CSV / MD ✓

Si, Ge, GaAs, GaN, ZnO, MgO, TiO2, NaCl, Al, Cu, Fe, Ni, graphite, h-BN, MoS2. Two tiers shipped (single-point + EOS-derived B); a third (DFPT phonons + dielectric function) is on the roadmap.

Full Results → Case Study Module Page

⚡ Carrier Mobility

6.6% / 21.7%
Hole median / MAPE
9.3% / 22.2%
Electron median / MAPE
15
Agreed-reference semiconductors
4
Material Families
  • Mobility at 300 K on 15 materials selected for agreement of two hole references within 1%; electrons scored against one reference ✓
  • 4 families: III-V, II-VI, IV-VI, elemental semiconductors ✓
  • MAPE 21.7% (holes) and 22.2% (electrons); 73% of hole predictions within ±15% ✓
  • The reference sets disagree on 8 further materials (e.g. SiC: 120 vs 40 cm²/V·s); those are not scored against either side ✓
Full Results → Module Page

⚛️ Atomic & Magnetic Properties

2.5%
EN MAPE
1.7%
IE MAPE
84
MM Materials
5
Properties
  • Electronegativity 2.5% MAPE (75 elements), ionization energy 1.7% (27), electron affinity 1.0% (28) ✓
  • Magnetic moment: 100% pass (84/84 materials); metallic intermetallics 3.1% MAPE ✓
  • Saturation magnetization: 100% pass (10/10 materials at ±50% tolerance) ✓
  • Atomic properties and magnetic subproperties carry separate basis notes ✓
Full Results → Module Page

📈 Spectroscopy

5.4%
UV-Vis Error
<1%
IR Error
0.3-0.5
NMR MAE (ppm)
119
UV-Vis Molecules
  • UV-Vis: 5.4% mean error, 119 molecules ✓
  • IR: <1% error, 32 NIST molecules validated ✓
  • NMR: 0.3-0.5 ppm MAE, 10 reference molecules, 1H and 13C ✓
Full Results → Module Page

⚗️ Mechanism Discovery

99.4%
Mechanism Accuracy
334/336
Cases Correct
7.5
kJ/mol MAE
74.4%
Barriers within ±10 kJ/mol (250/336)
  • 334/336 experimental test cases: SN2 115/115, SN1 90/90, E2 68/68, E1cb 16/16, E1 45/47 ✓
  • 10,000 random physical consistency tests ✓
  • Head-to-head comparison with DFT (B3LYP) ✓
  • Every prediction traceable and reproducible ✓
Full Methodology & Results → Module Page

MechanismOS

Real-time
Mechanism Steering
100%
GOLD Direct Ea (154/154)
94.94%
SILVER Arrhenius Ea (1256/1323)
Evidence pack
Audit-ready export
  • SOTA real-time mechanism steering: control surfaces, pathway boundaries, constraint optimizer, and evidence-pack export ✓
  • GOLD: 154/154 direct measured activation barriers passed under fixed benchmark criteria ✓
  • SILVER: 1256/1323 Arrhenius-derived barrier checks passed ✓
  • Official experimental source provenance documented per benchmark tier ✓
Full Results → Module Page

🧪 Synthesis Planning

3.1%
Barrier MAE
29/29
Reaction Types
200
Specific Reactions
<50ms
Per Plan
  • 29 reaction-type barriers at 3.1% MAE (100% pass rate) ✓
  • 200 specific reactions at <1% MAE (72 exact matches) ✓
  • 15 disconnection SMARTS patterns validated ✓
  • All barriers fully auditable and reproducible ✓
Full Results → Module Page

🔥 Reaction Enthalpy

NEW
3.5%
MAPE
157
Reactions Tested
89%
Within 5%
<1ms
Per Reaction
  • 157 reactions from NIST WebBook at 3.5% MAPE, 10.0 kJ/mol MAE ✓
  • 12 categories: combustion, radical, formation, halogen, nitrogen, ozone ✓
  • Hess’s law with documented species resolution + universal bond engine ✓
  • Phase notation: C(s), C(g), H2O(l) — disambiguates reference states ✓
Full Results →

⚡ Electron Transfer

26/26
Tests Pass
2–3×
Tunneling Enhancement
Literature
Decay constant match
~150ms
Per Pair
  • Marcus rate constants with FLUX tunneling corrections ✓
  • Through-bond decay constant matches literature ranges ✓
  • Normal, activationless, and inverted Marcus regimes ✓
  • All coupling deterministic and traceable ✓
Full Results → Module Page

🧪 Solvation

Explicit water
FreeSolv benchmark route
0.3295
MAE (kcal/mol)
642
FreeSolv Cases
4
Native Non-Water Carriers
  • Explicit-water screening benchmark: 0.3295 kcal/mol MAE on 642 experimental FreeSolv references ✓
  • Official packet includes summary JSON, case CSV/JSON, and methodology ✓
  • Water externally benchmarked; methanol, ethanol, acetonitrile, and DMSO tracked ✓
Full Results →

🧬 BioTarget

0.617
Pearson r (CASF-2016)
95%
MoA Accuracy
1.335
MAE (pKi)
10,065
Targets
  • Binding affinity: Pearson r = 0.617 on CASF-2016 (269 complexes, deposited crystal poses) ✓
  • MoA prediction: 95% accuracy on 100 ChEMBL pairs ✓
  • Selectivity profiling: planned ⏳
Full Results →

⚛ Chemistry

0.079%
Bond Length Error
0.289%
Bond Energy Error
453 / 906
Length / energy counts; separate panels
64
Elements
  • Atomic and molecular reference benchmark: 1,483 scalar targets across 11 families, 0.176% weighted raw MAPE ✓
  • Bond lengths: 453 bonds (391 single + 62 multiple), 0.079% mean error ✓
  • Bond energies: 906 bonds, 0.289% mean error, 870/906 within 1.0% ✓
  • Flux-encoded bond-family formulas with published benchmark provenance notes ✓
  • Coverage: 24 p-block + 30 d-block + 10 s-block elements ✓
Reference Benchmark → Chemistry Results Module Page

↻ Torsion Barriers

0.83
kJ/mol MAE (99 rotors)
0
Empirical torsion tables
11.6–16.8×
Lower MAE than Sage 2.2 / GAFF2 / MMFF94 (same panel)
14
Rotor Classes
  • 0.83 kJ/mol MAE across 99 experimentally-measured rotational barriers (panel used in development); 1.43 kJ/mol from SMILES on 35 molecules ✓
  • Same-set head-to-head: 11.6× lower MAE than OpenFF Sage 2.2, 13.2× than GAFF2, 16.8× than MMFF94 ✓
  • 72% of cases within ±1 kJ/mol, 90% within ±2 kJ/mol of experiment ✓
  • Covers alkanes, ethers, amines, peptide ω, esters, acrylates, halides, X-X rotors and aromatic carbonyls ✓
Full Results → Case Study

How to read the measurements

📋

Publish the methodology

How we test, what datasets we use, how we measure. No hidden assumptions.

⚖️

Show head-to-head comparisons

Against established tools where possible. Fair comparisons, same test sets.

🎯

Document validation scope

Where predictions are most and least reliable. We tell you the boundaries.

🔄

Provide evidence context

Methodology, scope notes, and public artifacts or reviewer-accessible evidence packets where appropriate.

Independent Validation

Independent evaluation can assess performance on your application.

FluxMateria is now opening external validation tracks so researchers can choose blind datasets, define metrics, and score frozen predictions independently.

Available tracks include chemistry core, materials holdouts, life-science / ADMET, reaction mechanisms, and experimental validation.

Open validation program →

Validation scope

Where FluxMateria predictions are most and least reliable.

We document the boundaries of reliable prediction space:

  • ⚠ Novel chemotypes far from validated chemical space
  • ⚠ Specific endpoints with limited experimental data
  • ⚠ Edge cases identified through validation

Confidence indicators in predictions reflect these boundaries. Low confidence = verify experimentally.

See the value on your data

Choose one property or ranking task on a small dataset you have already measured. Your team scores the frozen predictions against its own data.

Discuss a blind benchmark