🔬 FLUXMATERIA — MATERIALS

Every materials property,
from composition alone

Band gap, formation energy, lattice parameters, thermal conductivity, Curie temperature, magnetic order — all from a composition string, no DFT setup, no training set. 313 benchmark materials across 14 families. The hub that feeds the Battery, Catalyst, Semiconductor, Surface & Contact, Design Studio, and Discovery sub-modules.

Band gap Formation energy Thermal conductivity Magnetic order No ML
313
Benchmark materials across 14 families
0.228 eV
Band-gap MAE across 1,048-material public set
4.6%
Curie-temperature MAPE (107 materials, 17 families)
0.64%
Lattice median error on the 15-material DFT cross-check
0
Machine-learning models trained
The breakthrough

DFT-competitive materials properties, at sub-millisecond per query.

DFT gives you real materials properties but takes hours to days per cell. ML force fields are fast but blind outside their training distribution. The Materials engine predicts band gap, formation energy, lattice parameters, thermal conductivity, and magnetic order from composition alone — deterministic, re-runnable, and sub-millisecond per query. 313 materials validated, 14 families, no machine-learning training.

What the Materials engine computes

The universal property calculator that feeds every materials sub-module.

📐

Band gap

0.228 eV MAE on the 1,048-material public benchmark; 0.338 eV on the 587 gap-bearing materials.

⚡

Formation energy

0.08 eV / atom MAE — screening-grade accuracy for stability ranking and hull-distance calculations.

📏

Lattice parameters

0.64% median error on the 15-material DFT cross-check panel (6.6% MAPE including layered cells that need structure input), at a tiny fraction of DFT runtime.

🔗

Crystal bond lengths

3.84% MAPE from formula on 352 scored cases out of 354. Context-assisted panel: 1.93% on 351 materials spanning 65 structure categories. Sub-millisecond predictions. See benchmark →

🌡️

Thermal conductivity

r2 = 0.87 correlation with experiment. Screening-grade ranking for thermoelectric and heat-management work.

🧲

Curie temperature

4.6% MAPE on 107 magnetic materials across 17 families, with a fixed family calibration — ML baselines typically sit at 15–30%.

🎯

Metal vs insulator

0.879 ROC-AUC on the 4,921-compound matbench experimental metal/insulator task. Gap-sign check before any downstream physics.

🔎

Property / elasticity bundle

Bulk modulus, Young’s modulus, sound velocity, Debye temperature — 9 elastic / thermal properties per composition.

🧬

Magnetic order

Ferro / antiferro / paramagnetic classification + saturation magnetisation for the ferromagnetic class — same engine.

How a property is computed

From composition to full property card in sub-millisecond.

1

Composition

Enter formula or pick from the 313-material curated set. The engine parses composition and classifies structural family automatically.

2

Atomic descriptors

Element-level physics primitives (ionisation, electron affinity, polarizability) pulled from the shared atomic-property engine.

3

Property compute

Closed-form expressions evaluate band gap, formation energy, lattice, thermal, magnetic in parallel. Sub-millisecond per property.

4

Classify & bundle

Metal / insulator classification; magnetic order classification; family assignment. The output is a complete property card, not a single scalar.

5

Hand-off

Bundle feeds Battery (electrode screening), Catalyst (surface properties), Semiconductor (mobility / Baliga FoM), Design Studio / Discovery (evolutionary search).

Why you can trust it

Every property validated on a public benchmark with the test-set MAE or MAPE reported.

0.228 eV
Band-gap MAE on the 1,048-material public benchmark — 0.338 eV on the 587 gap-bearing materials, 0.088 eV on the 461 metals at exp = 0. Family-stratified reports on the benchmark page.
0.08 eV/atom
Formation-energy MAE across the 313-material benchmark — screening-grade for stability triage.
0.64%
Lattice median error on the 15-material DFT cross-check (6.6% MAPE including layered cells), sub-millisecond per structure.
4.6%
Curie-temperature MAPE on 107 magnetic materials in 17 families, fixed family calibration. ML baselines: 15–30%.
r2 = 0.87
Thermal-conductivity correlation with experiment — ranking-grade for thermoelectric and heat-management screens.
Bit-for-bit
Deterministic: the same composition returns the same property card every run.

How FluxMateria compares

Head-to-head against the standard approaches to materials-property prediction.

MetricFluxMateriaDFT (PBE / HSE)ML force fieldMaterials Project lookup
Band-gap MAE0.228 eV0.5–1.0 eVDepends on setReference (exact)
Lattice error (DFT panel, median)0.64%1–2%1–3%Reference (exact)
Curie T MAPE4.6%10–20%15–30%Very limited
Runtime per material< 1 msHours to daysSecondsInstant lookup
Predicts new compositionsYesYes (slow)Within distributionNo
Training dataNo ML training; reference-calibrated termsNoneLarge datasetsData is the tool
Magnetic-order classificationBuilt-inPer-runNot alwaysCatalogue entry
Family-stratified reportPublishedPer-paperUsually missingN/A

The key insight: DFT is the gold standard for materials properties but too slow for library-scale screens. ML force fields are fast within their training distribution and unreliable outside it. Materials Project is exact but only for materials someone already measured. FluxMateria’s engine gives DFT-competitive numbers from composition alone, sub-millisecond per query, with no training set required. See the full materials benchmark →

Where Materials wins

The Materials engine feeds the sub-modules you actually use for design & discovery.

Flagship sub-module

Solar Device Studio

Carry crystalline-Si material state through optics, transport, recombination, junction behavior, and contacts to terminal Jsc, Voc, fill factor, and efficiency. All four accuracy gates pass for the qualified Al-BSF scope.Open Solar Device Studio →

Sub-module

Band gap + full electronic transport

Composition-only band-gap predictor with the complete electronic-property bundle — 0.228 eV MAE on the 1,048-material public set, ~1 ms per query, no machine-learning training. One call returns band gap, dopant classification, carrier concentration, mobility, conductivity, and heterojunction band-edge alignment. Works ab initio on novel candidates with no measured-property lookups.Open Band Gap →

Sub-module

Battery electrochemistry

Electrode-electrolyte free-energy calculations, phase stability, voltage windows. 0.149 V holdout MAE on voltage prediction.Open Battery →

Sub-module

Catalyst scoring

9-phase readiness scoring, activation barriers, microkinetics, ASF product distribution. 93.4% pairwise-ranking accuracy.Open Catalyst →

Sub-module

Semiconductor design

Live drift-mobility + Baliga / Johnson figures of merit across doping, temperature, alloy. Median error 9.3% μe / 6.6% μh on 15 agreed-reference semiconductors.Open Semiconductor →

Sub-module

Surface & Contact

Work function, Schottky barriers, band alignment — 0.063 eV MAE on 35 interface stacks.Open Surface & Contact →

Sub-module

Design Studio

Realtime generative design: spec in, ranked candidates out. 313 curated materials seed every population.Open Design Studio →

Sub-module

Evolutionary Discovery

Round-based evolutionary search toward a spec — superconductor, thermoelectric, battery cathode, custom target.Open Discovery →

Materials in the product

Real captures from the live application. Click any image to zoom.

Composition input with structural family classification
Composition inputEnter formula, pick from 313-material curated set. Family is classified automatically.
Full property card with band gap, lattice, thermal, magnetic
Property cardBand gap + formation energy + lattice + thermal + magnetic order + elastic bundle — one view.
Per-family benchmark breakdown
Benchmark breakdownFamily-stratified MAE / MAPE for band gap, lattice, Curie, and the thermal-conductivity correlation.
Hand-off to device and discovery sub-modules
Hand-offOne-click pipe to Solar / Battery / Catalyst / Semiconductor / Surface / Design Studio / Discovery.

Scope & Limitations

Strengths

  • Band gap, lattice, formation energy and Curie temperature at sub-millisecond runtime, each with its benchmark and basis stated.
  • 313 curated materials across 14 families, 1,048-material public band-gap benchmark.
  • One composition input returns band gap, formation energy, lattice, thermal, magnetic, and elastic bundle — not a single scalar.
  • Family-stratified benchmarks published — you see where the model is strong and where it drifts.
  • Feeds seven specialised sub-modules (Solar, Battery, Catalyst, Semiconductor, Surface, Design Studio, Discovery).

Known limitations

  • Strong-correlation systems (heavy 4f lanthanide oxides, some TM oxides) drift wider than the average MAE — flagged in the benchmark.
  • Amorphous solids, organic semiconductors, and most MOFs sit outside the current structure library.
  • Thermal-conductivity is ranking-grade (r2 0.87), not quantitative — use for triage, refine on the top picks.
  • Non-collinear magnetic order is a 2026 roadmap item; today’s engine classifies collinear ferro / antiferro / paramagnetic.

Bring Materials to your program

Pilot access includes the universal Materials engine, all seven sub-modules (Solar, Battery, Catalyst, Semiconductor, Surface, Design Studio, Discovery), and a Workspace seat for audit-ready runs.

Request Pilot Access →