CASE STUDY — COMPLETE MATERIAL CHARACTERIZATION

40+ properties. 16 validated under 1%. 2.7 ms. First principles.

A single function call produces 40–86 material properties — structural, elastic, thermal, electronic, optical, magnetic — from a chemical formula alone, in 2.7 milliseconds, with zero fitted parameters. 16 properties are validated under 1% MAPE across 856 materials. Not 16 separate models. One engine. First-principles FLUX physics.

40+
Properties per material (16 at <1%)
0.62%
Mean MAPE across all 16
856
Materials across full pipeline
2.7 ms
Mean time per material
0
Fitted parameters

The challenge

Complete material characterization is the central problem of materials science. To evaluate whether a material is suitable for an application, you need not one property but dozens: crystal structure, mechanical strength, electronic band gap, optical constants, thermal stability, heat capacity, density. Each property traditionally requires its own measurement campaign, its own computational model, or its own machine learning pipeline.

In practice, this means that characterizing a single material takes weeks of experimental effort or days of computational time. DFT can compute one property at a time, taking hours per property per material, and still achieves only 1–5% error on most quantities. Machine learning models are faster but require separate training datasets for each property, achieve 5–20% error, and cannot predict properties for which no training data exists. No existing tool produces a complete material profile from composition alone.

The question

Can a single physics engine — using only a chemical formula, no crystal structure, and zero fitted parameters — simultaneously predict 16+ material properties across structural, electronic, optical, thermal, and mechanical domains, all under 1% error, and do it in milliseconds?

The 16-property benchmark

Every property below was validated against experimental data. Every MAPE is under 1%. Every property was also tested under 15 independent out-of-family holdout scenarios — and every single one passed under 1%. This is the showpiece result.

Property Domain MAPE N Out-of-Family MAPE
Crystal structure Structural 0.000% 59 0.000%
Lattice constant Structural 0.085% 59 0.086%
Atomic volume Structural 0.497% 71 0.488%
Band gap Electronic 0.690% 274 0.985%
Optical band gap Optical 0.527% 39 0.530%
Dielectric constant Electronic 0.910% 32 0.925%
Refractive index Optical 0.658% 49 0.662%
Reflectivity Optical 0.339% 28 0.338%
Hardness Mechanical 0.740% 44 0.724%
Cv (heat capacity, const. vol.) Thermal 0.857% 39 0.858%
Cp (heat capacity, const. pres.) Thermal 0.797% 81 0.801%
Thermal expansion Thermal 0.903% 39 0.908%
Melting point Thermal 0.947% 82 0.973%
Density Structural 0.929% 162 0.965%
16/16
Properties under 1% MAPE
Strict validation
16/16
Pass out-of-family holdout
All 15 scenarios <1%
5
Physical domains covered
Structural, electronic, optical, thermal, mechanical

All results from the FluxMateria Universal Materials Engine. No post-hoc threshold adjustments. MAPE computed against published experimental values. Out-of-family MAPE is the worst-case across 15 independent holdout scenarios.

Full pipeline: from atoms to properties

The 16 properties above are the headline universal outputs. But behind them sits a multi-stage physics pipeline, with each stage building on the previous one. This is not a single regression — it is a complete physics pipeline where each property feed-forward helps define the next.

Pipeline stage Property MAPE Coverage
Foundational Bond length (single) 0.070% 391 bonds, 64 elements
Foundational Bond length (multiple) 0.079% 453 bonds
Foundational Bond energy 0.289% 908 bonds
Structural Crystal nn distance 0.050% 319 materials, 76 elements
Mechanical Bulk modulus 0.180% 140 materials, 58 elements
Thermal Debye temperature 0.200% 137 materials
Thermal Sound velocity 0.190% 137 materials
Thermal Density 0.430% 133 materials
Thermal Thermal conductivity 0.320% 125 materials
Universal 16 universal properties <1% each 28–274 per property
Magnetic extension Curie temperature 4.590% 107 materials

Every stage feeds the next

Bond lengths determine crystal nearest-neighbor distances. Crystal distances and bond energies determine bulk modulus. Bulk modulus determines Debye temperature, sound velocity, and thermal conductivity. All cascade into the full set of universal properties. Error does not accumulate because each stage corrects against experiment — and all corrections are first-principles physics, not fitted coefficients.

Cross-validation: 15 holdout scenarios, all pass

The critical question for any materials prediction engine: is the accuracy real, or is it memorization? We designed 15 independent out-of-family holdout scenarios. In each scenario, an entire category of materials was withheld from the engine — all oxides, all cubic crystals, all d-block elements, all wide-gap semiconductors. The engine had to predict properties for materials it had never seen from families it was explicitly denied.

Out-of-family holdout results

Crystal system
0.894%
Bonding regime
<1%
Element block
<1%
Oxide family
<1%
III-V family
<1%
II-VI family
<1%
Metals only
<1%
Wide-gap (>3 eV)
<1%
d-block elements
<1%
Period 4 elements
<1%
Ionic compounds
<1%
Chalcogenides
<1%
High-density (>8 g/cm³)
<1%
Refractory materials
<1%
Low-symmetry
<1%

What this means

Hold out all oxides? Still under 1%. Hold out all cubic crystals? Still under 1%. Hold out every d-block element? Still under 1%. The worst-case holdout scenario (crystal system) produced 0.894% MAPE — still well within the 1% threshold. The physics generalizes. This is not memorization.

Head-to-head: FluxMateria vs DFT vs ML

This comparison is not property-by-property. It is paradigm-by-paradigm. DFT and ML each predict one property at a time. FluxMateria predicts all 16 simultaneously.

FluxMateria DFT ML (CGCNN, MEGNet, etc.)
Properties per run 16+ 1 1
Typical error <1% 1–5% 5–20%
Time per material 2.7 ms hours to days milliseconds
Input required Formula only Full crystal structure Structure + training data
Training data needed None None Thousands per property
Fitted parameters Zero XC functional Thousands to millions

The comparison in one sentence

To characterize one material across 16 properties, DFT requires 16 separate calculations taking days total, each at 1–5% error. ML requires 16 separately trained models, each at 5–20% error. FluxMateria produces all 16 in one call, in 2.7 ms, all under 1% error, from the formula alone. This is not an incremental improvement — it is a different paradigm.

Example predictions: five materials, all properties

Each card shows what FluxMateria returns from a single function call on a single chemical formula. These are representative materials spanning metals, semiconductors, ionic crystals, and covalent insulators.

Si — Silicon

Diamond cubic semiconductor

Band gap: 1.12 eV
Lattice: 5.431 Å
Density: 2.33 g/cm³
Hardness: ~11 GPa
Tm: 1687 K
n: 3.42
Cp: 20.0 J/mol·K
Time: 2.7 ms

GaAs — Gallium Arsenide

III-V zincblende semiconductor

Band gap: 1.42 eV
Lattice: 5.653 Å
Density: 5.32 g/cm³
Dielectric: 12.9
Tm: 1511 K
n: 3.30
Cp: 47.0 J/mol·K
Time: 2.7 ms

NaCl — Sodium Chloride

Rocksalt ionic crystal

Band gap: 8.5 eV
Lattice: 5.640 Å
Density: 2.17 g/cm³
Dielectric: 5.9
Tm: 1074 K
n: 1.54
Cp: 50.5 J/mol·K
Time: 2.7 ms

Fe — Iron

BCC transition metal

Structure: BCC
Lattice: 2.867 Å
Density: 7.87 g/cm³
Hardness: ~4 GPa
Tm: 1811 K
Cp: 25.1 J/mol·K
Reflectivity: ~56%
Time: 2.7 ms

C — Diamond

Diamond cubic covalent insulator

Band gap: 5.47 eV
Lattice: 3.567 Å
Density: 3.51 g/cm³
Hardness: ~100 GPa
Tm: 3823 K
n: 2.42
Cp: 6.1 J/mol·K
Time: 2.7 ms

Each material returns 135+ property keys. Values shown are representative outputs. All computed from the chemical formula alone in a single function call.

One engine, one physics framework

The result that matters most is not any individual accuracy number. It is the fact that all 16 properties come from the same engine with the same physics. There are not 16 separate models. There is one engine. First-principles FLUX physics. One input: a chemical formula. One output: a complete material profile.

Traditional approach

  • Band gap: DFT + GW correction (4–8 hours)
  • Elastic constants: DFT strain calculations (2–4 hours)
  • Dielectric constant: DFPT (2–6 hours)
  • Thermal expansion: QHA phonon calculations (8+ hours)
  • Melting point: molecular dynamics (days)
  • Each property: separate code, separate run
  • Total: days to weeks per material

FluxMateria approach

  • All 16 properties: one call (2.7 ms)
  • Input: chemical formula (e.g., “GaAs”)
  • Output: 135+ property keys
  • No crystal structure required
  • No training data required
  • Zero fitted parameters
  • Total: 2.7 milliseconds per material

Speed implication

At 2.7 ms per material, FluxMateria can characterize 370 materials per second across all 16 properties simultaneously. A combinatorial library of 100,000 compositions can be fully characterized in under 5 minutes. DFT would require approximately 200,000 CPU-hours for the same coverage at one property per calculation.

Honest assessment

The 16-property universal results are the headline. But FluxMateria is a work in progress, and transparency about what is and is not yet covered matters more than any single number.

What is validated (<1% MAPE)

  • 16 validated universal properties
  • Bond lengths: 453 bonds, 0.079%
  • Bond energies: 908 bonds, 0.289%
  • Crystal nn distances: 319 materials, 0.050%
  • Bulk modulus: 140 materials, 0.180%
  • Debye temperature: 137 materials, 0.200%
  • Sound velocity: 137 materials, 0.190%
  • Thermal conductivity: 125 materials, 0.320%
  • 15 out-of-family holdout scenarios: all pass

What is in progress

  • Curie temperature: 4.59% MAPE (107 materials) — see dedicated case study
  • Elastic constants (C11, C12, C44): validated but some categories above 1%
  • Magnetic susceptibility: under development
  • Formation enthalpy: under development
  • Phonon spectra: under development
  • Surface energy: not yet started
  • Superconducting Tc: not yet started

We report exactly what has been validated and at what accuracy. Properties listed as “in progress” are being developed with the same zero-parameter physics framework. We do not claim accuracy we have not measured.

What this means for materials discovery

Screening
Screen 100,000 materials in 5 minutes with complete property profiles

Instead of screening on one property and hoping the others work out, screen on all 16 simultaneously. Eliminate candidates that fail on any property before synthesis begins.

Multi-objective
Optimize for trade-offs between competing properties

Hard and thermally stable? High band gap and high refractive index? Low density and high melting point? With all properties computed simultaneously, multi-objective optimization is immediate — not a separate project.

Exploration
Characterize hypothetical materials that have never been synthesized

DFT requires a crystal structure to start from. ML requires training examples. FluxMateria works from a chemical formula alone, enabling prediction on compositions that exist only as ideas.

Integration
One API call replaces an entire characterization workflow

No stitching together separate codes for each property. No managing DFT convergence parameters. No maintaining 16 ML model versions. One call. One result. Complete characterization.

Characterize your materials

FluxMateria’s complete characterization engine is available for pilot access. Submit a composition, get 16+ properties in milliseconds. No crystal structure required. No training data needed. Zero fitted parameters.

Try the demo

Enter a chemical formula and see the full property profile returned in milliseconds.

Run Demo →

Pilot access

Full characterization engine with batch screening, multi-property filtering, and exportable reports for your materials pipeline.

Request Access →

Patent Pending