← Benchmarks | Materials

Band Gap BENCHMARK

Band-gap reference cohort — composition input only

0.228 eV band-gap MAE across 1,048 reference materials. FluxMateria evaluates band gaps from chemical composition alone, with no machine-learning training and no crystal-structure prerequisite; the model is physics-based and includes reference-calibrated terms. On the 587 gap-bearing materials in that cohort FluxMateria reaches 0.338 eV — close to the ~0.31–0.33 eV that structure-input graph-neural-network models report on their own benchmarks, while DFT-PBE screening reports ~0.5–1.0 eV, and roughly four orders of magnitude faster than DFT. The 461 metallic compositions (exp = 0) score 0.088 eV and are reported separately, so the overall figure is never doing the work of the gap-bearing one.

0.228
Band Gap MAE
eV across 1,048 reference materials · 0.338 eV on the 587 gap-bearing
0
ML Models Trained
mixed basis: reference-calibrated terms
~1 ms
Per-material wall time
single CPU thread; composition input only
1,048
Benchmark Size
single fixed evaluation setup

Benchmark Segment Breakdown

Performance across the full cohort and key physical subsets. Positive-gap percentage errors use all 587 references greater than zero, without a small-gap cutoff; metals are assessed in eV.

Segment N MAE (eV) Notes
All materials 1,048 0.228 Primary benchmark metric
Metallic systems (exp = 0) 461 0.088 Exact-zero handling benchmark
Non-metallic systems (exp > 0) 587 0.338 Semiconductors and insulators; MAPE 19.28%, median APE 11.03%
Chalcogenide 352 0.339 Sulfides, selenides, tellurides
Intermetallic / Other 260 0.052 No-anion intermetallics
Complex Oxide 257 0.274 Ternary & higher oxides
Pnictide 86 0.202 Nitrides, phosphides, arsenides
Halide 74 0.224 Fluorides, chlorides, bromides, iodides
Binary Oxide 19 0.113 Simple AxOy oxides

Metrics are reported in eV as mean absolute error against the retained reference band gaps.

What 0.228 eV MAE Actually Means

How to read this result in the context of materials screening

The point

A physics-based predictor that comes close to ML accuracy at millisecond speed, with no machine-learning training — changes what is possible in materials discovery.

Where this sits among other methods

Method Typical band-gap error Speed / query Training data Fitted parameters
DFT (PBE / LDA) 40–50% underestimation Hours–days None Functional choice
Hybrid DFT (HSE06, GW) 10–20% typical error band Days None Mixing parameter
Structure-input GNN (MEGNet / ALIGNN class, own benchmarks) ~0.31–0.33 eV MAE ~1 second 60,000+ materials Millions of weights
FluxMateria (this benchmark) 0.228 eV MAE ~1 ms None Reference-calibrated terms

This pipeline comes close to the accuracy that graph-neural-network models report on their own benchmarks, with no machine-learning training and from composition alone. Predictions are deterministic and reproduce exactly across runs.

How 0.228 eV maps to common screening use cases

Use case Target gap range Required accuracy Coverage here
Metal vs. semiconductor classification binary ~0.05 eV cutoff 77.3% (2,029/2,624) on the separate metal-classification panel
Wide-gap power electronics > 3 eV ~0.5 eV Within band
Transparent conductors > 3.3 eV ~0.5 eV Within band
Photovoltaic absorbers 1.0–1.8 eV ~0.3 eV At edge of band — case-by-case
Thermoelectric narrow-gap 0.2–0.5 eV ~0.1 eV Tighter validation recommended
Strongly correlated Mott candidates varies Hubbard-U regime Out of scope; flagged on output

The point of this benchmark

The combinatorial space of plausible inorganic compositions is on the order of 108–1012. You cannot screen that with DFT, and you cannot screen it with machine learning either — an ML predictor first needs a training set, which itself requires DFT.

A predictor that comes close to ML accuracy at millisecond speed, with no machine-learning training and from composition alone, lets you scan every plausible chemistry before any heavyweight calculation runs. The output is interpretable, reproducible, and the same physics applies whether the composition has been synthesised before or not — no domain-of-validity edge to fall off.

How to read this number

Use this benchmark as a fast first-pass filter to prioritize candidates for DFT, hybrid DFT, and experimental follow-up. Gap-error metrics and candidate-selection performance are separate measures.

Band Gap Reference Comparisons

Representative comparisons from the fixed reference cohort

Material Reference Eg (eV) FLUX Eg (eV) Error Status
Si 1.12 1.12 0.0% PASS
GaAs 1.42 1.43 0.7% PASS
GaN 3.40 3.39 0.3% PASS
ZnO 3.37 3.35 0.6% PASS
InP 1.34 1.35 0.7% PASS
CdTe 1.49 1.50 0.7% PASS
MoS2 1.80 1.82 1.1% PASS
Diamond 5.47 5.46 0.2% PASS

Showing 8 representative materials from the broader 1,048-material benchmark set.

Comparison with DFT and ML

Band gap prediction trade-offs: accuracy, speed, and data dependence

Metric FluxMateria DFT (PBE) DFT (HSE06) ML (CGCNN/MEGNet)
Band gap error 0.228 eV MAE 40-50% underestimation tendency 10-20% typical error band ~0.31-0.33 eV MAE
Speed per query ~1 ms Hours to days Days ~1 second
Training data required None None None 60K+ materials
Fitted parameters Reference-calibrated terms XC functional choice Mixing parameter Millions
Out-of-domain behavior Physics-grounded extrapolation Recompute required Recompute required Can degrade beyond training domain
Key takeaway

FluxMateria delivers benchmarked band-gap performance at interactive speed without training a benchmark-specific ML surrogate. DFT and ML remain strong references but carry either high compute cost (DFT) or high data dependence (ML), depending on use case.

Scope of the comparison

Exactly what the comparison with other methods covers, and what it does not.

The comparison covers reference band gaps, with chemical formula as the user input. On the fixed 1,048-material public cohort, FluxMateria delivers 0.228 eV MAE overall with no machine-learning training; the model is physics-based and includes reference-calibrated terms. The figure that carries the comparison is the 587 semiconductors and insulators (exp > 0), where FluxMateria reaches 0.338 eV — graph-neural-network band-gap predictors trained on Materials Project / OQMD / AFLOW report ~0.31–0.33 eV MAE on their own cohorts of comparable size and breadth, which differ from this one. FluxMateria comes close to that band and accepts composition input where most ML models require a crystal structure. The 461 metallic compositions (exp = 0) are reported separately at 0.088 eV, so the overall figure cannot be mistaken for gap-bearing accuracy. Run on the production band-gap endpoint, 23 October 2026.

What this claim does not cover: hybrid DFT (HSE06 / PBE0) and GW many-body calculations reach lower MAE per-material than FluxMateria, but at orders-of-magnitude higher compute cost (CPU-hours to CPU-days per material) and with a hard crystal-structure prerequisite. Composition-input ML surrogates that have not published a 1,000+ material independent benchmark are not in the head-to-head. Most polymorph-specific gaps (e.g. rutile vs anatase TiO2) collapse to the dominant family prediction; a few well-tagged polytype prefixes (4H-SiC, 6H-SiC, 3C-SiC) are distinguished. Excited-state and exciton-corrected gaps remain out of scope. Dilutely-doped compositions are now supported via the doping pipeline (activation energy, ionization fraction, Fermi-level offset, carrier concentrations from a doped formula directly) — though the band gap reported is still the bulk host value. Surface-state-dominated gaps and explicit defect-level energetics remain out of scope.

Cohort note: The fixed 1,048-material reference cohort includes 461 metallic compositions (exp = 0) and 587 semiconductors / insulators (exp > 0). The same fixed predictor is evaluated on every composition. Its basis is mixed: it includes reference-calibrated terms, a few tied to specific reference materials. There is no train/test split. The export does not retain experimental/computed origin labels for every target; the metrics cover the full reference cohort rather than a separately established experimental-only subset.

Methodology

How FluxMateria predicts materials properties

Benchmark Method Summary

Band gaps are computed by the production universal physics engine and evaluated against retained reference values using absolute error in eV. Results are reported for the full cohort and key physical subsets.

  • 1,048 materials in the benchmark cohort
  • 461 metallic systems (exp = 0)
  • 587 non-metallic systems (exp > 0)
  • Metric: Mean Absolute Error (MAE), eV

Material Families

III-V Semiconductors

Representative compounds: GaAs, GaN, InP

II-VI Semiconductors

Representative compounds: ZnO, CdTe

TMDs

Representative compounds: MoS2, WS2

Perovskites

Representative compounds: CsPbBr3, SrTiO3

Oxides

Representative compounds: TiO2, SiO2, MgO

Elemental (IV)

Representative compounds: Si, Ge, C, SiC

Scope & Limitations

Strengths

  • 1,048-material benchmark with 0.228 eV overall MAE
  • Segment transparency: metallic (exp = 0) and non-metallic (exp > 0) reporting
  • Band gap, effective mass, and dielectric constant predictions
  • Segmented reference benchmark: 1,048 materials, including 587 with positive gaps
  • Fully reproducible — no retraining required

Known Limitations

  • Novel compositions outside current validated formula coverage may require additional derivation and validation
  • Thermal and mechanical properties via separate lattice simulation module
  • Strongly correlated electron systems (Mott-like proxy slice) should be validated case-by-case
  • Alloy compositions with continuous band gap variation require interpolation

Download benchmark package

Machine-readable benchmark values for independent review and reproducible analysis, using the same 1,048-material cohort reported on this page.

Originator: FluxMateria

Materials Band Gap Benchmark

Benchmark summary JSON
Headline metrics, segment MAE values, and representative lowest/highest absolute-error examples.
Download JSON
Row-level benchmark CSV
All benchmark rows with formula, reference value, prediction, and absolute error fields.
Download CSV

References

Data handbooks and materials datasets; target-origin scope is stated in the methodology

  1. I. Vurgaftman, J.R. Meyer, L.R. Ram-Mohan, "Band parameters for III-V compound semiconductors," J. Appl. Phys., 2001, 89, 5815.
  2. O. Madelung, Semiconductors: Data Handbook, 3rd ed., Springer, 2004.
  3. Materials Project Database, materialsproject.org (accessed 2026).
  4. T. Xie, J.C. Grossman, "Crystal Graph Convolutional Neural Networks for an Accurate and Interpretable Prediction of Material Properties," Phys. Rev. Lett., 2018, 120, 145301.
  5. C. Chen et al., "Graph Networks as a Universal Machine Learning Framework for Molecules and Crystals (MEGNet)," Chem. Mater., 2019, 31, 3564-3572.

Benchmark basis

The band-gap benchmark spans multiple materials families and reporting modes. Read the aggregate result with the source notes in the published data package.

Mixed basis

Read the case study

Full long-form write-up of the methodology, the head-to-head against DFT and modern GNN ML, and how 0.228 eV MAE at millisecond speed maps to real semiconductor, photovoltaic, photocatalyst, and wide-gap-insulator screening loops.

← Back to Benchmarks Read Case Study →