0.237 eV band-gap MAE across 1,048 materials — without training on a single one.

The question this benchmark answers

The space of plausible inorganic compositions is on the order of 10⁸–10¹². DFT cannot screen that — hours to days per material. ML can run at interactive speed, but only after a training set itself generated by DFT, and only inside the chemistries it has already seen. A predictor that delivers DFT-competitive accuracy from composition alone, at millisecond speed, would let researchers scan the whole space before any heavyweight calculation runs.

We measured that predictor against an experimentally grounded 1,048-material cohort to find out.

The cohort

1,048 inorganic compounds with experimental band gaps, sourced from the Materials Project database. The cohort spans chalcogenides, oxides, halides, pnictides, intermetallics, hydrides, and carbides — with experimental gaps from 0 eV (metals) up to ~12 eV (wide-gap fluorides). Both metals and semiconductors are included so the predictor is tested on the full classification problem, not only on the wide-gap subset where most physics models look easy.

Full benchmark data — per-compound experimental, predicted, and absolute-error values for all 1,048 entries — are downloadable as JSON and CSV from the materials benchmark page.

Where 0.237 eV sits among other methods

Reading a band-gap MAE number in isolation is misleading. The right question is: at this accuracy level, what computational resources and training data did it cost? Every other method that reaches the ~0.2–0.3 eV accuracy band pays for it in one of two currencies — a large labelled training set, or hours-to-days of supercomputer time per material. This benchmark pays neither.

Landscape of band-gap predictors

The table consolidates published numbers across four method classes. Sources are linked inline; see also the consolidated reference list at the end of the case study.

Reported MAE values come from the methods' own publications or the MatBench leaderboard for the matbench_expt_gap task (Zhuo et al. 2018 cohort, 4,604 experimental gaps from composition). DFT timings are screening-mode estimates on a single CPU; ML timings are inference-only (training cost amortised separately).

How 0.237 eV maps to real screening use cases

A 0.237 eV mean absolute error is not the same thing across applications. Below is what this accuracy band buys you for the most common materials-design tasks.

For each use case the table reports the required accuracy band as a practical engineering tolerance, not a mathematical proof. The honest framing: this benchmark is a first-pass filter that clears ~99% of bad candidates in milliseconds, before any DFT, hybrid DFT, or wet-lab synthesis cost is incurred. For the tightest applications — thermoelectrics, narrow-gap photovoltaics, Mott physics — treat the output as a candidate list, then validate each finalist with a more expensive method.

What this benchmark does NOT claim

• Not better than HSE06 or GW on individual compounds where those methods have been carefully run.
• Not a replacement for an experimentalist measuring a real sample under controlled conditions.
• Not best-in-class on strongly correlated Mott insulators — that regime needs Hubbard-U physics still on the roadmap.
• Not magical on rare-earth f-electron compounds, where the orbital-occupation problem remains hard.
• Not a 100% accuracy claim — per-compound errors of several hundred meV occur, especially at the boundary between metallic and narrow-gap semi behaviour. The benchmark page lists the highest-absolute-error compounds explicitly.

Beyond the benchmark: continuous alloys and doping

The 0.237 eV MAE number is what the headline benchmark measures, but the real value of the predictor for industrial users is the continuous piece. Every working semiconductor device ships as an alloy or a doped composition, not a stoichiometric endpoint. The same single predictor handles both directly, with the same zero-training, zero-fitted-parameter contract.

Alloy band-gap engineering

Type a fractional composition and the predictor decomposes it into integer-stoichiometry endpoints, evaluates each via the same composition-only path that gives 0.237 eV on the cohort above, and interpolates with Vegard's law. No system-specific bowing constants; no fit. Sample verified outputs:

Doping → carrier concentration and Fermi level

A dilute dopant in a fractional formula is auto-classified as donor / acceptor / isoelectronic by comparing the dopant's bonding signature to the host site it replaces. The module returns activation energy, carrier concentration n or p at the user's temperature, Fermi-level offset from the band edge, and Burstein–Moss channel-filling shift — all in the same millisecond call:

Carrier concentration: 99.0% within factor 3 on 197 Hall measurements

Extending past type classification, the predictor scores against a curated 197-entry Hall-measurement cohort spanning Si, Ge, GaAs, InP, InAs, InSb, GaSb, GaN, AlAs, AlSb, GaP, ZnO, ZnS, ZnSe, ZnTe, CdS, CdSe, CdTe, PbS, PbSe, PbTe, SiC, diamond, ITO, AZO, FTO and their compensated combinations. Curated from Sze, Madelung, Adachi, Pearton, Look, NSM Ioffe.

Mobility & conductivity

The same single call returns electron and hole mobilities, and the full conductivity σ = q n μ, at any temperature. 100% of predictions land within a factor of 3 of literature mobilities across Si, Ge, GaAs, InP, InAs, GaN, CdTe, ZnO, and other canonical semiconductors. Pair with carrier-concentration output to size a transistor channel, rank transparent conductors, or triage thermoelectric leads.

Band-edge alignment for heterojunctions

Pass any two compositions and the predictor returns the conduction-band offset ΔE_c, valence-band offset ΔE_v, and the Type I / II / III alignment classification — the key inputs for HEMT, LED quantum well, multi-junction solar, and IR-detector stack design. ~88% type classification accuracy on canonical heterostructure benchmarks (AlGaAs / GaAs, AlGaN / GaN, CdSe / ZnS, InAs / GaSb broken-gap, SiGe, CdTe / ZnTe, etc.) with eV magnitudes inside typical DFT band-offset uncertainty (~0.3 eV).

Ab initio on materials that don't exist yet

A separate "novel-material" benchmark mode disables every measured-property lookup the module would normally use, and forces the predictor to derive the entire bundle — band gap, carrier concentration, mobility, conductivity, band-edge alignment — from composition alone. On the same 197-entry Hall cohort and the same heterostructure benchmark set, the novel-material mode reproduces the known-material accuracy to within a few percentage points on every metric:

This is the contract that separates this pipeline from interpolation-based surrogates: the prediction for a never-measured composition is generated by the same physics that scores against measured materials, and lands at the same accuracy. The known / novel split is exposed as a benchmark mode so users can verify it on any composition.

Reproduce and read the data

See also the DFT cross-check case study for a head-to-head comparison on 15 canonical materials with a local PBE-DFT setup, and the semiconductor mobility atlas for a full screening-application walkthrough.

References

Comparison numbers in the landscape table above are sourced from the methods' own publications and the publicly maintained MatBench leaderboards. The full list:

1. MatBench leaderboard (composition-only experimental band gap) — matbench_v0.1 matbench_expt_gap. The canonical leaderboard for composition-only experimental band-gap prediction; 5-fold CV on 4,604 materials from the Zhuo et al. (2018) cohort.
2. Zhuo, Y., Mansouri Tehrani, A., Brgoch, J. “Predicting the Band Gaps of Inorganic Solids by Machine Learning.” J. Phys. Chem. Lett. 9(7), 1668–1673 (2018). DOI: 10.1021/acs.jpclett.8b00124. The source dataset used by MatBench expt_gap.
3. arXiv:2501.02932 (Jan 2025) — “Predicting band gap from chemical composition: A simple learned model for a material property with atypical statistics.” arXiv:2501.02932. The closest direct comparator: composition-only, no structure, one fitted parameter per element, MAE = 0.575 eV on the Zhuo 4,603-material cohort.
4. DARWIN 1.5 (2024) — “Large Language Models as Materials Science Adapted Learners.” arXiv:2412.11970. Current #1 on MatBench expt_gap (0.287 eV); LLaMA-2-based, fine-tuned on 6M materials-science papers and 21 experimental datasets covering 49,256 materials.
5. Wang, A.Y.-T. et al. “CrabNet: Compositionally restricted attention-based network for materials property predictions.” npj Comput. Mater. 7, 77 (2021). npj Comput. Mater. article.
6. De Breuck, P.-P., Hautier, G., Rignanese, G.-M. “Materials property prediction for limited datasets enabled by feature selection and joint learning with MODNet.” npj Comput. Mater. 7, 83 (2021).
7. Choudhary, K., DeCost, B. “Atomistic Line Graph Neural Network for improved materials property predictions.” npj Comput. Mater. 7, 185 (2021). arXiv:2106.01829. ALIGNN (uses crystal structure).
8. Chen, C., Ye, W., Zuo, Y., Zheng, C., Ong, S.P. “Graph Networks as a Universal Machine Learning Framework for Molecules and Crystals.” Chem. Mater. 31(9), 3564–3572 (2019). MEGNet.
9. Xie, T., Grossman, J.C. “Crystal Graph Convolutional Neural Networks for an Accurate and Interpretable Prediction of Material Properties.” Phys. Rev. Lett. 120, 145301 (2018). CGCNN.
10. EOSnet (2025) — Embedded overlap structures for materials GNN; 0.163 eV MAE on MP DFT-PBE band gaps. PMC article.
11. Phillips, J.C. “Ionicity of the Chemical Bond in Crystals.” Rev. Mod. Phys. 42(3), 317–356 (1970). Phillips–Van Vechten ionicity framework, the canonical first-principles composition-only band-gap model for binary AB compounds.
12. Zaanen, J., Sawatzky, G.A., Allen, J.W. “Band gaps and electronic structure of transition-metal compounds.” Phys. Rev. Lett. 55(4), 418–421 (1985). The Mott–Hubbard / charge-transfer classification used in our pipeline mechanism tagging.
13. Materials Project — materialsproject.org. Source of all DFT-PBE / DFT-HSE band-gap reference values used in the cross-check case study.

The 1,048-material cohort used for this benchmark is sourced from Materials Project and includes 461 metallic (E_g = 0) and 587 non-metallic systems. The full per-material predictions and errors are available via the JSON / CSV download links above.