← Benchmarks | Electronic Properties

Metal / Insulator Classification BENCHMARK

Is a compound a metal or an insulator? FluxMateria answers from composition alone — 0.879 ROC-AUC across 4,921 experimentally-characterised compounds (matbench expt_is_metal, 5-fold cross-validation), with zero fitted parameters and zero training data.

0.879
ROC-AUC
5-fold cross-validated
4,921
Compounds Scored
from composition only
0
Fitted Parameters
no training data used
100%
Deterministic
same input → same output

Metal recall (sensitivity) 0.858 · insulator recall (specificity) 0.899 · overall accuracy 0.879. A dummy classifier scores 0.470 on this task; FluxMateria is well above the standard featurized random-forest baseline while using no labels at all.

Comparison with DFT and ML

Metal/insulator classification trade-offs: accuracy, speed, and data dependence

Metric FluxMateria DFT band structure Trained ML (featurized / GNN)
Metal/insulator ROC-AUC 0.879 (4,921 compounds) Reference-grade, per compound Featurized baselines ~0.82; best trained models into the low 0.9s
Speed per query Milliseconds Hours per compound ~1 second
Input required Composition only Full crystal structure Composition (and often structure)
Training data None None Thousands of labelled examples
Fitted parameters 0 fitted XC functional + U parameter Thousands to millions
Reproducibility Fully deterministic Deterministic (given settings) Seed- and split-dependent

Key takeaway: FluxMateria classifies metal vs. insulator at 0.879 ROC-AUC from composition alone — above standard featurized machine-learning baselines — while using zero training data and zero fitted parameters. Trained deep-learning models can reach somewhat higher, but only by fitting the full labelled dataset; FluxMateria predicts purely from first-principles physics, is fully deterministic, and runs in milliseconds with no crystal structure required.

Representative Predictions

Correct calls spanning intermetallics, oxides, semiconductors, and correlated systems

Compound Class of material Experiment FluxMateria Status
Al2CuIntermetallicMetalMetalPASS
AlNi2IntermetallicMetalMetalPASS
B4WTransition-metal borideMetalMetalPASS
Ag2BiO3Mixed-valence oxideMetalMetalPASS
Al2FeNiIntermetallicMetalMetalPASS
Ba2Ce2O5Rare-earth oxideMetalMetalPASS
NaClAlkali halideInsulatorInsulatorPASS
MgOIonic oxideInsulatorInsulatorPASS
GaAsIII–V semiconductorInsulatorInsulatorPASS
TiO2Closed-shell oxideInsulatorInsulatorPASS
NiOCorrelated (Mott) oxideInsulatorInsulatorPASS
Cr2O3Correlated oxideInsulatorInsulatorPASS

12 representative compounds from the 4,921-compound benchmark. Experimental classes from the matbench experimental dataset.

Physics Regimes Covered

Metal-vs-insulator is not one mechanism — FluxMateria resolves the distinct physics that decide it

Band metals & intermetallics

Alloys and metal-rich phases with delocalised conduction

Mott & charge-transfer insulators

Correlated oxides where partial filling still gaps (NiO, Cr2O3)

Mixed-valence conductors

Inter-valence conduction in non-integer-valence oxides

Closed-shell band insulators

Empty- or full-shell oxides and ionic salts (TiO2, MgO)

Rare-earth & heavy-fermion

f-electron oxides with valence-fluctuation conduction

Covalent & ionic semiconductors

III–V, II–VI, halides and main-group oxides

Methodology

How FluxMateria decides metal vs. insulator

Metallicity is determined by the production FluxMateria materials engine directly from composition — no crystal structure input, no training data, and no parameters fitted to the benchmark labels. The engine derives each compound's electronic character from first-principles physics and textbook atomic data alone, then classifies it as a metal or an insulator.

  • 4,921 compounds from the matbench experimental is-metal task
  • 5-fold cross-validation using the benchmark's canonical splits; the metric is ROC-AUC
  • Zero training: no benchmark labels are seen during prediction; training_data_used = none, fitted_parameters = 0
  • Deterministic: identical results across runs and machines — no random seeds anywhere in the pipeline
  • Composition only: no crystal structure required; millisecond runtime enables full-library screening

Scope & Limitations

Strengths

  • 0.879 ROC-AUC on 4,921 compounds — best-in-class among zero-parameter physics methods
  • Above standard featurized machine-learning baselines, with no training data
  • Composition-only input: no crystal structure required
  • Fully deterministic and reproducible — no random seeds
  • Millisecond runtime enables batch screening of large compositional spaces
  • Spans intermetallics, oxides, chalcogenides, nitrides, halides, and rare-earth systems

Known Limitations

  • Some metal/insulator distinctions hinge on crystal structure and fine band details that composition alone cannot fully resolve
  • Heavy 4d/5d narrow-gap semiconductors and certain strongly-correlated phases remain the hardest cases
  • Trained machine-learning models with the full labelled dataset can reach somewhat higher accuracy
  • Benchmark uses experimental room-temperature classifications; phase- and temperature-dependent materials are reported as measured

References

Benchmark and experimental data sources

  1. A. Dunn, Q. Wang, A. Ganose, D. Dopp, A. Jain, “Benchmarking materials property prediction methods: the Matbench test set and Automatminer reference algorithm,” npj Computational Materials 6, 138 (2020).
  2. Y. Zhuo, A. Mansouri Tehrani, J. Brgoch, “Predicting the band gaps of inorganic solids by machine learning,” J. Phys. Chem. Lett. 9, 1668 (2018) — experimental band-gap / metallicity source.
  3. R. Kingsbury et al., experimental band-gap compilation (independent cross-check of metallicity labels).
  4. Matbench leaderboard, matbench.materialsproject.org (accessed 2026).
  5. Materials Project Database, materialsproject.org.

Benchmark basis

Metallicity is computed from first-principles FluxMateria physics with zero parameters fitted to the benchmark. Experimental classifications are used only to score accuracy — never as model input.

Flux Physics

Try the Materials module

Classify any composition as metal or insulator — alongside band gap, elastic, thermal, and magnetic properties — all from composition, all in milliseconds.

← Back to Module Request Access