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.