📐 FLUXMATERIA — MATERIALS

State of the art for composition-input
electronic-property prediction.

Band gap, carrier concentration, mobility, conductivity, and band-edge alignment — from a formula and a temperature, in a single ~1 ms call. 0.237 eV band-gap MAE on the 1,048-material public cohort beats every other published composition-input method we've benchmarked against and matches the accuracy band of modern crystal-structure-input graph neural networks — without a training set. 99.0% of Hall carriers within a factor of 3 with 100% type accuracy, 100% of mobilities within a factor of 3. No crystal structure required, novel-material accuracy that matches the known-material number on every metric.

SOTA · composition-input methods End-to-end bundle: E_g / n / μ / σ / ΔE Composition + T input only ~1 ms per query Zero training data · zero fitted parameters Novel-material = known-material accuracy

Open band-gap benchmark Read case study Request Pilot Access

SOTA

Among composition-input methods on the 1,048-material public cohort

0.237 eV

Band-gap MAE · beats published GNN ML on the same cohort (~0.31–0.33 eV)

99.0%

Hall carriers within factor 3 (197-entry cohort, novel-material mode)

100%

Mobility within factor 3 · 100% conductivity-type accuracy

~1 ms

Per-material wall time, full bundle, single CPU thread

The breakthrough

An entire electronic-property bundle, in the time other tools take to open a crystal file.

For band gap alone, modern predictors fall into two camps. DFT-PBE delivers ~0.5–1.0 eV MAE and needs crystal structures, pseudopotentials, and a queue. Graph-neural-network ML reports ~0.31–0.33 eV MAE on the same 1,048-material cohort — and needs training data that does not exist for new chemistries. This module delivers 0.237 eV MAE at ~1 ms per material, from composition alone, with no training set, no fitted constants, and no crystal-structure prerequisite. Every prediction is a closed-form physics evaluation, not an interpolation between training examples.

The unique part isn't the band gap. It's the bundle. The same single call returns dopant classification, activation energy, carrier concentration n or p at any temperature, Fermi-level offset, Burstein–Moss shift, electron and hole mobility, total conductivity, and heterojunction band-edge alignment — from a composition string and a temperature. No published predictor produces Hall carrier concentration from composition + temperature alone at this accuracy. Every existing alternative needs either the measured activation energy as an input (which is what you were trying to predict), a DFT defect-formation calculation per dopant (hours to days each), or a per-family empirical fit (Si only, or GaN only, or ITO only). One ~1 ms call replaces all three.

To replicate this with existing tools

Band gap: HSE06 DFT — hours per material, requires crystal structure
Doping → carrier concentration: DFT defect-formation run per dopant — hours to days each
Mobility & conductivity: per-family empirical fit or full Boltzmann-transport DFT
Band-edge alignment: DFT slab calculation per interface — days, requires the supercell

FluxChem returns all four from a formula + temperature in roughly a millisecond.

And it works on novel candidates as well as it does on tabulated ones. A separate ab initio benchmark mode disables every measured-property lookup and forces the predictor to derive the entire bundle from composition alone. On the same 197-entry Hall cohort, novel-material mode reproduces the known-material accuracy on every metric — the rare case where the ab initio mode actually beats the table-fed mode (99.0% vs 98.5% within factor 3) because the underlying physics carries information the tables don't. This is the property that ML predictors fundamentally lack — they degrade out of distribution. A closed-form physics predictor doesn't have a distribution to be out of.

What the Band-Gap module gives you

One composition in. One closed-form prediction out. Documented physics every step.

📐

Composition-only prediction

Type a formula — GaAs, Cs2Al2B2O7, Bi2Te3 — and get a band-gap estimate in roughly a millisecond. No crystal structure required, no pseudopotentials, no SCF settings.

🧪

Continuous alloy composition

Accepts fractional formulas directly: In0.53Ga0.47As → 0.73 eV (telecom-grade 1.55 µm), Hg0.4Cd0.6Te → LWIR cutoff, mixed-halide perovskites, CIGS, InGaAsP quaternaries. Each endpoint comes from the same zero-training predictor; Vegard's law handles the interpolation. No system-specific bowing constants; no fit.

⚛️

Doping → carrier concentration No peer

Dilute dopants are auto-classified as donor / acceptor / isoelectronic. The module returns activation energy, carrier concentration n or p at any temperature, Fermi-level offset, Burstein–Moss shift, and conductivity type — from formula + temperature alone. Works for Si:P, GaN:Mg, ITO, AZO, p-CdTe:Cu and the rest of the canonical industrial dopant set. 99.0% of predictions within a factor of 3 on a 197-entry Hall-measurement cohort, with 100% type-classification accuracy. No published predictor returns Hall carrier concentration from composition + temperature alone at this accuracy — every alternative needs either the measured E_a as input, DFT defect-formation runs, or a per-family empirical fit.

🚂

Carrier mobility & conductivity

The same single call returns electron and hole mobilities and the total electrical conductivity σ = q n μ at any temperature. 100% of predictions within a factor of 3 against literature mobilities across Si, Ge, III-V, II-VI, IV-VI, and wide-gap families. Combine with carrier concentration to size a transistor, screen a transparent conductor, or rank thermoelectric leads.

🪜

Band-edge alignment for heterojunctions

Pass any two compositions and get the conduction-band offset ΔE_c, valence-band offset ΔE_v, and 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 heterostructures, with eV magnitudes inside the typical DFT band-offset uncertainty.

🆕

Novel candidates work too

The pipeline does not need to have seen the material before. A separate ab initio benchmark mode disables every measured-property lookup; the predictor reproduces its known-material accuracy to within a few percentage points on every metric. Carrier concentration, mobility, and band-edge alignment for materials that have never been synthesised, derived from composition only.

⚡

Sub-millisecond inverse search

Scan thousands of candidates per second against a target gap window. Wide-gap insulator? UV-active semiconductor? Visible-light photocatalyst? Set the spec; rank the library.

🧭

Metal / semi / insulator on one predictor

461 metals + 587 semis / insulators evaluated on a single fixed predictor. The same physics that delivers 0.320 eV MAE on semiconductors correctly identifies 461 metallic compositions as E_g = 0 with 0.130 eV MAE — no separate metal/semi classifier.

📊

Per-family transparency

Chalcogenide, oxide, halide, pnictide, intermetallic, binary oxide — the benchmark page reports family-stratified MAE so you can see exactly which chemistries the predictor owns.

📜

Deterministic and auditable

Same composition, same prediction, every time. Every result carries the mechanism class and family classification it was scored under, so any decision in your screening pipeline can be reproduced and explained months later.

⚙️

API + workspace export

REST endpoint, JSON/CSV batch export, Workspace run capture with deterministic versioning. Drop the predictor into your screening pipeline and reproduce any decision later.

Novel-material accuracy = known-material accuracy

The honest test of whether a predictor performs ab initio: does it still work when you take its lookup tables away?

The dual-mode benchmark runs the same 197 Hall measurements, 14 mobility cases, and 8 heterojunction band-offset cases under two configurations. Known-material mode lets the pipeline use every tabulated property it would normally have access to. Novel-material mode disables every measured-property lookup and forces the predictor to derive each output from composition alone — the configuration that matters for materials that don't exist yet.

Metric	Known-material mode	Novel-material mode
Band gap MAE (1,048-material public cohort)	0.237 eV	0.237 eV
Hall carrier within factor 3 (n = 197)	98.5%	99.0%
Hall carrier mean log₁₀ error	0.065 dex	0.053 dex
Conductivity-type classification (n / p / intrinsic)	100%	100%
Mobility within factor 3 (n = 14)	100%	100%
Band-offset Type I/II/III classification (n = 8)	87.5%	87.5%

Note the Hall-carrier row: novel-material mode beats known-material mode (99.0% vs 98.5%). This is not a typo. The closed-form physics carries information that the property tables do not — specifically the multi-valley density-of-states enhancement for elemental indirect-gap hosts — so disabling the lookups improves the result on those cases. ML predictors cannot make this kind of claim; they degrade out of distribution because they have no physics to fall back on. A closed-form physics predictor doesn't have a distribution to be out of.

How a band gap is computed

From composition to result in five deterministic steps. No fitting, no training, no DFT.

Enter the composition

Type a chemical formula. No crystal structure, no pseudopotentials, no SCF settings — just the composition. Works from the browser, the API, or a Workspace batch.

Family is identified

The engine places the composition in its physics family (oxide, chalcogenide, halide, pnictide, intermetallic, etc.). Family-stratified accuracy is reported on the public benchmark page.

Gap is evaluated

First-principles physics returns a band-gap value in eV in roughly a millisecond. Metals (E_g = 0), semiconductors, and wide-gap insulators are all handled on the same predictor.

Audit signal is attached

The mechanism class and family classification are returned alongside the value. The output is fully deterministic: the same composition always yields the same prediction.

Use or hand off

Filter against a target gap window, batch-screen a library via the Workspace, or pipe to the Semiconductor Design, Surface & Contact, or Battery modules for the rest of the property bundle.

Why you can trust it

Calibrated against 1,048 experimental band gaps. Reported headline metric, not a cherry-picked subset.

0.237 eV

Overall MAE on the 1,048-material public band-gap benchmark. Single fixed pipeline; every composition predicted under the same rules.

0.320 eV

Semiconductor / insulator MAE on the 587 non-zero-gap subset. Inside the band where modern GNN ML reports its own headline numbers — without their training set.

0.130 eV

Metallic-system MAE (exp = 0) on the 461 metals. The same predictor that scores semiconductors correctly identifies these as E_g = 0 — no separate metal/semi classifier.

Fitted parameters anywhere in the band-gap path. First-principles physics from end to end.

Training examples. The 1,048-material cohort is the evaluation set, never the training set. The same predictor is used on every new composition.

~1 ms

Per-material wall time on a single CPU thread. Hundreds of thousands of compositions per minute on a laptop.

Read the proof: the full per-material results, family-stratified MAE, head-to-head versus DFT and ML, and a downloadable JSON / CSV are on the band-gap benchmark page. The narrative case study walks through what 0.237 eV at millisecond speed means for screening loops in the band-gap case study.

How FluxMateria compares

0.237 eV MAE at ~1 ms, from composition alone, with no training data — head-to-head with the alternatives.

Property	FluxMateria	DFT (PBE screening)	GNN / ML surrogates	Experiment
Headline MAE on 1,048-mat set	0.237 eV	~0.5–1.0 eV	~0.31–0.33 eV	Reference (0)
Time per material	~1 ms	Minutes to hours	Seconds (after training)	Days to weeks per measurement
Input required	Composition only	Composition + crystal structure + pseudopotentials + SCF settings	Composition + crystal structure (most models)	Synthesized sample
Training data required	None	None	10³–10⁵ labelled examples	None
Fitted parameters	0	Functional-level (PBE)	10⁵–10⁷ weights	None
Out-of-distribution behaviour	Same physics, same accuracy	Same accuracy, longer queue	Silently degrades; no signal	N/A
Reproducibility	Deterministic	Deterministic given settings	Depends on training seed	Lab-to-lab variance
Auditable prediction path	Yes — mechanism class returned	Yes, via SCF logs	No (black-box weights)	Yes, via lab notebook

The honest one-line summary: on the same fixed 1,048-material cohort, this module reaches the accuracy band of modern graph-neural-network ML without ever training on a single example, and runs roughly four orders of magnitude faster than DFT-PBE screening. The headline number, all 1,048 per-material predictions, and the bench JSON are on the benchmark page.

Where the Band-Gap module wins

Screening loops where the question is the gap, not a single material.

Use case 1

Semiconductor library triage

Score a candidate library against a target gap window before committing to DFT or synthesis. Cut a 10,000-material list to a 100-material shortlist in seconds, then send the survivors to higher-fidelity tools.

Use case 2

Photovoltaic absorber search — CIGS, perovskite tandems

Sweep Cu(In(1−x)Ga(x))Se₂ across x to land on the Shockley–Queisser sweet spot (1.1–1.7 eV). Tune mixed-halide perovskite tandems CsPb(I,Br)₃ for the 1.6–2.0 eV top-cell. Composition steps under a millisecond each.

Use case 3

LED band-gap engineering — InGaN, AlGaInP

Continuously vary In(x)Ga(1−x)N or (Al(y)Ga(1−y))(0.5)In(0.5)P to hit the target emission wavelength. p-GaN:Mg and n-GaN:Si layers come with carrier-concentration and Fermi-level outputs for diode design in the same call.

Use case 4

Telecom photonics — InGaAs / InGaAsP at 1.3 / 1.55 µm

In(0.53)Ga(0.47)As lattice-matched to InP returns 0.73 eV (exp 0.74) directly from the fractional formula. The full quaternary In(x)Ga(1−x)As(y)P(1−y) is supported, with bilinear interpolation between four endpoints.

Use case 5

IR detectors — HgCdTe cutoff wavelength

Hg(1−x)Cd(x)Te cutoff tunes continuously from LWIR (high Hg) to MWIR to SWIR. Solve inverse: find x that gives a target gap for a specific detector band.

Use case 6

Transparent conductors — ITO, AZO, ATO

Sn-doped In₂O₃ (ITO), Al-doped ZnO (AZO), Sb-doped SnO₂ (ATO): the doping framework returns the n-type carrier concentration and the Burstein–Moss-shifted optical gap directly from the doped formula.

Use case 7

Thermoelectric alloy optimisation

Bi_2−xSb_xTe₃, Pb(Te,Se,S), half-Heusler alloys — sweep composition + dopant fraction in one call to find the gap and carrier concentration that maximise the thermoelectric figure of merit.

Use case 8

Power electronics — SiC, GaN, Ga₂O₃ channels

n-GaN:Si, p-GaN:Mg, doped β-Ga₂O₃, AlGaN HEMT channels — composition + dopant fraction + temperature give the band gap, carrier concentration, and conductivity type in the same call. Pair with the Semiconductor Design module for mobility / breakdown FoM.

Use case 9

Photocatalysts and water-splitting

Target the 1.6–2.4 eV visible-light window with composition-only screening across transition-metal oxide / sulfide / oxynitride libraries. Doped variants (NaTaO₃:La, BiVO₄:Mo) get carrier-type and concentration in the same pass.

Use case 10

Battery cathode interface coatings

Wide-gap insulator candidates for electrolyte interphase coatings: score 1,000+ oxide / halide / oxyhalide compositions in seconds, hand off the survivors to the battery-electrochemistry module.

Use case 11

Intermetallic / metallic-vs-semi triage

Tag metallic compositions (E_g = 0) versus narrow-gap semis automatically. Useful for catalyst, magnetic, and thermoelectric-adjacent searches where you need to know which compositions are conducting before anything else.

Use case 12

Heterojunction stack design

Pair any two compositions and get the conduction-band offset, valence-band offset, and Type I / II / III alignment in one call. Spec a HEMT channel, an LED quantum well, a multi-junction solar tandem, or a tunnel junction without first running DFT on each pair.

Use case 13

Mobility & conductivity screening

Electron and hole mobilities, total electrical conductivity σ = q n μ, at any temperature, in the same call as the band gap. Rank transparent-conductor candidates, transistor channels, or thermoelectric leads on the actual transport metric they will be evaluated on.

Use case 14

Novel materials — ab initio screening

The pipeline produces the full electronic-property bundle for materials that have never been measured. Switch on the ab initio mode and every prediction is derived from composition alone, with no measured-property fallback. Use this to rank materials that don't yet exist.

Questions FluxMateria can answer about a material

From a formula and a temperature — full breadth of the platform, no need to know which sub-module returns which value.

Question	What you get back	Latency
What's the band gap of `Cs₂Al₂B₂O₇`?	E_g in eV, metal / semi / insulator classification, mechanism family	~1 ms
What's the band gap of the alloy `In_0.53Ga_0.47As`?	Continuous-composition E_g (Vegard interpolation on the same predictor), no per-system bowing constant	~1 ms
*What's the optical* band gap (with exciton binding subtracted) of `GaAs`?**	Single-particle E_g, Wannier-Mott exciton binding E_b, optical-edge gap E_g^opt = E_g − E_b, and a regime flag (Wannier-Mott / Frenkel / unavailable)	~1 ms
What's the band gap of `GaAs` at 77 K? At 600 K?	Temperature-dependent E_g(T) over 0–600 K, returned alongside the high-T slope and the phonon-onset temperature for that composition. Median 24 meV accuracy on the v1 reference cohort — within the published-fit scatter itself.	~1 ms
Is `NiO` a Mott insulator or a band semiconductor? What about `TiO₂`?	Electronic-structure family classification: metal / Mott insulator / d⁰ band insulator / d¹⁰ band semiconductor / band semiconductor / band insulator / ionic insulator. Returned alongside the underlying mechanism and (cation, anion) taxonomy. 96% accuracy on a 1,629-material cohort — 100% metal-class precision, 100% non-metal recall.	~1 ms
Is `Fe` in `Si` a shallow donor or a deep transition-metal trap?	Dopant depth classification: shallow / deep-TM / deep covalent acceptor. Returned alongside the activation energy and which deep-level mechanism fires. 100% accuracy on a 99-material reference (Madelung Tables 5.3–5.8 + Pearton 2012 + NSM Ioffe Hall benchmark).	~1 ms
What's the carrier concentration of `Si:P` at 1×10¹⁷ dopants/cm³, 300 K?	Activation energy E_a, ionization fraction, n or p, conductivity type	~1 ms
What's the Fermi-level offset and Burstein–Moss shift for heavily-doped ITO?	E_F offset from band edge, B-M channel-filling shift, degenerate flag	~1 ms
What's the electron mobility of `GaAs` at 1×10¹⁷ donors/cm³?	μ_e, μ_h, ionized-impurity vs phonon breakdown, total conductivity σ	~1 ms
Is my compensated `Si:P+B` still n-type? By how much?	Compensation ratio, N_D/N_A ionized populations, net dopant, majority + minority carriers	~1 ms
Is my requested dopant level above the solubility limit?	Per-dopant solubility ceiling N_sol, headroom, above-ceiling warning	~1 ms
What's the conduction-band offset at the `AlGaN/GaN` interface?	ΔE_c, ΔE_v, Type I / II / III alignment classification	~1 ms
What's the effective mass and dielectric constant?	m^*/m_e, ε_r — derived alongside the bundle, no separate query	~1 ms
Rank 4.66 million semiconductor candidates by Baliga FoM.	Atlas explorer: composition + family + tier + FoM rank. Top-10 per FoM exported as a gold-candidate shortlist.	~225 s
*Does this work for a novel* material the predictor has never seen?**	Yes — novel-material mode disables every measured-property lookup; the bundle is derived from composition alone. Hall accuracy unchanged.	~1 ms

Every row above is one API call. The platform decides which physics path to take based on your input; you don't pick a module.

Where FluxMateria stops — and what to reach for instead

Honest pairing guide. For each question the platform doesn't answer, here's the canonical external tool to use.

Question outside FluxMateria's scope	Why	Canonical tool to use instead
*What's the gap of the rutile* polymorph specifically (vs anatase)?**	Composition input only; most polymorphs collapse to family value (a few SiC polytype prefixes are tagged)	DFT with the polymorph CIF (VASP, Quantum ESPRESSO, ABINIT, FHI-aims). Or look up the polymorph-specific entry on Materials Project or AFLOW
*What's the optical* gap with exciton binding included?**	✓ Now supported in FluxMateria. Wannier-Mott exciton physics with closed-form binding, ~6 meV mean error vs experiment on the v1 reference cohort. Frenkel-regime guard for ionic / molecular crystals.	For higher accuracy on strongly-anisotropic wurtzite or polaronic systems, still use GW + BSE (BerkeleyGW, YAMBO). For molecular Frenkel excitons, TD-DFT (Gaussian, NWChem).
What's the deep-trap defect level (e.g. EL2 in `GaAs`, DX center in `AlGaAs`, V_O in `TiO₂`)?	Shallow-dopant activation energies are handled, mid-gap defect-level diagrams are not	DFT defect-formation calculations with charge-state corrections (VASP defect workflows, NEMO for nitrides). Materials Project defect database for ~5K materials
What's the surface-state-modified gap on a reconstructed surface or 2D system?	Bulk gap only; surface reconstructions and edge states are not modelled	Slab DFT with explicit vacuum layers (VASP, QE). STM/STS simulation codes for surface DOS
What's the gap at a grain boundary or interface defect?	Bulk only; polycrystalline morphology not modelled	DFT supercells with explicit GB structures. ReaxFF molecular dynamics for grain-boundary morphology at scale
How does the gap shift with temperature?	✓ Now supported in FluxMateria. E_g(T) over 0–600 K with the Varshni form, median 24 meV accuracy on the v1 reference. Returns the temperature-dependent gap, the high-T slope, and the phonon-onset temperature per material.	For finer-grained T-dependence on strongly-anharmonic or quantum-paraelectric materials, still use DFPT (Quantum ESPRESSO ph.x, VASP DFPT).
How does the gap shift with hydrostatic pressure?	Room-pressure bulk values only; no explicit dE_g/dP derivative	High-pressure DFT (VASP under stress). Pressure derivative can be reconstructed externally from FluxMateria's deformation-potential output and the bulk modulus.
I need the full band structure / DOS / k-resolved effective mass tensor.	Returns scalar gap; scalar m^*/m_e; not full k-space	DFT band structure (any code) + BoltzTraP / AMSET / EMC++ for transport tensors
I need spin-orbit-coupling fine structure (Γ–L–X splittings, k·p Kane parameters).	Single scalar gap; SOC fine structure not resolved	DFT + SOC (VASP, QE with relativistic pseudopotentials). NEMO tight-binding for III-V quantum-confinement Kane parameters
I need < 0.15 eV MAE on a complex-oxide tail material (cuprate, manganite, Mott-Hubbard correlated oxide).	Family MAE is 0.29 eV for complex oxides; strongly-correlated tail cases carry the residual	Hybrid DFT (HSE06 / PBE0) for ~0.2–0.3 eV MAE at 10⁶× the compute cost. GW for ~0.15 eV MAE at 10⁸× the compute cost. DFT + DMFT for correlated-electron physics

The pairing logic: use FluxMateria for the screening question — "is this material in the right ballpark? which thousand candidates from my four-million-candidate space deserve a closer look?" Use the canonical tool above for the fine-grained question — "exactly what is the gap of this specific polymorph at 77 K under 2 GPa with this defect concentration?". FluxMateria is faster than any of those by 10⁴–10⁸×; those tools are tighter than FluxMateria on the specific question they exist to answer. Different jobs.

Scope & Limitations

Strengths

0.237 eV MAE on the 1,048-material public set — matches modern GNN ML on the same cohort without a training set.
~1 ms per query — roughly four orders of magnitude faster than DFT-PBE screening.
Composition input only. No crystal structure, pseudopotentials, or SCF settings required.
Zero fitted parameters and zero training data in the band-gap path. First-principles physics from end to end.
Same predictor handles metals (E_g = 0), semiconductors, and wide-gap insulators — no separate metal/semi classifier.
Family-stratified accuracy reported on the public benchmark page — chalcogenide, oxide, halide, pnictide, intermetallic.
Deterministic and auditable: every prediction returns the mechanism class and family it was scored under.
Electronic-structure family classification (correlation_class field): every prediction is tagged metal / Mott insulator / d⁰ band insulator / d¹⁰ band semiconductor / band semiconductor / band insulator / ionic insulator, with the underlying mechanism returned alongside. 96% accuracy on a 1,629-material reference, 100% metal-class precision, 100% non-metal recall.
Dopant depth classification (dopant_class field) on substitutional-doping queries: shallow / deep-TM / deep covalent acceptor, with the deep-level mechanism reason. 100% accuracy on a 99-material reference cohort.
Temperature-dependent E_g(T) over 0–600 K via the Varshni form: median 24 meV accuracy on the v1 reference cohort (15 canonical semiconductors × 9 temperatures = 135 sample points) — at or below the Varshni-fit scatter itself.

Known limitations

Composition input only — most polymorphs and crystal-structure-specific gaps (e.g. rutile vs anatase TiO₂) collapse to the dominant family prediction. A handful of well-tagged polytype prefixes (4H-SiC, 6H-SiC, 3C-SiC) are distinguished.
Optical band gap (Wannier-Mott exciton binding subtracted) is reported alongside the single-particle fundamental gap. ~6 meV mean error against a small experimental reference cohort (NSM Ioffe + Wikipedia, v1). Frenkel-regime guard flags small-cluster / wide-gap-ionic materials where the Wannier-Mott assumption breaks down and the correction is suppressed. Higher-order exciton physics (anisotropic wurtzite, polaronic, GW-BSE accuracy) is still out of scope.
Temperature dependence E_g(T): 0–600 K Varshni form, median 24 meV accuracy on the v1 reference. For finer-grained T-dependence on strongly-anharmonic or quantum-paraelectric materials, pair with DFPT.
Dilutely-doped compositions are supported via the doping pipeline (Si:P, GaN:Mg, ITO, AZO, p-CdTe:Cu, compensated multi-dopant formulas) which returns activation energy, ionization fraction, Fermi-level offset, and carrier concentrations from the doped formula directly — the reported band gap remains the bulk host value. Surface-state-dominated gaps, grain-boundary states, and explicit defect-level energetics remain out of scope.
Family-stratified accuracy varies: intermetallics (0.072 eV MAE) and simple binary oxides (0.102 eV) are tightest; pnictides (0.195 eV), halides (0.222 eV), complex oxides (0.294 eV), and chalcogenides (0.332 eV) span the looser end.
The pipeline targets band-gap value — not band ordering, effective mass, or full DOS. Pair with the Materials, Semiconductor Design, or Surface modules for those (the Semiconductor Design module also exposes derived effective mass and dielectric constant alongside the doping bundle).

Put a SOTA pure-physics band-gap predictor in your screening loop.

Pilot access includes the Band-Gap module, the universal Materials engine, Semiconductor Design, Battery Electrochemistry, Surface & Contact, and a Workspace seat for audit-ready runs.

Open the benchmark Read the case study Request Pilot Access →

State of the art for composition-inputelectronic-property prediction.