Screening the Gallium-Oxide Family for Power Electronics

The challenge: power figure of merit vs heat

A power transistor's voltage ceiling is set by its material's breakdown field, which rises steeply with band gap — so an ultra-wide-gap oxide can, in principle, exceed 4H-SiC and GaN on the Baliga figure of merit, the standard material-quality yardstick for power switching. Two well-known obstacles temper that promise: these oxides tend to conduct heat poorly, and most cannot yet be doped both n- and p-type. A fast physics-based screen is useful precisely for triaging a large composition space against these constraints before committing to crystal growth.

Question 1

Power figure of merit

Which members of the family — and the alloys among them — rank highest on breakdown-limited and thermal-limited figures of merit?

Question 2

Band-gap tuning

How far can the gap be engineered by alloying, and where does that trade against carrier mobility?

Question 3

Dopability

Which members show a valence band capable of hole transport — a route to the p-type behaviour β-Ga₂O₃ lacks?

The pipeline: three evaluation tiers, then two gates

The study is deliberately staged so that the expensive step is applied only where it matters. The three tiers differ in fidelity and in what they can be trusted to say.

10

Full-property evaluations. Ten pure anchor oxides run through the Universal Materials Engine (structure, band gap, mobility, elastic, thermal), ~2 s total.

1.2M

Interpolated alloy screens. Every sesquioxide alloy of the anchors (binary–quinary, 5% grid) scored by Vegard interpolation of the anchor properties — a coarse ranking pass, ~0.2 s, not a physics evaluation of each composition.

6

Candidate deep dives. Six finalists re-evaluated through the Universal Materials Engine to recover the properties interpolation cannot. The archived end-to-end campaign completed in approximately 2.4 s.

Two manufacturability gates then filter the ranking: a miscibility gate (textbook ionic-radius mismatch vs documented oxide-alloy solubility) that removes phase-separating combinations, and a measured thermal-conductivity table that replaces the engine's thermal proxy for the endpoints. The final shortlist below is the output after both gates.

The shortlist

After miscibility and thermal gating, the candidates fall into three evidence tiers. Baliga figures of merit are relative to silicon and computed from physics-derived properties (all figures apples-to-apples through the same framework); the incumbents are shown on the same basis.

Literature-supported — real, growable compound with independent experimental/computational literature Prospective lead — ranks well, plausible, not yet experimentally validated for this use Exploratory — screening output with known synthesizability concerns

Candidate	Family	Baliga FoM (× Si)	vs GaN	Evidence tier
Al-rich (Al,Ga)₂O₃	sesquioxide alloy	∼3,800	∼18×	Literature-supported
ZnGa₂O₄	spinel gallate	591	2.8×	Literature-supported
LiGaO₂	alkali gallate	2,188	10×	Prospective lead
NaGaO₂	alkali gallate	2,157	10×	Exploratory
— GaN (incumbent)	incumbent	214	1.0×
— 4H-SiC (incumbent)	incumbent	155	0.7×

The two literature-supported entries are the strongest results: Al-rich (Al,Ga)₂O₃ is the alloy the β-Ga₂O₃ community already uses to engineer the gap,^2,19,20 and ZnGa₂O₄ is an actively pursued ultra-wide-bandgap oxide with demonstrated bipolar (n- and p-type) self-doping^12,13 — directly relevant to Question 3. The Baliga figure of merit is a material-quality ceiling, not a device guarantee: realizing it still depends on doping and heat extraction that are unproven for most of these.

How the shortlist survived its own first draft

The shortlist above is what remained after the screen was made to confront its own weaknesses. That process is worth showing, because it is where the credibility comes from.

First-pass failure

A breakdown-only optimum that cannot be grown

Ranked on the breakdown figure of merit alone, the top alloy was Al-rich (Al,In)₂O₃. But aluminium and indium have a large ionic-size mismatch, and the published record shows Al-rich (Al,In)₂O₃ tends to amorphize or phase-separate at that composition.^14,15 A figure-of-merit optimum is not a synthesizability guarantee — the interpolated screen has no term for whether two cations will share one lattice.

The fix

Two gates the screen was missing

A miscibility gate (Shannon ionic-radius mismatch vs documented solubility) removes phase-separating combinations — on this space, 94% of the raw alloy count, including the (Al,In)₂O₃ and boron-bearing mixes. A measured thermal-conductivity table replaces the engine's thermal proxy, which is unreliable for light and glass-forming oxides. After both gates, the optimum moves to the growable (Al,Ga)₂O₃ alloy.

Validation scope, property by property

The conclusions rest on several properties with different evidence behind them. Only the band gap is validated against a literature reference dataset in this study; the others are disclosed here at their true level rather than bundled under one claim.

Property	Evidence level	Basis
Band gap	Validated here	17 wide-gap oxides scored against literature reference gaps — 10 from measured primary sources, 5 from a semiconductor-property compilation, 1 from a secondary source, and 1 (NaGaO₂) a many-body GW computation where no measurement was located. No DFT-PBE values are used as targets. The reference gaps mix optical, fundamental, and direct/indirect transitions (labelled per-entry in the download), so individual comparisons are approximate; mean absolute error ~0.6 eV, material-specific.
Thermal conductivity (endpoints)	Literature-supplied	Room-temperature values curated per anchor oxide from the literature, with source URL and confidence disclosed per endpoint in the download. Alloys are interpolated; unsupported Sc₂O₃/Y₂O₃/Lu₂O₃ endpoint values are excluded.
Electron mobility	Benchmarked elsewhere	Derived by the Universal Materials Engine; benchmarked against literature in the semiconductor mobility atlas study (not re-validated on this oxide set).
Breakdown field	Estimated	Derived from the band gap via an empirical scaling (E₋ ∝ E₀^2.5), standard for figure-of-merit ranking; not an independent calculation.
Dielectric constant	Predicted	Engine prediction, not independently validated on this set.
Dopability (n/p-type)	Proxy + literature	Hole-transport screen is a proxy, not validated; the ZnGa₂O₄ bipolar-doping claim is literature-supported.^12,13
Miscibility	Empirical gate	Ionic-radius mismatch vs a documented oxide-alloy solubility boundary — a literature-grounded heuristic, not a first-principles free-energy calculation.
Melting point	Not reliable — omitted	The engine's melting prediction is unreliable for these oxides (it under-predicts β-Ga₂O₃ substantially¹⁸) and is excluded from the datasheet below.

β-Ga₂O₃4.53 eVmeas. 4.8¹

Al₂O₃ (sapphire)8.70 eVmeas. 8.7²

Sc₂O₃6.0 eVmeas. 6.0⁵

ZnGa₂O₄4.4 eVmeas. 4.7⁴

LiGaO₂6.3 eVmeas. 5.4³

Bi₂Ga₄O₉4.2 eVmeas. 2.9⁶

Band-gap error is roughly unbiased across the set (material-specific, not a systematic over- or under-shoot), which supports using the rankings while individual gaps remain approximate.

Convergence with the published record

A screen is more credible if it lands where independent evidence lands. FluxMateria's Universal Materials Engine derived these candidates using proprietary FLUX Theory physics and textbook atomic data, without access to the power-electronics literature; the ranking was compared to that literature only afterward.

Candidate	Status in the literature	Reference
ZnGa₂O₄	Actively pursued UWBG oxide for “energy electronics”, with demonstrated bipolar self-doping (n- and p-type); crystal growth ongoing.	Chikoidze 2020;¹² Lyle 2021¹³
(Al,Ga)₂O₃ alloy	The established gallium-oxide bandgap-engineering alloy: α-(AlGa)₂O₃ spans 5.4–8.6 eV by MBE, and β-(AlGa)₂O₃ is grown by MOCVD and bulk Czochralski.	Sci. Adv.;² Cryst. Growth Des.;¹⁹ J. Appl. Phys.²⁰
MgGa₂O₄	UWBG confirmed with noted power potential; worked on to date for photodetectors and phosphors, not yet power devices.	Ref. 16
LiGaO₂	Known UWBG oxide, bulk crystals grown; the Li–Ga–O family is being explored for power (related LiGa₅O₈ demonstrated as p-type with β-Ga₂O₃).	Ref. 3, 17
NaGaO₂	No power-electronics literature located — the one genuinely under-explored candidate.	—

Recovering directions the field is independently pursuing — including ZnGa₂O₄'s bipolar doping, the specific escape from β-Ga₂O₃'s p-type wall — supports the pipeline's ranking credibility and motivates prospective validation. It does not, by itself, establish that the under-explored predictions (e.g. NaGaO₂) will hold; that is what the experimental plan below is for.

What the same search would require with other methods

The relevant comparison is not one FluxMateria prediction against one DFT calculation. It is the cost of traversing the same 1,201,617-entry composition grid, applying multi-property filters, and reducing it to a shortlist suitable for higher-cost validation.

FluxMateria campaign

Physics-led triage at interactive speed

2.4 s

Archived wall time for the staged campaign: full-property anchor evaluations, 1.2-million-composition interpolation and gating, then candidate re-evaluation. No project-specific ML model was trained for this target.

Output: a four-candidate public shortlist, including two directions independently supported by the literature and one under-explored lead.

~5.5×10⁸
CPU-hours

Literal DFT coverage of 1.2 million entries

A historical Materials Project scale — 15 million CPU-hours for 33,000 compounds — extrapolates to roughly 546 million CPU-hours, or about 62,000 CPU-years.²¹ Even 10,000 continuously utilized cores would imply about six years. This is an order-of-magnitude illustration, not a modern-code benchmark, and it is optimistic because each disordered alloy requires explicit structures and often multiple configurations.

Days +
HPC

Realistic DFT plus surrogate workflow

No serious program would run DFT on every grid point. It would calculate a selected structure set, build a cluster-expansion or ML surrogate, screen cheaply, and return finalists to higher fidelity. Cluster expansion exists precisely because exhaustive first-principles treatment of disordered configurational spaces is impractical.²² Published Materials Project workflows describe new-compound electronic and thermodynamic turnaround on the order of days, while multi-fidelity optimization can reduce total cost by about threefold rather than eliminate the reference calculations.^21,23

Fast inference,
large pre-cost

Pretrained machine-learning screening

ML can also score millions of candidates rapidly once a suitable model and structural representation exist. The trade is upstream: training data, domain coverage, uncertainty control, and high-fidelity verification. GNoME, for example, used iterative graph-network filtering followed by DFT verification to explore millions of crystal structures; its target was stability, not this power-electronics property stack.²⁴

Days to years
in the lab

Combinatorial or autonomous experiment

A published thin-film workflow reports eight hours just to composition-map 342 measurement areas.²⁵ Literal extrapolation to 1.2 million points is about 28,100 hours, or 3.2 years, for that measurement alone. At the shortlist end, an autonomous laboratory synthesized 36 of 57 inorganic targets during 17 days of continuous operation.²⁶ Neither figure includes power-device fabrication or full electrical and thermal validation.

Interpretation: no alternative method is guaranteed to return an identical shortlist because each represents the search space differently. DFT evaluates explicit structures, ML inherits its training domain, and experiments sample specific processing conditions. The defensible result is that FluxMateria reduced a million-scale compositional hypothesis space to a validation-sized set in seconds. DFT, defect calculations, crystal growth, and device testing now belong on that reduced set — not on all 1.2 million entries.

Limitations

Stated plainly, so the results are used for what they are.

The 1.2M figure is an interpolated ranking pass over the anchors, not 1.2M physics evaluations.
Only the band gap is validated on this oxide set; other properties carry the evidence levels tabulated above.
The miscibility gate is an empirical solubility heuristic, not a computed mixing free energy; its threshold is grounded in documented phase behaviour, not derived.

Predicted thermal conductivity is the engine's least-settled output; endpoints use measured values, alloys are interpolated.
Breakdown field is an empirical function of the band gap, not an independent calculation.
Sc-rich alloy rankings are withheld because the located mixed-sesquioxide measurement does not establish the pure Sc₂O₃, Y₂O₃, and Lu₂O₃ endpoint conductivities required for that interpolation.¹⁰
Bi₂Ga₄O₉ is a visible-light absorber, not a transparent conductor; the exploratory NaGaO₂ and the (Al,In)₂O₃ alloy carry specific synthesizability concerns.

What happens next: a staged validation plan

For any candidate the screen elevates, the path to a defensible result is a sequence of measurements, each gating the next. The screen's role is to make this sequence sharp and directed — not to replace it.

Phase & miscibility confirmation
Grow or deposit the composition; confirm single-phase solid solution (XRD, EDS/EPMA) — the first gate for any alloy candidate, especially the flagged mixes.
Optical band gap
Measure the gap (spectroscopic ellipsometry / diffuse reflectance) against the predicted value.
Hall mobility & carrier transport
Hall-effect measurement of carrier type, concentration, and mobility on doped samples.
Thermal conductivity
Direct measurement (time-domain thermoreflectance / laser flash) — the property most in need of confirmation for a power device.
Dopability & defect characterization
Establish achievable n- and (where claimed) p-type doping and the limiting compensating defects.
Breakdown testing
Measure the critical field on test structures rather than inferring it from the gap.
Prototype device fabrication
Only after the above: a simple diode or transistor to test the figure of merit in a real device.

Method, data & references

This is a hybrid screening pipeline built on FluxMateria's Universal Materials Engine, not an all-first-principles calculation. The engine derives electronic properties (band gap, mobility, dielectric constant) from its underlying proprietary FLUX Theory physics plus textbook atomic data, with zero parameters fit to the target property; the final ranking then combines those outputs with literature-measured thermal conductivity, an empirical breakdown-field scaling (E₋ ∝ E₀^2.5), and an empirical miscibility gate. The figures of merit (Baliga for power; the thermal-limited Keyes metric) are the standard published metrics, computed from that combined property set.

Download the underlying data

Case-study results — JSON, CC BY 4.0

The 17-material band-gap set (predicted vs literature reference, each row tagged with its source type — measured, compilation, or computed), the thermal-conductivity table, the shortlist, the full-property deep-dive datasheet, the supported hardened result, the methods comparison, the figure-of-merit leaderboard, prior-art status, and the reference list — each entry tagged with its evidence level.

Download JSON ↓

References

Recent Advanced Ultra-Wide Bandgap β-Ga₂O₃ Material, Adv. Electron. Mater. (2023). doi:10.1002/aelm.202300844
Crystal-orientation-dictated epitaxy of UWBG 5.4–8.6 eV α-(AlGa)₂O₃, Sci. Adv. (2021). doi:10.1126/sciadv.abd5891
Optical properties of lithium gallium oxide, Appl. Surf. Sci. (2017). sciencedirect.com/…/S0169433217300995
Optical band-gap values of spinel AB₂O₄ (A=Mg,Zn; B=Al,Ga), ACS Org. Inorg. Au (2023). doi:10.1021/acsorginorgau.3c00030
The optical energy gap of scandium oxide, OSTI 4014535. osti.gov/biblio/4014535
Structure, electronic and optical properties of mullite-type Bi₂M₄O₉ (M=Al,Ga), Inorg. Chem. (2016). doi:10.1021/acs.inorgchem.6b00330
QSGW band structures of LiGaO₂ and NaGaO₂, arXiv 2010.15934. arxiv.org/abs/2010.15934
Klimm et al., The thermal conductivity tensor of β-Ga₂O₃ from 300 to 1275 K, Cryst. Res. Technol. (2023). doi:10.1002/crat.202200204
Thermal conductivity of B₂O₃ glass under pressure (0.52 W·m⁻¹·K⁻¹ at ambient), Int. J. Thermophys. doi:10.1007/BF00522148
Czochralski growth of a mixed cubic Lu₂O₃–Sc₂O₃–Y₂O₃ crystal, including a 4.1 W·m⁻¹·K⁻¹ room-temperature measurement for one Er-doped mixed composition, IUCr (2021). This result is not used as a pure-endpoint conductivity. doi:10.1107/S2052520621005321
Thermal conductivity of bulk In₂O₃ single crystals, arXiv 2008.13519. arxiv.org/abs/2008.13519
Chikoidze et al., p-Type ultrawide-band-gap spinel ZnGa₂O₄: new perspectives for energy electronics, Cryst. Growth Des. (2020). doi:10.1021/acs.cgd.9b01669
Bipolar self-doping in ultra-wide bandgap spinel ZnGa₂O₄, Mater. Today Phys. (2021). sciencedirect.com/…/S2542529321001279
Liu et al., Structural and electronic properties of corundum and monoclinic (Al₁₋ₓInₓ)₂O₃ alloys, Adv. Theory Simul. (2025). doi:10.1002/adts.202401407
Influence of Al content on ternary Al₂ₓIn₂₋₂ₓO₃ alloy films (MOCVD; phase behaviour), ScienceDirect (2015). sciencedirect.com/…/S0025540815003219
Beta-to-spinel phase transition of magnesium gallium oxide thin films, ACS Appl. Electron. Mater. (2024). doi:10.1021/acsaelm.4c01079
Ultrawide-bandgap LiGa₅O₈/β-Ga₂O₃ heterojunction p–n diodes, APL Electron. Devices. pubs.aip.org/…/016115
Ga₂O₃ and related ultra-wide-bandgap power semiconductor oxides (review of the UWBG-oxide class). PMC8838167. β-Ga₂O₃ melting point ~2093 K is reported widely in the crystal-growth literature.
MOCVD epitaxy of ultrawide-bandgap β-(AlₓGa₁₋ₓ)₂O₃ with high-Al composition on (100) β-Ga₂O₃ substrates, Cryst. Growth Des. (2020). doi:10.1021/acs.cgd.0c00864
Alloyed β-(AlₓGa₁₋ₓ)₂O₃ bulk Czochralski single crystals and property trends, J. Appl. Phys. 131, 155702 (2022). pubs.aip.org/…/155702
Jain et al., The Materials Project: a materials genome approach to accelerating materials innovation, APL Mater. (2013). Reports several hundred CPU-hours for basic properties of one material, more than 15 million CPU-hours for over 33,000 compounds, and days-scale turnaround for new-compound calculations. doi:10.1063/1.4812323
Chang et al., CLEASE: a versatile implementation of cluster expansion for disordered materials, J. Phys.: Condens. Matter (2019). doi:10.1088/1361-648X/ab1bbc
Fare et al., A multi-fidelity machine-learning approach to high-throughput materials screening, npj Comput. Mater. (2022). doi:10.1038/s41524-022-00947-9
Merchant et al., Scaling deep learning for materials discovery, Nature (2023). doi:10.1038/s41586-023-06735-9
Ludwig, Discovery of new materials using combinatorial synthesis and high-throughput characterization of thin-film materials libraries, npj Comput. Mater. (2019). doi:10.1038/s41524-019-0205-0
Szymanski et al., An autonomous laboratory for the accelerated synthesis of inorganic materials, Nature (2023). doi:10.1038/s41586-023-06734-w

Every band-gap and thermal-conductivity reference value, and the growth/doping claims for the shortlisted materials, were located from the cited sources during this study; DOIs and links are provided for verification, and each downloadable data row carries its own source URL and source type (measured / compilation / computed).

Test the screen against your own data

Have a gallium-oxide or wide-gap-oxide dataset with hidden targets — your own measurements, an internal benchmark, an unpublished campaign? We will run the screen blind and return predictions for comparison, with optional independent scoring.

Submit a blind benchmark → Explore the atlas

Screening the gallium-oxide family for power electronics