Gallium nitride is a wide-bandgap semiconductor whose AlGaN/GaN heterostructures underpin high-efficiency RF amplifiers and power switches; the current research frontier is moving it from lateral devices on foreign substrates to vertical devices on cheap silicon wafers, where interface resistance is still the binding constraint.
Summary
Gallium nitride (GaN) is a compound semiconductor with a wide bandgap and high electron velocity. Its most commercially important trick is the heterostructure: stacking a thin aluminium gallium nitride (AlGaN) or aluminium nitride (AlN) barrier on a GaN channel creates a two-dimensional electron gas at the interface without intentional doping, giving a high-electron-mobility transistor (HEMT) that switches faster and with lower loss than a comparable silicon device. This makes GaN attractive in two distinct markets: radio-frequency power amplifiers, and power conversion (chargers, inverters, data-centre supplies). The same material family also emits blue and ultraviolet light, which is why GaN epitaxy and micro-LED work sit alongside the transistor work.
The engineering problem is almost entirely about crystal growth on substrates that are not GaN. Bulk GaN wafers are scarce, so devices are grown epitaxially on silicon or sapphire, which have different lattice constants and thermal expansion. Today’s mainstream devices are lateral: current flows sideways through the 2DEG, and the buffer layer that absorbs the lattice mismatch does not have to conduct. Vertical devices, which would carry current through the substrate and scale better to high voltage and to dense micro-LED arrays, are blocked precisely because those conventional buffer layers are electrically resistive. One 2026 approach forms a sub-nanometre silicide template on Si(111) by sputtering plus rapid thermal annealing, producing an amorphous-like interlayer that relaxes strain while remaining ohmic, and reported to work across 25 different metal species before MOCVD overgrowth. Related growth work is still at the level of controlling where GaN nucleates on patterned sapphire as a function of mask growth temperature.
The parameters that decide GaN device quality are visible in the device papers. For HEMTs: channel sheet resistance, threshold voltage and its stability, on/off current ratio, on-resistance and gate leakage. N-polar AlN-based HEMTs with a 5.2 nm GaN channel reached on/off ratios of 10,000 and saturation currents over 240 mA/mm, but on-resistance of about 20 Ω·mm and the high-frequency performance were all limited by a 2000 Ω/square channel sheet resistance. For circuits, the figure of merit is efficiency at back-off: GaN HEMT Doherty amplifier prototypes measured over 44.2 dBm saturated output with peak drain efficiency above 71.2% across 2.6-2.8 GHz, and 64% at 6 dB back-off with ACLR better than -51.3 dBc after digital predistortion.
A supporting theme in the 2026 literature is that GaN process development is data-limited, and machine learning is being used to compensate. A hybrid classical-quantum neural network trained on 468 fabricated recessed-gate AlGaN/GaN MIS-HEMTs across 17 process splits predicted six electrical targets from a 24-dimensional process vector, cutting normalised RMSE by 24.4% against a classical ANN baseline. Similar tooling appears in materials simulation (machine-learning-accelerated pathway calculations including GaN, up to 7-fold speedup) and in contact-free doping metrology by terahertz spectroscopy covering SiC, Si and GaN layer stacks.
Evidence base
Open questions
- Does the sputtered sub-nanometre silicide template on Si(111) survive to a complete vertical device with measured blocking voltage, leakage and switching loss, and at what wafer diameter?
- Can the 2000 Ω/square channel sheet resistance in N-polar AlN/GaN HEMTs be reduced without losing the abrupt heterostructure that delivered the 10,000 on/off ratio?
- Does the 24.4% surrogate-model accuracy gain on 468 MIS-HEMTs generalise to other process flows and to yield prediction, or is it dataset-specific?
- How does GaN compare with silicon carbide and silicon on cost per amp and per volt at equivalent reliability? No supplied source addresses this.
Assessment drafted 2026-08-31 from up to 18 KB sources using the technology-scorecard framework; scores are a draft read pending review.