Faruk Sami BilginNotes on quantum devices
A teardown · Notes on quantum devices

Inside a moiré qubit

Two crystals a few atoms thick, stacked with a small twist, make a lattice of identical traps for single electrons; here one such device comes apart from the chip down to the trap, followed by six devices from my own calculations and the numbers that decided them.

Faruk Sami Bilgin · November 2026

A moiré material is two atomically thin crystals laid on top of each other with a small twist. Across the stack, the places where the atoms of the two sheets line up and the places where they fall between each other alternate slowly, and that slow alternation, the moiré pattern, repeats over a distance far larger than the spacing of the atoms (the geometry is in chapter 1 of my notes). To an electron the pattern is a superlattice of identical wells, each several nanometres across. Each well can hold one electron or hole, and a trapped carrier is a candidate qubit: a quantum dot made by the crystal rather than by lithography.

The first part of this page takes a typical gated moiré device apart, from the chip on its carrier to the potential that a single hole feels. The second part builds six specific devices from my research, each with the computed numbers that decided whether it works.

Every device on this page is a model built from the geometries in my manuscripts, not a picture of fabricated hardware. None of them has been made. Where a dimension is not in my sources it is drawn at a generic value and marked illustrative.

Part I

The common device

A twisted bilayer of molybdenum disulfide, encapsulated in hexagonal boron nitride and gated from above and below, on a silicon chip. Twelve steps take it from millimetres to ångströms. The layer order is the usual one for such devices; the thicknesses, except where a step says otherwise, are generic.

Scroll, drag the slider at the bottom of the figure, or use ← → to step. Drag the model to turn it.

Step 1 of 12

The chip on its carrier

1 / 12

Going back up, the order of the layers is the logic of the device. Everything above and below the bilayer is there to shape the landscape of the last step or to keep it clean: boron nitride because it is flat and nearly free of charged defects, graphite because it screens the charges in the oxide, metal gates because a voltage on them can deepen or tilt the wells beneath. The landscape itself comes from the twist alone. That is the promise of moiré qubits, and, as the six devices below show, also where the difficulties begin.

Part II

Six devices

Each of the following is a device concept from my research, drawn as a model and tied to what I computed about it. The badge says whether the device is my own proposal or a published proposal that I analysed; all of them are computed, none fabricated. Each figure can be turned by dragging, and each has one control that changes the quantity that decided the verdict.

01
Twisted hBN

Artificial atoms in twisted boron nitride

Analysed device · computed

Twist two sheets of hexagonal boron nitride and the moiré makes a triangular array of identical wells, one per cell. A recent proposal (Song, Xu and Chang) treats each well as an artificial atom: an electron trapped at the AA site would have a ladder of levels 77.8 meV apart, a "magic angle" at 1.69°, and linewidths below 10 MHz up to 254 K.

Everything in that picture follows from one object, the moiré potential, so I computed it directly, with density-functional theory at 21 stacking registries. Toggle the figure between the two. The computed well sits where the proposal puts it, at AA, and has the same symmetry, but its equivalent amplitude is \(V_0 = 40\) meV instead of 160 meV, three to four times shallower, and the functional form the proposal uses is forbidden by the symmetry of the stack. The level spacing falls to 26–40 meV, and there is no magic angle.

A shallow well decides how long an excited electron stays excited. To fall from the first excited level (E) to the ground level (A₁) it must hand about 32 meV to the lattice. A single phonon with that energy carries more momentum than the smooth trapped wavefunction can easily take up, so it couples only weakly; two phonons sharing the energy, each with less momentum, escape that suppression. With them included, \(T_1 \approx 6\)–11 ps. A logical qubit needs about 30 ns.

What survives is still useful: the wells exist, they are identical, and the A₁→E transition carries a giant dipole, so the array remains interesting as an infrared emitter. As an orbital qubit it does not work.

Analysed device · computed
Moiré potential
Peak to peak1136 meV E₂₁ of the drawn rings79.5 meV Orbital T₁not computed for this well
Figure 1. Potential energy of an electron over a few moiré cells of twisted bilayer hBN at θ = 1.09° (moiré period 13.2 nm), drawn as height; both surfaces share one vertical and one colour scale. Proposed: Eq. 1 of the proposal, V₀ = 160 meV, with the sign that puts the well at AA (the repository notes that the sign is a convention). Computed: the three-star cosine fit to registry-resolved DFT, P₁ = −56.6, P₂ = +17.6, P₃ = +3.6 meV. Rings: the A₁ ground level and the E doublet solved on each surface in the repository, 79.5 meV apart on the proposed surface and 25.8 meV on the three-star fit (the curvature-constrained four-star fit gives 32.4 meV). The wave packets leaving the computed well are a cartoon of two-phonon emission.
+More — in my workPreprint and two repositories

The registry calculation, the symmetry argument against the proposed functional form, the levels and the absence of a magic angle are in the preprint. The relaxation budget is a separate study: one-phonon, two-phonon, flexural and sequential channels in two independent implementations, calibrated against the measured lifetime of phosphorus donors in silicon. An earlier internal draft that concluded the opposite rested on a one-phonon budget; its errors are documented in the repository.

02
Twisted MoS₂

A hole spin in twisted MoS₂

Proposed device · computed

Twisted bilayer MoS₂ is dark. Its valence band peaks at Γ and its conduction band at K, so for the lowest excitons to emit light the moiré would have to supply about 16 of its reciprocal vectors at 2°; it supplies about one. It cannot be an optical qubit. But the same band structure puts the top of the valence band where a hole has no valley and no spin–orbit coupling: a plain spin ½ in every moiré well, a natural quantum dot whose envelope, about 1.2 nm in radius, covers about 320 nuclei.

Those nuclei are the problem. A quarter of natural molybdenum atoms carry a nuclear spin (⁹⁵Mo and ⁹⁷Mo), and their random sum acts on the hole as a magnetic field it cannot tell from a real one: \(T_2^* \approx 42\) ns. Isotope purification, ⁹⁸Mo and ³²S, which have no nuclear spin, removes them. Move the slider: \(T_2^*\) grows as one over the square root of the fraction that remains.

It grows until the encapsulation takes over. Boron and nitrogen have no spin-free isotope, so the hBN nuclei just outside the bilayer set a floor: 0.79 µs with the field perpendicular to the plane, 1.3 µs with it in plane. Independent noise sources add as

$$\frac{1}{T_2^{*2}} = \frac{1}{T_{2,\mathrm{MoS_2}}^{*2}} + \frac{1}{T_{2,\mathrm{hBN}}^{*2}},$$

so a highly enriched, hBN-encapsulated device reaches 0.7–1.0 µs, the level of natural silicon. Going further would need a spinless spacer between the bilayer and the hBN.

Proposed device · computed
0.25 (natural)
Magnetic field
Spinful host nuclei drawn— T₂*, MoS₂ nuclei— T₂*, with hBN—
Figure 2. One Bernal well of twisted bilayer MoS₂ at θ = 2°: the atoms within 4.5 nm of its centre, Mo teal and S yellow, with one hBN layer drawn on each side (B pink, N blue; the top one cut away over the front half); vertical distances stretched 2.5 times. The red arrow is the hole spin, the red cloud its envelope, |ψ|² ∝ exp(−r²/ℓ²) with ℓ = 1.22 nm. Small arrows mark nuclei with spin: each host nucleus is drawn spinful at random with the probability set by the slider (Mo: f; S: f × 0.0076/0.2547, the scaling of the repository's scan); every hBN nucleus has spin. T₂* is the repository's quasi-static estimate, 42.4 ns × (0.2547/f)^½ for the MoS₂ nuclei, combined in quadrature with the hBN floor.
+More — in my workPreprint · Twisted MoS₂ cannot talk to light

The preprint establishes the momentum-dark band character from four independent lines of evidence, including my own DFT (the Γ valence maximum lies 0.477 eV above K), solves the moiré dot (barrier 77.4 meV, ħω = 57.0 meV) and builds the coherence budget, with a GaAs probe test of the same formula (6.5 ns against about 10 ns measured). It predicts a 17–23-fold Ramsey contrast between natural and highly enriched devices.

The Γ-valley hyperfine constant was not computed directly; published K-valley constants were transferred through DFT orbital weights, and a deliberately pessimistic case gives 0.14 µs.

03
Twisted WSe₂

Valley qubits: by wire or by light

Analysed devices · computed

In twisted WSe₂ the qubit can live in the valley of a hole, which of two corners of the Brillouin zone it occupies. There are two ways to control it. A gate-defined dot, proposed by Saleem and co-workers, confines a hole under electrodes; the dot's lowest level splits into a valley doublet, and gate voltages drive it. The alternative uses light: at small twist one hole sits in each moiré cell (a Mott insulator), and two laser beams rotate its spin–valley state through a trapped trion, a Raman gate.

I compared both in the same material on one pipeline, checked against the magic angle (1.42° predicted, 1.43° published) and against the proposal's own ladder of valley splittings. The comparison is not symmetric in angle: each paradigm works at a different twist of the same crystal. The gate dot sits at 5.08°; the Mott insulator needs small angles, where the ratio of interaction to bandwidth reaches 10²–10³, and at 5.08° that ratio has collapsed to 2.7. Toggle the figure between the two.

Electrically, the doublet splitting depends steeply on the size of the dot: 382, 55.3 and 3.41 µeV for radii of 5, 7 and 10 moiré periods. Optically, a π gate takes 0.22 ns at an error of 4 × 10⁻³, but only if the branching ratio β, the fraction of trion decays that return to the other qubit state, is near one; spin–valley locking keeps it small, and no knob is known that makes it large. The lasers need 0.02–11 mW per site.

Twist disorder hits both. At a specification of σθ = 0.036°, the electrical dots spread by 4.5 % and need per-site calibration; the optical sites spread over about 17 frequency channels. At the disorder actually measured in twisted WSe₂, more than 1° over micrometres, both budgets fail.

Analysed devices · computed
Control
5 aM = 18.7 nm
1
Twist5.08° Moiré period3.74 nm Valley splitting Δ₁S382 µeV At σθ = 0.036°4.5 % spread
Figure 3. Electrical: a gate-defined dot at θ = 5.08° (moiré period 3.74 nm), drawn with split gates, two barriers and a plunger leaving an opening of the computed dot radius; the glow is a schematic 1S envelope. Optical: the ν = 1 Mott lattice at θ = 1.43° (period 13.3 nm), one hole per site, under a pair of focused Raman beams; the spot covers many sites, and the red one is the site resonant with the beams. Gate shapes, gate height (10 nm), beam size and the boron nitride, which is not drawn, are illustrative. Δ₁S and the Raman gate times and errors are from the repository (trion linewidth 50 µeV, corrected AC-Stark sign).
+More — in my workTwo repositories, one in revision

An internal audit found a sign error in the AC-Stark term of the Raman budget; every optical number here is from the corrected budget, and the earlier ones are withdrawn. Neither side's T₂ is measured or reliably computable, which bounds what the comparison can claim.

A processor architecture built on the electrical dots (singlet–triplet pairs at a 40 nm pitch, exchange 1.6–63 µeV in the target window) is in revision; only its audited results are public.

04
MoSe₂ / WSe₂

An exciton emitter in a nanobeam cavity

Analysed device · computed

In a heterobilayer of two different crystals, MoSe₂ on WSe₂, an electron in one layer and a hole in the other bind into an interlayer exciton. The moiré traps it, and each trapped exciton is a quantum emitter. On its own it is slow: 139 ns in my calibrated pipeline, against about 200 ns measured. A cavity can speed it up, the Purcell effect, by giving the photon a resonant mode to go into.

In the textbook, the enhancement grows with the cavity's quality factor \(Q\). A real emitter, though, has a linewidth \(\gamma\), and once the cavity line \(E/Q\) is narrower than the emitter's, more \(Q\) buys nothing. The dephasing-corrected budget is

$$F_\mathrm{eff} = \frac{3}{4\pi^2}\,\frac{\xi}{V}\,\frac{E}{E/Q + \gamma},$$

which saturates at \(Q_{1/2} = E/\gamma \approx 1.4\times10^4\) for the best linewidth reported, 100 µeV, and at \(2.1\times10^3\) for the 640 µeV line measured in a cavity at 10 K. Slide \(Q\) and watch the curve flatten.

So the currency is not \(Q\) but the mode volume \(V\) and the overlap \(\xi\), the share of the peak field the emitter actually sits in. My 3D FDTD computation of a suspended silicon-nitride nanobeam gives \(Q = 7.9\times10^3\) and \(V = 0.19\,(\lambda/n)^3\), but a crystal laid on top sees only \(\xi = 0.032\): a tenfold penalty for sitting on the surface, which the small volume only partly repays.

Analysed device · computed
7.9 × 10³
Emitter linewidth γ
Feff— Lifetime 139 ns / Feff— Q / Q₁/₂—
Figure 4. The silicon-nitride nanobeam of the FDTD computation: 500 nm wide and 200 nm thick, holes at a period of 290 nm with radius 0.28 of the period, tapering over six holes to radius 0.19 and period 0.86 at the centre, ten mirror holes on each side. The supports and the trench below are illustrative, as are the lateral size and thickness of the hBN / MoSe₂–WSe₂ / hBN flake on top. The mode glow is schematic. Feff, the emission cone and the plot are evaluated in your browser from the budget formula with E = 1.37 eV, V = 0.19 (λ/n)³ and ξ = 0.032.
+More — in my workCalibration, disorder, and two studies in revision

The same pipeline passes eight experimental anchors within pre-declared tolerances, and shows that an aligned WSe₂/WS₂ moiré, whose period is set by lattice mismatch rather than twist, suppresses twist-disorder scatter 42-fold at σθ = 0.1° (47 µeV against 1.9 meV). The budget also explains the one published moiré-exciton cavity experiment as collective coupling of 70–235 emitters.

Two related studies are in revision and report only audited results: an erasure-biased architecture that uses the position of a dark interlayer exciton in a double trap as a qubit, and a DFT study of twisted Janus WSSe and MoSSe, where the valence maximum and conduction minimum sit on the same site (K–K phase separation 174° and 178°).

05
Magic-angle graphene

A twisted-graphene junction in a sub-terahertz cavity

Proposed device · computed

Twisted bilayer graphene at its magic angle, 1.025° in my model, where the flat bands are 2.9 meV wide, superconducts. Gate a strip of it between two superconducting leads and you have a Josephson junction whose current–phase relation is skewed by its high transparency, and a skewed relation shows half-integer Shapiro steps under microwaves (chapter 3 of my notes; the junction itself is explained in the long read).

The design puts the junction in the defect of a silicon photonic-crystal slab: three missing holes, an L3 cavity, in a lattice of 587 µm, which traps a mode at 145 GHz with an in-situ \(Q_i = 1.46\times10^4\). The cavity is millimetres across, the junction micrometres; the slider zooms from one to the other. A bare van der Waals stack would barely feel the cavity, about 1 µV of drive instead of 600 µV, so an engineered collection structure 20–25 µm long is part of the design.

At \(I_c = 300\) nA and \(R_n = 200\) Ω, the half step is \(0.200\,I_c\), 60 nA, wide: four times the floor set by a sinusoidal junction. It survives heating only while \(E_J/k_BT > 55.5\), so the electrons must be colder than 129 mK; a parametric estimate puts them near 0.6 K at the reference drive.

A companion study draws the limits. The step measures transparency, not mechanism: trapped vortices and multiple Andreev reflection mimic its dependence on filling and twist. A quantum-geometric readout is ruled out, the Higgs mode is overdamped (\(Q_H \approx 1.5\)), and 145 GHz sits at or above the pair-breaking edge, \(\hbar\omega/2\Delta = 0.99\)–1.9. A safe, adiabatic drive is 15–29 GHz.

Proposed device · computed
view —
Lattice constant586.74 µm Drive ħω / 2Δ0.99–1.9 Half-step survival at 100 mK0.585
Figure 5. High-resistivity silicon slab with a triangular lattice of holes (a = 586.74 µm, radius 0.3a, thickness 0.5a; 9 rows drawn) and an L3 defect: three missing holes, the end holes shifted outward by 0.20a, the surrounding ring of holes enlarged 1.15 times. The red glow marks the 145 GHz mode, schematically, slowed down enormously. At the centre: an hBN-encapsulated twisted-graphene channel between two superconducting leads, fed by a collection structure about 22 µm long. Its bow-tie shape, the channel size and the lead material are illustrative.
+More — in my workThree repositories on one platform

The fractional-Shapiro study chains four modules (material, cavity, junction, integration), each re-derived independently, and turns the floor-corrected half step into an electron thermometer for 36–715 mK. The limits study records two errors its validation protocol caught and withdrew.

Shunted into the transmon regime, the same junction would be a gatemon at f₀₁ = 40.6 GHz; the drive cavity cannot read it out, and the Shapiro experiment and the qubit cannot run on the same chip at once.

06
Bilayer graphene

Shaking a quantum dot does not quiet it

Proposed device · computed · refuted

In bilayer graphene a perpendicular electric field opens a band gap, and gates can carve a quantum dot out of the gapped sheet. Its weakness is charge noise: any fluctuating charge nearby shifts the dot's levels. I asked whether shaking the confinement, modulating the dot potential at a frequency well above the qubit's, could average that noise away and make a dynamical sweet spot.

The model is a four-band continuum with a Gaussian dot 15 nm wide and 20 meV deep at a displacement gap of 0.167 eV, where the two lowest levels are 0.989 meV apart; an independent atomistic solver reproduces that to three digits. Averaged over one period of the drive, the modulation seemed to help: T₂* rose by 9 % at an amplitude of 50 meV.

It does not survive. The averaging argument needs the level splitting to bend down as a function of the field; it bends up (curvature +0.48). And when the drive is followed in time, in the full space of states, it throws the electron out of the two qubit levels: 38–95 % of the population leaves every period at the amplitudes where the protection appeared. Where the qubit survives, the effect is negligible. The slider steps through the computed amplitudes.

What did survive is symmetry: a mirror-symmetric double dot, and a spin qubit inside one valley, are protected; the tunable sweet spots were not.

Proposed device · computed
0 meV
Leakage out of the qubit per period0 % Driveħω = 6.6 Δ₀₁
Figure 6. A gate-defined dot in hBN-encapsulated bilayer graphene: split gates leave a channel, finger gates across it (lifted 50 nm here so that the dot can be seen) define the dot; graphite back gate below. Gate shapes and spacings are illustrative. The dot is the model's Gaussian, U₀ = 20 meV, Ldot = 15 nm; the cloud is drawn with a schematic threefold warping (the computed states show C₃ trefoils). The drive is slowed down enormously; the share of the cloud that escapes each period is the computed Floquet leakage.
+More — in my workA refuted hypothesis, documented

The repository keeps both the original claim and its refutation: the period-averaged scans, the convexity check, the full-Hilbert Floquet calculation with its time-step convergence, and six candidate sweet spots that dissolved under convergence tests.

07
Scale

Why none of these is a logical qubit yet

A logical qubit is not a better physical qubit. It is hundreds of physical qubits, each controlled individually, running error correction together. I asked whether a moiré platform could reach 300 logical qubits, starting from its best material, isotopically purified twisted MoS₂ (device 2), and attacking every design with an independent reviewer who wrote separate code. The material and the qubit type survived. The control layer did not.

The reason is a mismatch of scales. To address one bond of the array, a control line must set an energy to within 5 µeV to 0.23 meV, depending on the scheme; the moiré sites themselves scatter by about 15 meV. A metal line 3–20 nm above the plane spreads its potential over 16–44 nm, wider than the 8 nm bond it is meant to address. The least-bad scheme needs 607–809 thousand analog lines for 300 logical qubits; an active matrix would cut the wiring, but it needs a memory node at a 12.5–14 nm pitch that no technology offers. Until site disorder falls to 3–5 meV and such a node exists, moiré qubits remain a research platform, and the main line for a large processor is a conventional gate-defined one.

Further in my notes