The Josephson junction
How two films of aluminium separated by a nanometre of oxide become the heart of a superconducting qubit, explained with figures you can move.
A superconducting quantum processor is, at first sight, an unremarkable object: a square of silicon a few millimetres across, patterned with metal films a few hundred nanometres thick. Nothing is trapped inside it and no laser points at it. Its qubits are electrical circuits, made of capacitors, inductors and wires, cooled to about ten millikelvin, where the metal superconducts and the electrical oscillations of each circuit behave quantum mechanically.
Capacitors and wires alone are not enough. A circuit built only from them is a harmonic oscillator, whose energy levels are evenly spaced like the rungs of a ladder. A microwave pulse that lifts it from the first rung to the second lifts it from the second to the third just as well, so there is no way to keep it inside two levels and use it as a qubit. What breaks the even spacing is a single nonlinear element: the Josephson junction, two superconductors separated by an insulating barrier one or two nanometres thick.
In this article I'll explain what a Josephson junction does, why it behaves like an inductor unlike any coil, and how a junction and a capacitor together make the transmon, the most widely used superconducting qubit today. Each idea comes with a figure you can drag or slide. Let me start with the object itself, a transmon of the kind made in many laboratories. Drag it to change the viewing angle; the first slider takes you from the whole chip down to the junction, and the second pulls the junction apart:
The two large rectangles are the capacitor pads. Each is a few hundred micrometres across, and together they form a capacitor of about 80 femtofarads. In the middle of the gap between them, a thin lead from each pad reaches the junction, which is a thousand times smaller than the pads. The meandering line on the right is a readout resonator, a length of superconducting transmission line whose resonant frequency shifts slightly with the state of the qubit; probing it through the straight feedline beside it is how the qubit is measured.
At the highest zoom you can see how the junction is made. A narrow strip of aluminium is deposited first, and its surface is oxidised in a controlled amount of oxygen, which grows a layer of amorphous aluminium oxide, AlOx. A second strip is then deposited across the first, usually without breaking vacuum, by evaporating from a second angle through the same resist mask. Everything in the rest of this article happens in the small patch where the two strips overlap, across a barrier only a few atoms thick.
Superconductors and phase
In an ordinary metal each conduction electron goes its own way, scattering off impurities and lattice vibrations; that scattering is electrical resistance. Below its critical temperature, about 1.2 kelvin for aluminium, a superconductor is different. Its electrons bind into Cooper pairs, which all condense into one quantum state extending through the whole piece of metal. That state can be described by a single macroscopic wavefunction,
$$\Psi(\mathbf r) = \sqrt{n_s}\;e^{i\theta(\mathbf r)},$$where \(n_s\) is the density of Cooper pairs and \(\theta\) is the phase of the wavefunction. In a piece of superconductor at rest, the phase is the same everywhere. You can picture it as a clock hand that points in the same direction throughout the metal.
Now bring two superconductors close together with a thin insulator between them. Each has its own phase, \(\theta_L\) on the left and \(\theta_R\) on the right. Classically no charge can cross the insulator, but quantum mechanically a Cooper pair can tunnel through it. In 1962 Brian Josephson predicted that this tunnelling carries a current with no voltage across the junction, and that the current depends only on the phase difference \(\varphi = \theta_L - \theta_R\):
$$I = I_c\sin\varphi .$$The critical current \(I_c\) is the largest supercurrent the junction can carry. It grows with the area of the overlap and falls exponentially with the thickness of the barrier. In the figure below each clock is one superconductor, and its hand shows that side's phase. Drag either hand and watch the current between them:
Turning both hands together changes nothing: only the difference between them is physical. The current is largest when the hands are a quarter of a turn apart, and it reverses when the difference passes half a turn.
Now apply a voltage \(V\) across the junction with the last slider. A Cooper pair carries charge \(2e\), where \(e\) is the electron charge, so the pairs on the two sides now differ in energy by \(2eV\), and in quantum mechanics an energy difference makes the phase difference wind steadily in time:
$$\frac{d\varphi}{dt} = \frac{2eV}{\hbar},$$where \(\hbar\) is the reduced Planck constant. A constant voltage therefore produces an alternating current, at the frequency \(f_J = 2eV/h\). The constant \(2e/h\) is 483.6 GHz per millivolt, so even the largest voltage on the slider, 10 µV, makes the current oscillate 4.8 billion times per second. I have let each hand take half of the winding; again, only the difference can be measured.
The Josephson relations
The two equations we have met are the Josephson relations, and together they contain everything about the junction we will need:
$$I = I_c\sin\varphi,\qquad V=\frac{\hbar}{2e}\frac{d\varphi}{dt}.$$Combine them and something familiar appears. If the current through the junction changes, the phase must change, and a changing phase is a voltage. Differentiating the first relation in time and substituting the second gives \(V = L_J\,dI/dt\), the defining equation of an inductor, with inductance
$$L_J = \frac{\hbar}{2e\,I_c\cos\varphi}.$$So a junction is an inductor, but an unusual one: its inductance depends on the phase across it, and therefore on the current through it. A junction with \(I_c = 25\) nA, typical for a transmon, has \(L_J \approx 13\) nH at zero current, and its inductance grows without limit as the current approaches \(I_c\). This is the nonlinearity a qubit needs, and it comes without dissipation. In practice \(I_c\) is checked through the junction's room-temperature resistance \(R_n\): for a tunnel junction \(I_cR_n = \pi\Delta/2e\), with \(\Delta \approx 0.18\) meV the energy gap of aluminium, so 25 nA corresponds to \(R_n \approx 11\) kΩ.
There is a useful mechanical picture of a junction. A real junction also has a capacitance \(C\), from the two films facing each other across the barrier and anything connected in parallel, and some resistance \(R\) to its surroundings. Write Kirchhoff's law for a bias current \(I\) shared between these three elements and you get an equation for the phase that is Newton's law for a ball, of mass proportional to \(C\) and with friction proportional to \(1/R\), rolling on the potential
$$U(\varphi) = -E_J\cos\varphi - \frac{\hbar I}{2e}\,\varphi .$$Here \(E_J = \hbar I_c/2e\) is the Josephson energy, the depth of the cosine wells. The bias current tilts the potential, which is why it is called the tilted washboard. Measure time in units of \(1/\omega_p\), where \(\omega_p = \sqrt{2eI_c/\hbar C}\) is the plasma frequency at which the ball oscillates at the bottom of a well, and current in units of \(I_c\). The equation of motion becomes
$$\ddot\varphi + \gamma\,\dot\varphi + \sin\varphi = i ,$$where the dots are derivatives with respect to this dimensionless time, \(i = I/I_c\) is the bias and \(\gamma = 1/(\omega_p R C)\) is the damping. This is the resistively and capacitively shunted junction model. The figure below integrates it as you watch. The voltage across the junction is proportional to how fast the ball moves, \(V = (\hbar/2e)\,\dot\varphi\).
While \(i < 1\) the washboard still has dips, and the ball settles into one of them. The phase is then constant, so by the second Josephson relation there is no voltage: this is the supercurrent. Raise \(i\) past 1 and the dips disappear; the ball rolls, \(\varphi\) grows steadily, and a voltage appears. The panel on the right records the bias against the average voltage, in units of \(I_cR\). At large bias the ball's speed is limited only by friction, and the junction behaves like the resistor \(R\), along the dashed line.
Now lower the bias with small damping. A rolling ball carries momentum, and it keeps rolling long after the dips have returned. It is caught only when the energy it gains from the tilt over one period of the washboard no longer makes up for what friction takes, which for small damping happens at \(i \approx 4\gamma/\pi\). Between that retrapping current and \(i = 1\) the junction has two possible states, and which one you find depends on its history: the current–voltage curve is hysteretic. Increase \(\gamma\) and the loop closes. Aluminium tunnel junctions at millikelvin temperatures are strongly underdamped and their current–voltage curves are strongly hysteretic, although the resistance of a real tunnel junction is itself far from constant.
From a ball to a wave
A transmon is a junction shunted by a capacitor, the two pads of Figure 1, with no bias current at all. The washboard is flat, \(i = 0\), and the ball sits in a cosine well. Classically it could rest perfectly still at the bottom, or oscillate with any amplitude.
Quantum mechanically it can do neither. The phase \(\varphi\) and the number \(n\) of Cooper pairs that have tunnelled through the junction are conjugate variables, like the position and momentum of a particle, and they obey an uncertainty relation. The energy stored in the capacitor plays the role of kinetic energy and the junction's \(-E_J\cos\varphi\) that of potential energy, and the Hamiltonian is
$$H = 4E_C\,(n-n_g)^2 - E_J\cos\varphi .$$Here \(E_C = e^2/2C\) is the charging energy, the cost of putting one extra electron on the capacitor (a Cooper pair costs \(4E_C\)), and \(n_g\) is an offset charge, in units of \(2e\), induced by stray electric fields around the device. I'll set \(n_g = 0\) for now. With \(E_C/h = 0.25\) GHz, which corresponds to \(C \approx 78\) fF, and the 25 nA junction from above, \(E_J/h \approx 12.4\) GHz, we have \(E_J/E_C \approx 50\): a typical transmon.
The figure solves this Hamiltonian as you move the slider. It writes \(H\) as a matrix in the basis of definite Cooper-pair number, \(n = -25, \dots, 25\), in which the charging term is diagonal and \(\cos\varphi\) moves one pair across the junction. It finds the eigenvalues and eigenvectors of that 51 × 51 matrix and turns each eigenvector back into a wavefunction of the phase, \(\psi(\varphi) = \sum_n c_n e^{in\varphi}\).
Each line is an allowed energy, with its wavefunction drawn on top of it. The dashed ticks on the right show where the levels of a harmonic oscillator with the same curvature at the bottom would be, evenly spaced by \(\hbar\omega_p = \sqrt{8E_JE_C}\). The real levels fall progressively below them, because a cosine well widens faster than a parabola as you go up.
The two lowest levels are the qubit, and the frequency of the transition between them, \(f_{01}\), is the qubit frequency. The next gap, \(f_{12}\), is smaller, and the difference \(\alpha = f_{12} - f_{01}\) is the anharmonicity. For a deep well \(f_{01} \approx \sqrt{8E_JE_C}/h - E_C/h\), and \(\alpha\) approaches \(-E_C/h\). At \(E_J/E_C = 50\) the computed values are \(f_{01} = 4.74\) GHz and \(\alpha = -0.29\) GHz: small compared with the qubit frequency, but enough for a microwave pulse at \(f_{01}\) to leave the 1→2 transition alone, as long as the pulse is not too short.
Slide \(E_J/E_C\) down towards 1 and the picture changes. The levels rise far above the top of the well, the wavefunctions spread over all phases, and each state comes close to a state of definite charge. This is the regime of the Cooper-pair box, the earliest superconducting qubit. Its anharmonicity is large, but as we'll see, its frequency depends strongly on the stray charge \(n_g\). The transmon gives up some anharmonicity in exchange for living at large \(E_J/E_C\).
Talking to the qubit
Restricted to its two lowest levels, the transmon is a qubit. Its state \(c_0|0\rangle + c_1|1\rangle\) can be drawn as a point on a sphere, the Bloch sphere, with \(|0\rangle\) at the north pole, \(|1\rangle\) at the south pole and equal superpositions on the equator, where the longitude is the relative phase of \(c_0\) and \(c_1\).
To control the qubit we send it a microwave pulse at an angular frequency \(\omega_d\) close to its own, \(\omega_q = 2\pi f_{01}\). Seen from a frame that rotates at the drive frequency, and neglecting terms that oscillate at \(2\omega_d\) (the rotating-wave approximation), the Hamiltonian becomes
$$\frac{H}{\hbar} = \frac{\Delta}{2}\,\sigma_z + \frac{\Omega}{2}\,\sigma_x ,$$where \(\sigma_x\) and \(\sigma_z\) are Pauli matrices, \(\Delta = \omega_d - \omega_q\) is the detuning of the drive from the qubit, and \(\Omega\), the Rabi frequency, is proportional to the drive amplitude. In this frame the state does something simple: it precesses about the axis \((\Omega, 0, \Delta)\) at the rate \(\sqrt{\Omega^2+\Delta^2}\). The figure integrates the equations of motion of the Bloch vector \((x, y, z)\),
$$\dot x = -\Delta\,y - \frac{x}{T_2},\qquad \dot y = \Delta\,x - \Omega\,z - \frac{y}{T_2},\qquad \dot z = \Omega\,y - \frac{z-1}{T_1},$$in which the terms with \(T_1\) and \(T_2\) describe relaxation; they are switched off until you tick the box. Press play:
On resonance, \(\Delta = 0\), the state turns about the \(x\) axis, from \(|0\rangle\) through the equator to \(|1\rangle\) and back. These are Rabi oscillations: the probability of finding the qubit in \(|1\rangle\), \(P_1 = (1-z)/2\), swings between 0 and 1. Driving for exactly \(t = \pi/\Omega\), a π pulse, takes \(|0\rangle\) to \(|1\rangle\), the quantum version of a NOT gate; half as long makes an equal superposition. Detune the drive and the rotation axis tilts towards the pole, so the state no longer reaches \(|1\rangle\): the largest \(P_1\) falls to \(\Omega^2/(\Omega^2+\Delta^2)\).
This is where the anharmonicity of Figure 4 earns its keep. A short pulse contains a spread of frequencies, roughly the inverse of its length. If that spread reaches \(f_{12}\), part of the population leaks into \(|2\rangle\), outside the qubit. With \(|\alpha| \approx 0.3\) GHz this keeps single-qubit gates longer than several nanoseconds; in practice they take some tens of nanoseconds, with pulse shapes designed to suppress the leakage.
What goes wrong
Tick the relaxation box in Figure 5 and the trajectory spirals inwards. Two processes are at work. Energy relaxation, with time constant \(T_1\), is the qubit giving its energy to the environment and falling back to \(|0\rangle\). Dephasing scrambles the relative phase between \(|0\rangle\) and \(|1\rangle\) without any exchange of energy, because the qubit frequency fluctuates slightly. The phase memory \(T_2\) combines both, \(1/T_2 = 1/2T_1 + 1/T_\varphi\), where \(T_\varphi\) is the time for pure dephasing; this is why \(T_2\) can never exceed \(2T_1\). In the figure both are a microsecond or two so that you can watch them. The best transmons today keep their energy for hundreds of microseconds, and a few for more than a millisecond.
Where does the energy go? Largely into defects. The amorphous oxide of the junction barrier, and the thin oxides and residues on every surface and interface of the chip, contain microscopic two-level systems, atoms or charges that can tunnel between two configurations, and these absorb energy at microwave frequencies. Because the qubit's electric field is strongest near the surfaces, a few nanometres of material can dominate the loss. Much of the progress in coherence has come from spreading the field out, one reason the pads of Figure 1 are so large, and from cleaner interfaces and better materials. Other channels include quasiparticles (broken Cooper pairs) tunnelling through the junction, radiation, and decay through the readout resonator.
Dephasing comes from anything that moves the qubit frequency: magnetic flux noise in tunable designs, stray photons in the readout resonator, and charge noise. Charge noise is where \(E_J/E_C\) matters. Go back to Figure 4: the last readout under it gives how far \(f_{01}\) moves when the offset charge goes from \(n_g = 0\) to \(\tfrac12\), the largest shift a stray charge can cause. At \(E_J/E_C = 1\) it is almost 0.8 GHz; at 50 it is about 10 kHz. It falls exponentially, roughly as \(e^{-\sqrt{8E_J/E_C}}\), while the anharmonicity falls only slowly towards \(-E_C/h\). That trade is the idea of the transmon.
Finally, the \(\sin\varphi\) we started from is itself an idealisation. It holds for tunnel barriers, in which a very large number of conduction channels each transmit with a small probability. In junctions made of graphene, semiconductor nanowires or other weak links, a few channels can transmit with probability \(\tau\) close to one, and the current–phase relation becomes skewed towards \(\varphi = \pi\). Such a junction still makes a qubit, one whose \(E_J\) can be tuned with a gate voltage, but its anharmonicity is smaller, by a factor of about \(1 - 3\tau/4\) for a single channel.
Further in my notes
- A junction is a nonlinear inductor — the current–phase relation, and how transparency reshapes it.
- Listening to a junction with microwaves — Shapiro steps, and what a half step tells you.
- A qubit is the bottom of a ladder — the levels against offset charge, from charge qubit to transmon.
- What a qubit forgets, and how fast — relaxation, dephasing and coherence budgets.
- The machine around a qubit — the dilution refrigerator, its wiring and the sample box, in 3D.
- Taking a quantum computer apart — one slider, from the closed cryostat down to the atoms of the junction.
- Inside a moiré qubit — a gated moiré device taken apart in 3D, and six device concepts with their verdicts.
- josim-cpr-bridge — carries a junction's current–phase relation from the JoSIM circuit simulator to the qubit spectrum.
- superconducting-qubit-chip-design — a transmon design flow, from layout and readout to gates and a repetition code.
- jjqubit-studio — a schematic-to-solver simulator for superconducting circuits.