betaFeedback

WorksPhysics and Astronomy

Non-ergodic Area Scaling of Andreev Transport in Al/AlOx_x/Al Tunnel Junctions

Abstract

State-of-the-art superconducting qubits rely on Al/AlOx_x/Al Josephson junctions (JJs) as a nonlinear element with an ideal sinusoidal current-phase dependence, with its inductance set by the junction area and oxidation conditions. This idealization holds only if the many conduction channels through the amorphous AlOx_x barrier transmit weakly and almost identically. Here, we measure the second-order subgap Andreev current, which sets the second Josephson harmonic, in Al/AlOx_x/Al JJs. For junction sizes relevant to qubits, we find that this current is dominated by ≲300\lesssim 300 effective channels whose transparencies exceed the ideal tunneling limit (≲10−6\lesssim 10^{-6}) by orders of magnitude. For small and high-aspect-ratio junctions, increasing the junction size does not average out the effect of these high transparency channels but instead adds additional channels of even greater transparency. The trend inverts once both the width and length of the junction exceed a critical length scale, above which the effective transparency decreases with area. This non-ergodic scaling of the Andreev current is reproduced by simulations and attributed to non-self-averaging statistics of the transparency distribution, which arise from variations in oxide thickness. The effect is largely invisible to the normal-state conductance used to target the inductance and reveals unexpected geometric and size-related trade-offs for scaling superconducting qubit systems to large numbers of junctions.

The Josephson junction is the cornerstone of superconducting quantum information processing, providing the nonlinearity essential for quantum operations [1, 2]. In particular, aluminum-oxide (AlOx_x)-based tunnel Josephson junctions (JJs) have enabled superconducting hardware to become a leading platform for fault-tolerant quantum computing [3, 4, 5, 6], and form crucial elements for quantum-limited amplifiers [7], microwave photon detectors [8] and SQUID magnetometers [9]. As quantum-processor architectures scale toward 10510^5 junctions [10, 11, 12] and ground-plane materials improve [13, 14, 15], the junction itself becomes a limiting source of device variability and decoherence [15, 16].

The supercurrent in JJs is carried by NchN_{\mathrm{ch}} conducting channels, each of which forms an Andreev bound state (ABS). Summing over the NchN_{\mathrm{ch}} contributions in the zero-temperature limit kBT≪Δk_{\mathrm{B}}T \ll\Delta appropriate to our measurements, we obtain the junction current-phase relation (CPR):

I(ϕ)=eΔ22ℏsin⁡ϕ∑i=1NchτiEiA(ϕ,τi)=∑m=1∞amsin⁡(mϕ),(1)I(\phi)=\frac{e\Delta^2}{2\hbar}\sin\phi\sum_{i=1}^{N_{\mathrm{ch}}}\frac{\tau_i}{E_i^{\mathrm{A}}(\phi,\tau_i)}=\sum_{m=1}^{\infty}a_m\sin(m\phi), \tag*{(1)}

where ±EiA=±Δ1−τisin⁡2(ϕ/2)\pm E_i^{\mathrm{A}}=\pm\Delta\sqrt{1-\tau_i\sin^2(\phi/2)} are the energies of the ithi^{\mathrm{th}} ABS pair [17, 18], ϕ\phi is the superconducting order parameter phase difference, and τi\tau_i is the ithi^{\mathrm{th}} channel transparency. The final equality is obtained through a Fourier series expansion and emphasizes the presence of higher harmonics within a general JJ. A measure of junction non-ideality is the CPR skew, given by the ratio of the second to first harmonics ∣a2/a1∣\lvert a_2/a_1\rvert, which results from the presence of leaky channels with large transparencies. Non-zero skew causes unpredictable deviations in transmon excited states [16], reduces readout fidelity through measurement-induced transitions [19], and increases dephasing from trapped quasiparticles and critical-current fluctuations [20].

For a junction obeying Eq. (1), expanding to second order in τ\tau and summing over channels gives ∣a2/a1∣=⟨τ2⟩/(8⟨τ⟩)\lvert a_2/a_1\rvert=\langle\tau^2\rangle/(8\langle\tau\rangle). Since the Josephson energy EJmE_{Jm} of a higher harmonic is proportional to am/ma_m/m, a ratio of ∣EJ2/EJ1∣=⟨τ2⟩/(16⟨τ⟩)\lvert E_{J2}/E_{J1}\rvert=\langle\tau^2\rangle/(16\langle\tau\rangle) is expected purely from the CPR skew. Previous work extracted higher harmonic content from the transition spectra of transmon qubits [16, 21]. To disentangle the ∣EJ2/EJ1∣\lvert E_{J2}/E_{J1}\rvert contributions intrinsic to the junction from those of the circuit inductance, the former was assumed to be independent of junction area AjjA_{\mathrm{jj}} [21] and thus distinguishable from the inductive contribution, which is proportional to AjjA_{\mathrm{jj}}. Fundamentally, this assumption relies on the spatial averages of τ\tau and τ2\tau^2 converging to their ensemble values independently of junction area. That is, the ratio ⟨τ2⟩/⟨τ⟩\langle\tau^2\rangle/\langle\tau\rangle is constant and area-independent, and the junction scales ergodically: its size does not influence the nature of first- and second-order Andreev transport. Figure 1 illustrates this concept by comparing the tunneling transparency map of one small junction of area A0A_0 with that of a junction of twice the area, 2A02A_0. With ergodic scaling, the larger junction has effectively the same transparency distribution and reproduces the hot-spot (high-transparency) statistics of the smaller junction; both ⟨τ⟩\langle\tau\rangle and ⟨τ2⟩\langle\tau^2\rangle are essentially unchanged, because the smaller region is statistically representative of the larger area. In the non-ergodic case, rare and extreme high-transparency channels become more common as the area grows. The tunneling transparency statistics are then non-self-averaging [22]: the most transparent channels disproportionately affect ⟨τ2⟩∝EJ2\langle\tau^2\rangle\propto E_{J2}, resulting in a strong area dependence. Conversely, ⟨τ⟩∝Gn\langle\tau\rangle\propto G_n, where GnG_n is the normal-state conductance of the JJ, is essentially unchanged. This implies that the room-temperature conductance measurements used in frequency targeting are insensitive to this area scaling, making it challenging to discriminate from inductance-related effects in transmon experiments.

Figure 1. Ergodic (self-averaging) versus non-ergodic (non-self-averaging) area scaling of the tunneling transparency (model schematic). a, Exploded view of one Al/Al2_2O3_3/Al junction of area Ajj=A0A_{\mathrm{jj}} = A_0 and the transparency map of its amorphous barrier (colour: local transparency τ\tau, white low to red high). b,c, shows how a slice of the sample twice in size would look, under the two hypotheses. Below each map, the histogram of channel transparencies of the A0A_0 region alone (grey, identical in b and c) and of the full 2A02A_0 junction (colour). Self-averaging, (b) implies that any sub-area is statistically representative of the whole so both ⟨τ⟩\langle\tau\rangle and ⟨τ2⟩\langle\tau^2\rangle are unchanged. In contrast, non self averaging sample, (c) the larger area samples even deeper into the right tail of the fat-tailed distribution, and adds channels of increasing transparency. These are the rare events responsible for the non-self-averaging behavior; ⟨τ⟩\langle\tau\rangle barely changes while ⟨τ2⟩\langle\tau^2\rangle increases with area.

Here, we characterize the underlying transparencies and junction scaling by directly probing the second-order Andreev current using DC subgap transport measurements of 134 JJs with varied area (AjjA_{\mathrm{jj}}), oxide thickness (dd), and aspect ratio (w/lw/l, where ww, ll are the junction width and length). The large and diverse sample set allows observation of unexpected scaling of the second-order Andreev current, which depends on both the junction area and aspect ratio. For small junctions and junctions with large aspect ratio, the dependence is consistent with the non-ergodic scaling in Figure 1. We find that an emergent threshold length scale lthl_{\mathrm{th}} exists whereby junctions with both ll, w>lthw > l_{\mathrm{th}} display improved oxide thickness uniformity, reduced effective transparency, and behavior closer to the ideal junction with a sinusoidal CPR. For junction sizes typically relevant to transmon qubits, a small number ≲300\lesssim300 of rare high transparency channels dominate the contribution to ⟨τ2⟩\langle\tau^2\rangle and therefore EJ2E_{J2}. Our results provide an alternative interpretation of recent area-dependent EJ2/EJ1E_{J2}/E_{J1} observations, and demonstrate how junction microscopics and the underlying distribution of transparencies constrain the design of scalable qubit architectures.

The 134 Al/AlOx_x/Al JJs studied here are fabricated following standard shadow-evaporation processes, with bottom electrode width ww and top electrode width ll ranging from 0.1 μ\mum up to 2.0 μ\mum, aspect ratios 0.25<w/l<100.25 < w/l < 10, and oxidation exposures E\mathcal{E}, the product of the static O2_2 pressure and the oxidation time, from 4 to 480 Torr·min. Junction areas Ajj=lw+2ltbA_{\mathrm{jj}} = lw + 2lt_{\mathrm{b}} include the overlap of the top electrode with the two sidewalls of the bottom electrode of thickness tbt_{\mathrm{b}}, while ll and ww are SEM-measured for Square Set A and design values otherwise (Supplementary Section S1). The junctions with w/l≠1w/l \ne1 comprise three different fabrication batches and are referred to as high-aspect-ratio (HAR) Sets A, B, and C. Sets A and B each contain devices with a fixed short dimension (l=200l = 200 or 300 nm), while Set C varies both ll and ww. For square devices, set A was fabricated in a single deposition step while set B underwent multi-step fabrication. DC measurements are taken at 10 mK in a Bluefors LD dilution refrigerator, with RC- and powder-filtered lines (Figure S1).

Figure S1. Schematic of the cryogenic DC measurement circuit. Temperature stages are indicated by horizontal dashed lines (50 K, 4 K, cold plate (CP), and mixing chamber (MXC) at ∼10\sim10 mK, enclosed by the red box). The junction (cross) is measured in a four-probe configuration: the DC bias line (right) and the current-measurement line (left) drive the two ends of the junction wire, and the two inner lines sense the voltage across it. Each line passes through a two-stage RC low-pass filter at MXC with component values RA=50 ΩR_A = 50\ \Omega, RB=100 ΩR_B = 100\ \Omega, RC=220 ΩR_C = 220\ \Omega, C1=47C_1 = 47 nF, C2=33C_2 = 33 nF, C3=15C_3 = 15 nF, followed by a Basel MFT silver-powder filter. Room-temperature electronics are a Basel LNHR DAC (DC voltage source), an SRS SR560 voltage preamplifier and an SRS SR570 current preamplifier, each read out by a Keithley 7510 digital multimeter (V). All DC wiring uses shielded twisted-pair cable. The sample package (gray) has a copper base and an aluminum light-tight lid.

The resulting voltage-driven idci_{\mathrm{dc}}–vdcv_{\mathrm{dc}} measurements (Figure 2a) show behavior typical of tunnel JJs, with a supercurrent peak at vdc=0v_{\mathrm{dc}} = 0 and the onset of a quasiparticle current for vdc≥2Δ/ev_{\mathrm{dc}} \ge2\Delta/e, where Δ=180\Delta= 180 μ\mueV is the Al order parameter. Figure 3a,Figure 3d plot GnG_{\mathrm{n}} as a function of AjjA_{\mathrm{jj}} for the rectangular junction sets and for square junctions at four oxidation conditions. In every set the normal-state conductance GnG_{\mathrm{n}} increases linearly with AjjA_{\mathrm{jj}}, and it decreases with oxidation exposure according to a power-law dependence (Figure S3). When plotted on a logarithmic scale (Figure 2b), a clear idci_{\mathrm{dc}} step becomes visible near vdc=Δ/ev_{\mathrm{dc}} = \Delta/e. The current steps I1I_1 and I2I_2 (dashed lines in Figure 2b) result from sideband scattering of quasiparticles mediated by ABSs, a process which is represented schematically in Figure 2c. For first order scattering n=1n = 1, a quasiparticle is excited by an applied voltage vdc≥2Δ/ev_{\mathrm{dc}} \ge2\Delta/e, crossing the insulating barrier with transmission probability τi\tau_i (i∈{1,2,…,Nch}i \in\{1, 2, \ldots, N_{\mathrm{ch}}\} conduction channels). For n=2n = 2, a right-moving quasiparticle is excited by Δ/e≤vdc<2Δ/e\Delta/e \le v_{\mathrm{dc}} < 2\Delta/e, scattering into a sideband channel on the opposite electrode with probability τi\tau_i while, coherently, a hole is excited and scattered with probability τi\tau_i into the n=2n = 2 sideband above the gap edge, leaving one cooper pair on the right electrode. Thus, the nn coherent excitations on the ladder result in current step In∝τinI_n \propto\tau_i^n [17, 18, 23, 24, 25, 26, 27, 28]. In practice, only the first- (I1I_1) and second-order (I2I_2) current steps are visible in our tunneling junctions due to the relatively small step size.

Figure 2. DC transport characteristics of Al/AlOx_x/Al Josephson junctions. a, Signed idci_{\mathrm{dc}}–vdcv_{\mathrm{dc}} characteristics for three junctions with Ajj=0.054A_{\mathrm{jj}} = 0.054, 0.0800.080, and 0.153 μm20.153\ \mu\mathrm{m}^{2} at a fixed oxidation exposure of 120 Torr·min. Dotted vertical lines mark vdc=±2Δ/ev_{\mathrm{dc}} = \pm2\Delta/e; the dashed guide indicates the normal-state conductance GnG_{\mathrm{n}}. b, ∣idc∣|i_{\mathrm{dc}}| versus vdcv_{\mathrm{dc}} on a logarithmic scale for the same three junctions as in a. Dashed lines denote the extraction of the step amplitudes I1I_{1} and I2I_{2}. c, Schematic of the two lowest-order Andreev tunneling processes. For eV>2ΔeV > 2\Delta a single quasiparticle excitation and barrier traversal occurs, giving I1∝⟨τ⟩I_{1} \propto\langle\tau\rangle. For Δ<eV<2Δ\Delta< eV < 2\Delta transport requires coherent crossings of a hole-like (left-traveling) and electron-like quasiparticle (right-traveling), resulting in an additional cooper pair (2e2e) on the right electrode. This process has amplitude I2∝⟨τ2⟩I_{2} \propto\langle\tau^{2} \rangle, with each coherent channel crossing ii contributing a factor τi\tau_{i}.

Figure 3. Normal-state conductance, effective transparency (τe=I2/I1\tau_{\mathrm{e}} = I_{2}/I_{1}) and effective channel number (NeN_{\mathrm{e}}) versus junction area for the high-aspect-ratio (HAR) sets (a-c) and Square Set A (d-f). a,d, GnG_{\mathrm{n}} versus AjjA_{\mathrm{jj}}. HAR Sets A and B have a short dimension of either l=200l = 200 nm or l=300l = 300 nm and correspond to different fabrication batches while set C varies both ll and ww. Square Set A comprises four oxidation exposures (15, 30, 120, and 240 Torr·min). Dashed lines are one-parameter fits Gn=CAjjG_{\mathrm{n}} = C A_{\mathrm{jj}} (∝Ajj\propto A_{\mathrm{jj}}) to each series. Insets: false-color SEM images of a representative HAR and square junction with the electrode widths ll, ww marked (scale bars 500 nm). b,e, τe\tau_{\mathrm{e}} versus AjjA_{\mathrm{jj}}. τe\tau_{\mathrm{e}} was previously assumed to be constant with junction area [21]. Here it increases with area in every HAR set, whereas the square junctions are non-monotonic with peak near Ajj≈0.15A_{\mathrm{jj}} \approx0.15 µm2^{2}. Large junctions show decreasing τe\tau_{\mathrm{e}} as ∝Ajj−1\propto A_{\mathrm{jj}}^{-1} (dashed guide). c,f, Point-wise NeN_{\mathrm{e}} (Eq. (3)) versus AjjA_{\mathrm{jj}}. The HAR samples show a strongly sublinear NeN_{\mathrm{e}}, while square devices have constant Ne≈50N_{\mathrm{e}} \approx50 for Ajj<0.15A_{\mathrm{jj}} < 0.15 µm2^{2} and then Ne=(3.3±0.2)×103Ajj2.07±0.07N_{\mathrm{e}} = (3.3 \pm0.2) \times10^{3} A_{\mathrm{jj}}^{2.07 \pm0.07} (AjjA_{\mathrm{jj}} in µm2^{2}) for larger junctions (∝Ajj2\propto A_{\mathrm{jj}}^{2}, dashed guide). Insets: τe\tau_{\mathrm{e}} versus GnG_{\mathrm{n}} for junctions with Gn<100G_{\mathrm{n}} < 100 µS. Independent of oxidation condition, the points are collinear with inverse slope corresponding to a constant Ne=274±5N_{\mathrm{e}} = 274 \pm5 (fit to HAR Sets A and B) and Ne=47±4N_{\mathrm{e}} = 47 \pm4 (Square Set A, solid).

Figure S3. Effective transparency (τe=I2/I1\tau_e = I_2/I_1) and effective channel number (NeN_e) versus oxidation exposure. a, τe\tau_e versus oxidation exposure EE on log–log axes for Square Set A junction areas from 0.054 μm2\mu\mathrm{m}^{2} to 1.71 μm2\mu\mathrm{m}^{2}. b, Channel density Ne/AjjN_e/A_{jj} versus oxidation exposure on log–log axes. No consistent trend is observed. Variation for the smallest area (Ajj=0.054A_{jj} = 0.054 μm2\mu\mathrm{m}^{2}) is attributed to large propagated error in I2I_2 due to its small amplitude.

Having established the transport mechanism, the two primary observables reflecting the behavior of ⟨τ⟩\langle\tau\rangle, ⟨τ2⟩\langle\tau^2\rangle are the effective number of channels NeN_{\mathrm{e}} and the effective transparency τe\tau_{\mathrm{e}}, which are defined as:

τe≡I2I1=⟨τ2⟩⟨τ⟩(2)\tau_{\mathrm{e}} \equiv\frac{I_2}{I_1} = \frac{\langle\tau^2\rangle}{\langle\tau\rangle} \tag*{(2)}
Ne=hGn2e2I1I2.(3)N_{\mathrm{e}} = \frac{hG_{\mathrm{n}}}{2e^2}\frac{I_1}{I_2}. \tag*{(3)}

The normal-state conductance GnG_{\mathrm{n}} is extracted from a linear fit to the ohmic branch well above 2Δ/e2\Delta/e, and I1I_1, I2I_2 are extracted as in Figure 2b. The values of τe\tau_{\mathrm{e}} and NeN_{\mathrm{e}} from experimental idci_{\mathrm{dc}}-vdcv_{\mathrm{dc}} data are plotted in Figure 3, with additional datasets available in the Supplementary Information (§ 2.4). Due to the starkly differing area dependencies, we segregate the data from rectangular junctions with w/l≠1w/l \ne1 (Figure 3a-Figure 3c) and square junctions with w/l=1w/l = 1 (Figure 3d-Figure 3f). Baseline GnG_{\mathrm{n}} vs AjjA_{\mathrm{jj}} curves in Figure 3a,Figure 3c, show typical linear behavior and it is only by looking at the subgap that the scaling behavior begins to differ. Focusing on HAR in Figure 3b, we first observe that the τe\tau_{\mathrm{e}} data from HAR C are co-linear, even though the set spans junctions with a range of top- and bottom-electrode widths and lengths. More strikingly, however, is that τe\tau_{\mathrm{e}} depends on AjjA_{\mathrm{jj}}–increasing by an order of magnitude across the area range studied in all HAR datasets. The square junctions in Figure 3e also show an unexpected, area-dependent τe\tau_{\mathrm{e}}, but here the trend is non-monotonic. An emergent threshold length scale lthl_{\mathrm{th}} appears such that for l,w>lthl,w > l_{\mathrm{th}}, τe\tau_{\mathrm{e}} peaks and begins to decrease as Ajj−1A_{\mathrm{jj}}^{-1}. We observe that lthl_{\mathrm{th}} is process

dependent, such that Square Set A inverts at Ajj≈0.14 μm2A_{\mathrm{jj}} \approx0.14\ \mu\mathrm{m}^{2} (lth≈0.34 μml_{\mathrm{th}} \approx0.34\ \mu\mathrm{m}), while square junctions on a Set B inverts near Ajj≈0.4A_{\mathrm{jj}} \approx0.4–0.5 μm20.5\ \mu\mathrm{m}^{2} (lth≈0.67 μml_{\mathrm{th}} \approx0.67\ \mu\mathrm{m}) (Figure S2). Within the sets of devices, the value of lthl_{\mathrm{th}} is also independent of the oxidation exposure.

Figure S2. Supplemental fit example from an independent fabrication batch (Square Set B), demonstrating that the non-monotonic τe\tau_e–AjjA_{jj} trend reproduces across processes. Fits (solid lines) use θI=θII=1.15\theta_I = \theta_{II} = 1.15 eV. a, τe\tau_e versus AjjA_{jj} for oxidation exposures of 4, 87, and 480 Torr·min. The single smallest-area 4 Torr·min die (Ajj=0.018A_{jj} = 0.018 μm2\mu\mathrm{m}^{2}, drawn) is excluded from the joint fit as an outlier. Inset: I2/AjjI_2/A_{jj} versus AjjA_{jj}, showing the saturation of I2I_2 above Ajj=AthA_{jj} = A_{\mathrm{th}}. b, NeN_e versus AjjA_{jj}. Inset: τe\tau_e versus GnG_n for the small-junction regime (87 and 480 Torr·min dies), whose inverse slope τe=hNe2e2Gn\tau_e = \frac{h}{N_e 2e^2 G_n} gives a constant Ne=349±27N_e = 349 \pm27. c, Extracted Region-I disorder σI\sigma_I (bars) and mean Region-I and Region-II oxide thicknesses dId_I, dIId_{II} (markers) for the Square Set A exposure series, HAR Sets A, B, and C, and Square Set B. The HAR Sets are fit as pure Region I, so no dIId_{II} is shown for them.

The source of this area dependence can be traced to the dependence of I2I_{2} on junction area, since all junction sets have Gn∝I1∝AjjG_{\mathrm{n}} \propto I_{1} \propto A_{\mathrm{jj}} (Figure 3a,Figure 3d). The inversion of τe\tau_{\mathrm{e}} in square junctions results from I2I_{2} saturating at Ajj=AthA_{\mathrm{jj}} = A_{\mathrm{th}}, the area of a junction of side lthl_{\mathrm{th}} (Eq. (4)), to a constant value as AjjA_{\mathrm{jj}} increases (Figure S4) while I1I_{1} continues to increase thus lowering τe\tau_{\mathrm{e}}. We also note that the size of the edge-overlap between the top and bottom electrodes (“rising edge”) does not affect τe\tau_{\mathrm{e}} appreciably, since HAR A, B, and C all display a similar trend.

Figure S4. Subgap current I2I_2 versus junction area for the rectangular and square junction families. a, Raw I2I_2 versus AjjA_{\mathrm{jj}}, one marker per junction; thin lines connect per-area medians. b, Subgap current density I2/AjjI_2/A_{\mathrm{jj}}, with each series normalized to its value at the smallest measured area (within HAR Sets A and B, the l=300l = 300 nm sweep shares its own set’s l=200l = 200 nm anchor, so each set reads as one band); the dotted line marks the reference level of 1. Compared with the near-linear I1∝AjjI_1 \propto A_{\mathrm{jj}} (see GnG_n vs AjjA_{\mathrm{jj}} in Figure 3a,Figure 3d), I2/AjjI_2/A_{\mathrm{jj}} varies by more than an order of magnitude and dominantly contributes to τe(Ajj)\tau_e(A_{\mathrm{jj}}).

Next, Using Eq. (3), NeN_{\mathrm{e}} is plotted as a function of AjjA_{\mathrm{jj}} in Figure 3c,Figure 3f. NeN_{\mathrm{e}} is flat or increases sub-linearly for Ajj<AthA_{\mathrm{jj}} < A_{\mathrm{th}}, and is independent of oxidation exposure for devices within a given fabrication batch. Fitting log⁡(Ne)\log(N_{\mathrm{e}}) versus log⁡(Ajj)\log(A_{\mathrm{jj}}) for square junctions with Ajj>AthA_{\mathrm{jj}} > A_{\mathrm{th}}, we find Ne=(3.3±0.2)×103Ajj2.07±0.07N_{\mathrm{e}} = (3.3 \pm0.2) \times10^{3} A_{\mathrm{jj}}^{2.07\pm0.07} with AjjA_{\mathrm{jj}} in μm2\mu\mathrm{m}^{2}. For comparison, in a homogeneous tunnel barrier NeN_{\mathrm{e}} should scale linearly with area, up to Ne=NchN_{\mathrm{e}} = N_{\mathrm{ch}} for an ideal junction. Zooming in on the small area region, we plot τe\tau_{\mathrm{e}} directly as a function of GnG_{\mathrm{n}} (insets of Figure 3c,Figure 3f). Data points from the small junctions (l,w<lthl,w < l_{\mathrm{th}}) fabricated in the same lot collapse to a common exposure-invariant line when Gn<100 μSG_{\mathrm{n}} < 100\ \mu\mathrm{S}. The inverse slope gives constant Ne=47±4N_{\mathrm{e}} = 47 \pm4 and Ne=274±5N_{\mathrm{e}} = 274 \pm5 for the Square Set A and HAR junctions, respectively. This conductance range corresponds to EJ/h<14 GHzE_{\mathrm{J}}/h < 14\ \mathrm{GHz} (considering Δ=180 μeV\Delta= 180\ \mu\mathrm{eV}), which overlaps with typical EJE_{\mathrm{J}} in transmon circuits. A similar analysis was used previously in small break junctions with few channels, but colinearity is unexpected in planar SIS tunnel junctions with τe≪1\tau_{\mathrm{e}} \ll1 [25].

The experimental data above indicates that small square devices with l,w<lthl,w < l_{\mathrm{th}} and all HAR devices have increasing τe\tau_{\mathrm{e}} and effectively constant NeN_{\mathrm{e}}. This behavior is consistent with increased area resulting in the addition of channels with increasing transparency that dominate the transport. Conversely, junctions with both l,w>lthl,w > l_{\mathrm{th}} have τe∝Ajj−1\tau_{\mathrm{e}} \propto A_{\mathrm{jj}}^{-1} and Ne∝Ajj2N_{\mathrm{e}} \propto A_{\mathrm{jj}}^{2}. Since I2I_{2} is constant or slowly varying for l,w>lthl,w > l_{\mathrm{th}} (Figure S4), we infer that only low transparency channels are added with increased area.

Based on these trends, we introduce an analysis framework which utilizes the area-dependence to extract and characterize the underlying channel transparency distribution ρ(τ)\rho(\tau). The data suggests two distinct disorder regions in the junctions, denoted Region I and Region II and shown schematically in Figure 4a. Region I corresponds to the small square devices and HAR junctions. Its oxide is assumed to have greater spatial variability, with mean thickness dId_{\mathrm{I}} and thickness standard deviation σI\sigma_{\mathrm{I}}. Region II emerges in larger square junctions and has more uniform thickness, resembling an ideal oxide. The oxide thickness distribution in this region has mean dIId_{\mathrm{II}} and standard deviation σII\sigma_{\mathrm{II}}, with σII<σI\sigma_{\mathrm{II}} < \sigma_{\mathrm{I}}. The observed constant I2I_{2} (Figure S4), τe∝Ajj−1\tau_{\mathrm{e}} \propto A_{\mathrm{jj}}^{-1} and Ne∝Ajj2N_{\mathrm{e}} \propto A_{\mathrm{jj}}^{2} therefore reflect the crossover from Region I to Region II. The disordered Region I remains fixed in size, while only low-disorder Region II is added for l,w>lthl,w > l_{\mathrm{th}} at a rate proportional to AjjA_{\mathrm{jj}}. Algebraically, the areas of Region I (Region II), AIA_{\mathrm{I}} (AIIA_{\mathrm{II}}), are then defined as

Figure 4. Two-region disorder model and fits to the experimental data. a, Schematic. Region I is a disordered oxide with thickness distribution ρI(d)\rho_{\mathrm{I}}(d) (large σ\sigma) occupying AI=min⁡(Ajj,Ath)A_{\mathrm{I}}=\min(A_{\mathrm{jj}},A_{\mathrm{th}}): the whole junction when min⁡(l,w)<lth\min(l,w)<l_{\mathrm{th}} and a fixed area AthA_{\mathrm{th}} otherwise (Eq. (4)). Region II is a more uniform oxide, ρII(d)\rho_{\mathrm{II}}(d) (small σ\sigma), covering the remaining area AII=Ajj−AIA_{\mathrm{II}}=A_{\mathrm{jj}}-A_{\mathrm{I}}. Arrows indicate the transparency of the channels within each region, larger arrows corresponding to higher transparencies. b,c, Simulated τe\tau_{\mathrm{e}} and NeN_{\mathrm{e}} versus channel number Nch=nchAjjN_{\mathrm{ch}}=n_{\mathrm{ch}}A_{\mathrm{jj}} for a junction composed only of Region I (AI=AjjA_{\mathrm{I}}=A_{\mathrm{jj}}), for σI=0.10\sigma_{\mathrm{I}}=0.10, 0.14, 0.18, 0.22, 0.26, and 0.30 nm (color bar) at fixed dI=1.40d_{\mathrm{I}}=1.40 nm. For small σI\sigma_{\mathrm{I}}, τe\tau_{\mathrm{e}} is size-independent and Ne∝NchN_{\mathrm{e}}\propto N_{\mathrm{ch}} (gray dashed line, Ne=NchN_{\mathrm{e}}=N_{\mathrm{ch}}); increasing σI\sigma_{\mathrm{I}} makes both observables strongly size-dependent. Curves are terminated where the thinnest channel of the distribution reaches τ=1\tau=1. The shaded band marks Nch=nchlth2N_{\mathrm{ch}}=n_{\mathrm{ch}}l_{\mathrm{th}}^2 for lth=0.1l_{\mathrm{th}}=0.1–1 μ\mum, the range swept in d,e. d,e, Region I →\rightarrow Region II transition, sweeping lthl_{\mathrm{th}} from 0.1 to 1 μ\mum (color bar, shaded band) at fixed ρI\rho_{\mathrm{I}} with (dI,σI)=(1.40,0.25)(d_{\mathrm{I}},\sigma_{\mathrm{I}})=(1.40,0.25) nm and ρII\rho_{\mathrm{II}} with (dII,σII)=(1.05,0.10)(d_{\mathrm{II}},\sigma_{\mathrm{II}})=(1.05,0.10) nm. Open circles mark Nch=nchlth2N_{\mathrm{ch}}=n_{\mathrm{ch}}l_{\mathrm{th}}^2, where the Region-I channel count saturates (the simulation carries no sidewall term); dashed lines are the pure Region-I (AI=AjjA_{\mathrm{I}}=A_{\mathrm{jj}}, blue) and pure Region-II (AII=AjjA_{\mathrm{II}}=A_{\mathrm{jj}}, orange) limits. f,g, Experimental τe\tau_{\mathrm{e}} and NeN_{\mathrm{e}} versus GnG_{\mathrm{n}} (markers) with model fits (lines) for the four Square Set A exposure series and HAR Sets A and B (100 Torr·min). The extracted parameters are listed in Table 1. All simulations use θ=1.15\theta=1.15 eV, nch=106n_{\mathrm{ch}}=10^6 μm−2\mu\mathrm{m}^{-2}, and deterministic order statistics.

AI={Ajj,min⁡(l,w)<lthAth=lth2+2lthtb,min⁡(l,w)≥lthAII=Ajj−AI,(4)\begin{aligned} A_{\mathrm{I}} &= \begin{cases} A_{\mathrm{jj}}, & \min(l,w)<l_{\mathrm{th}}\\ A_{\mathrm{th}}=l_{\mathrm{th}}^2+2l_{\mathrm{th}}t_{\mathrm{b}}, & \min(l,w)\ge l_{\mathrm{th}} \end{cases} \qquad A_{\mathrm{II}}=A_{\mathrm{jj}}-A_{\mathrm{I}}, \tag*{(4)} \end{aligned}

where AthA_{\mathrm{th}} is the area, sidewalls included, of a square junction of side lthl_{\mathrm{th}}. We model the tunneling behavior in Regions I and II assuming spatially inhomogeneous direct tunneling across a trapezoidal energy barrier shape, hence applying the Simmons model at vdc≈0v_{\mathrm{dc}}\approx0 [29, 30]. The low-voltage transparency at zero temperature is given by

τ(d)=Cde−2κd,(5)\tau(d)=\frac{C}{d}e^{-2\kappa d}, \tag*{(5)}

where dd is the barrier thickness, κ≡2meθ/ℏ=5.12 nm−1eV−1/2θ\kappa\equiv\sqrt{2m_{\mathrm{e}}\theta}/\hbar=5.12\ \mathrm{nm}^{-1}\mathrm{eV}^{-1/2}\sqrt{\theta}, mem_{\mathrm{e}} is the electron mass, θ\theta is the potential energy barrier height, C=2meθ/(2h nch)C=\sqrt{2m_{\mathrm{e}}\theta}/(2h\,n_{\mathrm{ch}}), and nch=106 μm−2n_{\mathrm{ch}}=10^{6}\ \mu\mathrm{m}^{-2} is the areal channel density assumed for this simulation. The number of simulated conducting channels is then given as Nch=nchAjjN_{\mathrm{ch}}=n_{\mathrm{ch}}A_{\mathrm{jj}}. In this model, the tunneling transmission e−2κd/de^{-2\kappa d}/d acts as a sieve whereby the channels located in regions with the thinnest barrier have an over-sized influence on ⟨τ⟩\langle\tau\rangle, compounded at second order in I2∝⟨τ2⟩I_{2}\propto\langle\tau^{2}\rangle. Choosing an appropriate thickness distribution ρ(d)\rho(d), is therefore critical to the model accuracy, and we show in § 2.5 that a normal ρ(d)\rho(d) more accurately reproduces the observed experimental results. Refer to § 2.5 for comparisons of functional forms of ρ(d)\rho(d) and any details of the simulations and fitting methodology.

To connect the model to the observed trends, we begin with simulations of NeN_{\mathrm{e}} and τe\tau_{\mathrm{e}} for representative parameters (Figure 4b-Figure 4e). For the single-region case (HAR devices and small-area square junctions, which contain only Region I), we fix dI=1.40 nmd_{\mathrm{I}}=1.40\ \mathrm{nm} and θ=1.15 eV\theta=1.15\ \mathrm{eV} and vary σI\sigma_{\mathrm{I}}. Within a proper parameter range, an area dependence in τe\tau_{\mathrm{e}} is reproduced for κσ≳0.8\kappa\sigma\gtrsim0.8 (and strongly so for κσ≳1\kappa\sigma\gtrsim1), which corresponds to σ≳0.15 nm\sigma\gtrsim0.15\ \mathrm{nm} (≳0.18 nm\gtrsim0.18\ \mathrm{nm}) at θ=1.15 eV\theta=1.15\ \mathrm{eV} in Figure 4b,Figure 4c. Likewise, NeN_{\mathrm{e}} is linear with NchN_{\mathrm{ch}}, but flattens out as σI\sigma_{\mathrm{I}} is increased. Next, we simulate the behavior of large square junctions by adding a Region II with much lower disorder. Extending the results from Figure 4b,Figure 4c, we fix parameters dI=1.40 nmd_{\mathrm{I}}=1.40\ \mathrm{nm}, σI=0.25 nm\sigma_{\mathrm{I}}=0.25\ \mathrm{nm}, dII=1.05 nmd_{\mathrm{II}}=1.05\ \mathrm{nm} and σII=0.10 nm\sigma_{\mathrm{II}}=0.10\ \mathrm{nm}. The threshold length lthl_{\mathrm{th}} is used as the floating parameter with results shown in Figure 4d,Figure 4e. Both τe,Ne\tau_{\mathrm{e}},N_{\mathrm{e}} transition from the pure Region-I curve (Ajj=AIA_{\mathrm{jj}}=A_{\mathrm{I}}) toward the pure Region-II curve (Ajj=AIIA_{\mathrm{jj}}=A_{\mathrm{II}}). In Region II, where σII<0.2 nm\sigma_{\mathrm{II}}<0.2\ \mathrm{nm}, τe\tau_{\mathrm{e}} decreases and NeN_{\mathrm{e}} eventually becomes linear in NchN_{\mathrm{ch}} similar to the experimental data. Interestingly, this area dependence arises intrinsically from the sampling of the transparency distribution ρ(τ)\rho(\tau). Because ρ(τ)\rho(\tau) is fat-tailed, i.e., its right tail does not quickly decay to zero as τ\tau increases, sampling more channels (or, equivalently, increasing the junction area) preferably picks up high-τ\tau outliers, producing a scaling similar to that shown in Figure 1c. τe\tau_{\mathrm{e}} is therefore governed by non-self-averaging behavior: its expectation value does not reflect the bulk of the channels, but is instead determined by rare, high-transparency outliers.

Assuming HAR junctions are pure Region I, experimental data is fit using a cost-minimization algorithm as described in the Supplementary Information, § 2.6. For square junctions, the two smallest areas are assumed to be Region-I dominant (for the 240 Torr-min series the second- and third-smallest, the smallest being an outlier), thus parameters (dI,σI)(d_{\mathrm{I}},\sigma_{\mathrm{I}}) are first fit per oxidation exposure using data from just these two areas. The first-stage result then seeds a joint fit over the entire area range, which determines lthl_{\mathrm{th}} and dIId_{\mathrm{II}}. Since the measured τe\tau_{\mathrm{e}} keeps decreasing and does not saturate at large AjjA_{\mathrm{jj}} (Figure 4f), σII\sigma_{\mathrm{II}} can only be bounded from above, σII≲0.14\sigma_{\mathrm{II}} \lesssim0.14 nm. The bound is set by the Region-II τe\tau_{\mathrm{e}} versus AjjA_{\mathrm{jj}} curves, which must satisfy τe(Ajj)>τe(AII)\tau_{\mathrm{e}}(A_{\mathrm{jj}}) > \tau_{\mathrm{e}}(A_{\mathrm{II}}) at every area. All fits below this bound are equivalent (Figure S9), therefore σII\sigma_{\mathrm{II}} is fixed at 0.10 nm during the optimization. The fit results are shown in Figure 4f,Figure 4g and the extracted parameters are listed in Table 1. The model accurately reproduces the trends in both τe\tau_{\mathrm{e}}, NeN_{\mathrm{e}} as function of AjjA_{\mathrm{jj}}, GnG_{\mathrm{n}}, including the sharp change of trends at lthl_{\mathrm{th}}.

Table 1

DeviceE\mathcal{E} (Torr·min)RnAjjR_{\mathrm{n}}A_{\mathrm{jj}} (Ω\Omega µm2^{2})dId_{\mathrm{I}} (nm)σI\sigma_{\mathrm{I}} (nm)σI/dI\sigma_{\mathrm{I}}/d_{\mathrm{I}}dIId_{\mathrm{II}} (nm)
Square A156761.4110.2800.1990.974
Square A309181.4440.2840.1970.993
Square A12019061.4930.2830.1891.070
Square A24030901.5450.2900.1881.091
HAR A10014951.3060.2250.173—
HAR B10013331.2920.2220.172—
HAR C (ll, ww var.)9612431.3880.2610.188—

Table 1. Region-I and Region-II fit parameters at θ=1.15\theta= 1.15 eV on the total junction area Ajj=lw+2ltbA_{\mathrm{jj}} = lw + 2lt_{\mathrm{b}}. E\mathcal{E} is the oxidation exposure (static O2_2 pressure × oxidation time); RnAjjR_{\mathrm{n}}A_{\mathrm{jj}} is the median normal-state resistance·area product of the junctions in each series (Supplementary Table S1). The joint Square Set A fit gives lth=340l_{\mathrm{th}} = 340 nm (Ath=0.141A_{\mathrm{th}} = 0.141 µm2^{2}); σII=0.10\sigma_{\mathrm{II}} = 0.10 nm.

Figure S9. Bounding the Region-II roughness σII\sigma_{\mathrm{II}}: the two-region joint fit for the 240 Torr·min Square Set A condition (16 junctions over six areas; open circles are the per-area medians the cost is scored on, small dots individual junctions), re-solved with σII\sigma_{\mathrm{II}} pinned at twelve values from 0.005 to 0.30 nm (colorbar), refitting lthl_{\mathrm{th}} and dII(E)d_{\mathrm{II}}(E) at each. a, τe\tau_e versus AjjA_{\mathrm{jj}}. Solid curves lie inside the 1σ1\sigma cost interval (σII≤0.14\sigma_{\mathrm{II}} \le0.14 nm) and form a single indistinguishable bundle: a 28× range of σII\sigma_{\mathrm{II}}, compensated by ≲60\lesssim60 pm of dIId_{\mathrm{II}}, moves τe\tau_e by only 6%. Dashed curves are excluded by the data. The vertical line marks AthA_{\mathrm{th}}; below it the junction is pure Region I and σII\sigma_{\mathrm{II}} has no channels to describe. b, NeN_e versus AjjA_{\mathrm{jj}}, same encoding. Inset: the fit cost versus σII\sigma_{\mathrm{II}}, flat to 1% over 0.005–0.14 nm and then rising steeply (3× at 0.20 nm, 60× at 0.30 nm, marked ×); the dotted line is the 1σ1\sigma threshold. The transport data therefore set only an upper bound on σII\sigma_{\mathrm{II}}; the convention σII=0.10\sigma_{\mathrm{II}} = 0.10 nm (thick curve; boxed value on the colorbar) lies inside the allowed plateau.

The fitted parameters in Table 1 can then be used to generate the corresponding transparency distribution for a typical junction of a desired area. As an example, we choose the Square Set A 30 Torr·min parameters (Table 1) to calculate representative histograms of ρI(τ)\rho_{\mathrm{I}}(\tau) (Figure 5a). These histograms are used to estimate I1I_{1} and I2I_{2} for square junctions across the range 0.01≤Ajj≤2.00.01 \leq A_{\mathrm{jj}} \leq2.0 µm2^{2}. In large junctions, although the total number of channels in Region II far exceeds that in Region I, all 50 of the most transparent channels still lie in Region I. Figure 5b quantifies the fractional contribution of the top 10, 50, and 1000 channels to I1I_{1} and I2I_{2}. Surprisingly, the top 10 contribute ≳90%\gtrsim90\% of I2I_{2} at every area. Even in the largest devices, where Region II supplies ∼86%\sim86\% of the channels and nearly all of I1I_{1}, I2I_{2} remains carried by a small number of Region-I channels. This dominance of I2I_{2} by so few channels is consistent with the small, slowly varying NeN_{\mathrm{e}}. For a more direct comparison with transmon results, we plot the experimental τe/16=EJ2/EJ1\tau_{\mathrm{e}}/16 = E_{\mathrm{J2}}/E_{\mathrm{J1}} data at EJ/h=8E_{\mathrm{J}}/h = 8 GHz and overlay contours of the same EJ2/EJ1E_{\mathrm{J2}}/E_{\mathrm{J1}} ratio in Figure 5c. Each fitted curve reduces to a single point in this plane, with all ∣EJ2/EJ1∣\lvert E_{\mathrm{J2}}/E_{\mathrm{J1}}\rvert values falling between 10−410^{-4} and 2×10−32 \times10^{-3}. These values are of similar magnitude to those in earlier work, but point to an alternative interpretation [16, 21]. Kim et al. carefully attributed the harmonics primarily to parasitic series inductance, under the assumption of a fixed, area-independent junction contribution (τe=constant\tau_{\mathrm{e}} = \mathrm{constant}). Our methodology shows that the HAR geometries used in that study represent a special case in which the linear trend in ∣EJ2/EJ1∣\lvert E_{\mathrm{J2}}/E_{\mathrm{J1}}\rvert can equally arise from non-ergodic scaling within the junction itself.

Figure 5. Simulated channel statistics and implications for qubit junctions. a, Histograms of channel transparencies τi\tau_i simulated with the fitted Square Set A 30 Torr·min parameters (Table 1) for devices of size Ajj=0.033A_{\mathrm{jj}} = 0.033 (top) and 1.0 μm21.0\ \mu\mathrm{m}^{2} (bottom). For both areas, all 50 of the most transparent channels (dark navy) are found in Region I, despite the large device being ∼86%\sim86\% Region II channels. b, Fraction of I1I_1 and I2I_2 carried by the top n=10n = 10, 50, and 1000 most-transparent channels versus AjjA_{\mathrm{jj}} for the 30 Torr·min simulation. The top 10 channels already carry ≳90%\gtrsim90\% of I2I_2 at every area, while their share of I1I_1 decreases rapidly with area. c, Josephson-harmonics optimization map: contours of ∣EJ2/EJ1∣=τe/16\lvert E_{J2}/E_{J1}\rvert= \tau_e/16 in the (σ,d0)(\sigma,d_0) plane at θ=1.15 eV\theta= 1.15\ \mathrm{eV}, evaluated at fixed EJ/h=8 GHzE_J/h = 8\ \mathrm{GHz}. Markers show the fitted Region-I (dI,σI)(d_I,\sigma_I) parameters of each device family in this work (Square Set A and B, and HAR Sets A, B, and C). The right bar places each family on the resulting ∣EJ2/EJ1∣\lvert E_{J2}/E_{J1}\rvert scale.

The area dependence observed provides a design handle for superconducting circuits, showing that higher Josephson harmonic content is inherently set by junction geometry for fixed oxide properties. The ensemble ρ(τ)\rho(\tau) is a property of the oxide, and as more junctions are placed in a circuit the law of large numbers dictates that the extreme τ\tau values will be sampled. In systems with thousands of qubits, designers will need to account for those atypical junctions with very transparent channels sampled from the extreme right tail of ρ(τ)\rho(\tau) and with anomalously large higher harmonic content. The only way to eliminate these channels is to reduce oxide thickness variability so that ρ(τ)\rho(\tau) never develops a fat tail and remains self-averaging.

The data indicate that this disorder is inherited from the metal rather than generated by the oxidation, since σI\sigma_I is independent of oxidation exposure while dId_I grows with it (Table 1). Likewise, ℓth\ell_{\mathrm{th}} is exposure-independent yet differs between processes. The fitted roughness σI≈0.22\sigma_I \approx0.22–0.29 nm0.29\ \mathrm{nm} is comparable to a single atomic step of Al (0.200.20–0.23 nm0.23\ \mathrm{nm}), whereas σII≲0.14 nm\sigma_{II} \lesssim0.14\ \mathrm{nm} is sub-monolayer, suggesting that Region I is oxide grown over a stepped or faceted bottom-electrode surface and Region II is oxide grown over flatter terraces. Grain boundaries are the natural candidate for such a surface, and cross-sectional TEM and atomistic simulations indicate that oxide grown over boundaries between crystallographic domains and along film edges differs in thickness and structure from that over the domain interiors [31, 32, 33, 34, 35, 36]. Sidewall roughness of the climbing edge is a second candidate, but the rising-edge overlap has no appreciable effect on τe\tau_e (HAR C). Moreover, a purely perimeter or edge contribution would grow as l+wl+w and make I2I_2 increase at least as Ajj1/2A_{jj}^{1/2}, whereas we observe saturation. Rather than an edge effect, the fixed size of A1A_1 may therefore reflect a systematic roughness of the Al surface that arises when electrodes are fabricated with critical dimensions below lthl_{\mathrm{th}}.

Finally, because the measurement and analysis techniques presented are insensitive to series inductance, our area-dependent subgap characterization provides a geometry- and process-agnostic screen for the harmonic content intrinsic to Josephson junctions, and a powerful methodology for comparing junction fabrication processes. Applied consistently, the optimization map of Figure 5c enables such comparisons across processes, provided the geometric trends identified in this work are taken into account.

Supplementary information. Supplementary Information is available for this paper.

Acknowledgements. We thank John Teufel, Matt Matheny and Bastien Dassonneville for helpful discussions; Simon Luong, Lawrence Wong and Matthew Fling for hardware and cryogenics support; and the staff across the AWS Center for Quantum Computing (CQC) who enabled this work.

Declarations

  • Funding: Not applicable.

  • Competing interests: The authors declare no competing interests.

  • Data availability: Data supporting the findings of this study are available from the corresponding author upon reasonable request.

  • Author contribution: D.J.P. contributed to the experimental design, developed the instrumentation and model, performed the primary measurements, analyzed the data, and wrote the manuscript. A.B., M.C., and G.R. contributed to the analysis, played a critical role in contextualizing the results and framing the discussion, with A.B. and M.C. also helping to draft the manuscript. M.P., H.L., and C.C. contributed to the experimental design and measurements, and provided editorial support. Y.S. provided supervision and contributed to the experimental design and manuscript editing.

References

Supplementary Information

Non-ergodic Area Scaling of Andreev Transport in Al/AlOx_x/Al Josephson Junctions

This document provides supplementary information for “Non-ergodic Area Scaling of Andreev Transport in Al/AlOx_x/Al Josephson Junctions,” including fabrication details, measurement setup, supplemental DC transport data, details of the direct tunneling model with oxide thickness disorder, and model fitting procedures.

Symbol Definitions

  • d\boldsymbol{d}: oxide barrier thickness (nm).

  • d0d_0: mean oxide thickness (nm) of a generic thickness distribution ρ(d)\rho(d), used where no region is specified; the fitted Region-I and Region-II values are dId_{\mathrm{I}} and dIId_{\mathrm{II}}.

  • σ\sigma: standard deviation of the oxide thickness distribution (nm).

  • θ\theta: average barrier height (eV) of the AlOx_x insulator.

  • κ=2meθ/ℏ\kappa= \sqrt{2m_{\mathrm{e}}\theta}/\hbar: tunneling wavefunction decay constant (nm−1^{-1}); the transmission probability decays as e−2κde^{-2\kappa d}.

  • ω=κσ\omega= \kappa\sigma: dimensionless tunneling disorder parameter. ω=0.5\omega= 0.5 is the fat-tail onset (Eq. (16)); for ω≳0.8\omega\gtrsim0.8 the sampled τe\tau_e becomes size-dependent over the experimental range, and for ω>1\omega> 1 the thinnest channels exponentially dominate the higher moments (extreme-value regime).

  • ρ(d)\rho(d): oxide thickness probability distribution.

  • nchn_{\mathrm{ch}}: areal density of independent tunneling channels (μm−2\mu\mathrm{m}^{-2}).

  • Nch=Ajj⋅nchN_{\mathrm{ch}} = A_{\mathrm{jj}} \cdot n_{\mathrm{ch}}: total number of channels in a junction.

  • Mk=(1/Nch)∑iτikM_k = (1/N_{\mathrm{ch}}) \sum_i \tau_i^k: kk-th moment of the transparency distribution.

  • τe=M2/M1\tau_e = M_2/M_1: effective transparency.

  • Ne=NchM12/M2N_e = N_{\mathrm{ch}} M_1^2/M_2: effective number of channels.

  • mm, ss: log-normal shape parameters, m=ln⁡d0−s2/2m = \ln d_0 - s^2/2, s=ln⁡(1+σ2/d02)s = \sqrt{\ln(1 + \sigma^2/d_0^2)}.

  • AjjA_{\mathrm{jj}}: total junction area (μm2\mu\mathrm{m}^2).

  • lthl_{\mathrm{th}}: threshold length (μm\mu\mathrm{m}); junctions with min⁡(l,w)<lth\min(l,w) < l_{\mathrm{th}} behave as pure Region I.

  • tbt_b: bottom-electrode thickness.

  • Ath=lth2+2lthtbA_{\mathrm{th}} = l_{\mathrm{th}}^2 + 2l_{\mathrm{th}}t_b: area, sidewalls included, of a square junction of side lthl_{\mathrm{th}}; the Region-I area saturates at AthA_{\mathrm{th}}.

  • AIA_{\mathrm{I}}, AIIA_{\mathrm{II}}: areas of Regions I and II (main text, Eq. (4)); AI=AjjA_{\mathrm{I}} = A_{\mathrm{jj}} for min⁡(l,w)<lth\min(l,w) < l_{\mathrm{th}} and AI=AthA_{\mathrm{I}} = A_{\mathrm{th}} otherwise, AII=Ajj−AIA_{\mathrm{II}} = A_{\mathrm{jj}} - A_{\mathrm{I}}.

  • dId_{\mathrm{I}}, σI\sigma_{\mathrm{I}}: mean thickness and disorder of Region I (the disordered, bounded region).

  • dIId_{\mathrm{II}}, σII\sigma_{\mathrm{II}}: mean thickness and disorder of Region II (the more uniform region).

  • θI\theta_{\mathrm{I}}, θII\theta_{\mathrm{II}}: barrier heights of Regions I and II (eV).

  • ϕ\phi: superconducting phase difference across the junction.

  • a1a_1, a2a_2: amplitudes of the first and second Fourier harmonics of the current-phase relation (CPR).

  • ∣a2/a1∣|a_2/a_1|: CPR skew ratio; ∣a2/a1∣≈τe/8|a_2/a_1| \approx\tau_e/8 to leading order.

Fabrication

See Table S1. All junctions are Manhattan-style Al/AlOx/Al junctions, shadow-evaporated in a UHV electron-beam evaporator (Angstrom Engineering, HAR A in a Plassys) on high-resistivity Si. The tunnel barrier was formed by static oxidation in O2_2; the oxidation condition is quoted as the exposure EE, the product of the static O2_2 pressure and the oxidation time (Torr⋅\cdotmin), which sets the barrier thickness and specifies the process. For RnAjjR_nA_{\mathrm{jj}} the median value is given and nn the number of junctions analysed. Individual RnAjjR_nA_{\mathrm{jj}} values span ≈320\approx320–4190 Ω μm24190\ \Omega\ \mu\mathrm{m}^2, which covers standard fabrication conditions reported in the literature [13, 37, 38, 39, 40]. Areas are Ajj=lw+2ltbA_{\mathrm{jj}} = lw + 2lt_b, with ww the bottom- and ll the top-electrode width and tbt_b the bottom-electrode thickness; the second term is the overlap of the top electrode with the two sidewalls of the bottom electrode, over which the top electrode climbs. ll, ww are design values except for Square Set A, whose face dimensions were measured by SEM on a duplicate wafer, since the as-fabricated dimensions differed from design. The tabulated AjjA_{\mathrm{jj}} and RnAjjR_nA_{\mathrm{jj}} include the sidewall term.

SetJunction geometry (μm\mu\mathrm{m})E\mathcal{E} (Torr·min)AjjA_{\mathrm{jj}} (μm2\mu\mathrm{m}^{2})RnAjjR_{\mathrm{n}}A_{\mathrm{jj}} (Ω μm2\Omega\,\mu\mathrm{m}^{2})nn
Square Al=w=0.15l=w=0.15–1.21.2150.054–1.716768
Square Al=w=0.15l=w=0.15–1.21.2300.054–1.7191816
Square Al=w=0.15l=w=0.15–1.21.21200.054–1.71190617
Square Al=w=0.15l=w=0.15–1.21.22400.054–1.71309016
Square Bl=w=0.1l=w=0.1–1.01.040.018–1.083407
Square Bl=w=0.2l=w=0.2–1.21.2870.055–1.53149710
Square Bl=w=0.15l=w=0.15–1.21.24800.034–1.53347412
HAR Al=0.2l=0.2 or 0.3, w=0.15w=0.15–2.02.01000.045–0.62149522
HAR Bl=0.2l=0.2 or 0.3, w=0.15w=0.15–2.02.01000.045–0.62133316
HAR Cl=0.1l=0.1–1.21.2, w=0.2w=0.2–0.750.75960.038–0.45124310

Table S1. Fabrication and geometry of the measured junction sets.

Measurement Setup and Wiring

All measurements were performed in a dilution refrigerator (Bluefors LD) at a base temperature of approximately 10 mK. Junctions were mounted in an Al/Cu light-tight sample package. A schematic of the DC measurement circuit is shown in Figure S1.

DC bias lines were heavily filtered using a combination of Basel MFT silver-powder filters and modified QDevil RCRC filters, configured in two stages. The low-pass RCRC component values were chosen such that the total line resistance satisfies Rline<RjjR_{\mathrm{line}} < R_{\mathrm{jj}} while maintaining a 3 dB cutoff frequency below 50 kHz.

A Basel LNHR DAC served as the low-noise DC voltage source. Voltage was measured using an SRS SR560 low-noise voltage preamplifier followed by a Keithley 7510 digital multimeter. Current was measured using an SRS SR570 low-noise current preamplifier and a second Keithley 7510.

Supplemental DC Transport Data

Every type of junction analyzed showed the geometry- and AjjA_{\mathrm{jj}}-dependent behavior. Here we show further examples, with fits, which demonstrate that the phenomenology remains the same, but the absolute magnitudes of τe\tau_{\mathrm{e}}, NeN_{\mathrm{e}} can depend on the individual fabrication process. Figure S2 shows data from Square Set B with oxidation exposures of 4, 87, and 480 Torr·min.

τe\tau_{\mathrm{e}} and NeN_{\mathrm{e}} versus oxidation exposure

Figure S3 plots τe=I2/I1\tau_{\mathrm{e}} = I_{2}/I_{1} (see Figure 2b of the main text for definitions of I1,2I_{1,2}) versus oxidation exposure E\mathcal{E} for fixed junction area. For all areas, the observed power-law dependence is consistent with a Mott–Cabrera oxidation process [41, 42], wherein the oxide thickness (d0d_{0}) grows logarithmically with exposure. Substituting d0=a+bln⁡Ed_{0}=a+b\ln\mathcal{E} into an exponential tunneling model, τe∝e−2κd0\tau_{\mathrm{e}}\propto e^{-2\kappa d_{0}}, yields the power law τe∝E−2κb\tau_{\mathrm{e}}\propto\mathcal{E}^{-2\kappa b}, which linearizes on the log–log axes of Figure S3. Here bb is the logarithmic growth coefficient of the oxide, with units of length, so that the exponent 2κb2\kappa b is dimensionless. Fitting the Region-I thicknesses dId_{\mathrm{I}} of main-text Table 1 against ln⁡E\ln\mathcal{E} gives b=0.046b=0.046 nm, i.e., 0.11 nm of additional oxide per decade of exposure (Region

II gives 0.045 nm, or 0.10 nm per decade, essentially the same), hence a predicted exponent 2κb=0.502\kappa b = 0.50. This agrees with the measured slopes dln⁡τe/dln⁡Ed\ln\tau_e/d\ln\mathcal{E}, which have a median of −0.50-0.50 across the six areas. This result serves as confirmation of fabrication consistency between sample lots, which is critical for reliably extracting the conclusions presented in the main text.

Figure S3b plots Ne/AjjN_e/A_{jj} for fixed junction areas as a function of oxidation exposure. Many of the 120 and 240 Torr⋅\cdotmin data points for the smallest devices have extremely small I2I_2 values, which propagate into correspondingly large error bars in NeN_e. No

correlation is observed between NeN_e and oxidation exposure over the range of areas and exposures covered in this study.

Determining the form of ρ(d)\rho(d) based on the area dependence of τe\tau_e, NeN_e

The oxide thickness distribution in a Josephson junction has been evaluated with TEM metrology and described by both normal [32, 43] and log-normal distributions [35], with the latter argued to be more physical due to the presence of thick regions serving as an upper tail. The right-skew of the oxide thickness distribution ρ(d)\rho(d) is less critical than the shape of the left tail due to the exponential tunneling dominance of thin channels. TEM on cross-sectional cuts is useful for correlation, but cannot easily distinguish the subtle difference in the extreme limits of the left tail structure. Therefore, we develop an analysis procedure that includes (i) introduction of moment generating functions, (ii) order statistics, (iii) tail enhancement, and (iv) moment convergence using order statistics.

The experimental results in the main text indicate both geometric and area dependencies which are unexpected in ergodic systems. For the square devices, τe\tau_e is both area-dependent and non-monotonic, while NeN_e has two distinct regions with different AjjA_{\mathrm{jj}} dependence. For both small-area junctions and high-aspect-ratio junctions, τe\tau_e increases with area, while NeN_e is constant or increases sub-linearly. Here, the experimental results are compared with possible solution spaces of log-normal and normal oxide thickness distributions to test whether each can recreate the AjjA_{\mathrm{jj}} dependence observed for τe,Ne\tau_e,N_e. The overarching goal of this section is to define the correct distribution, understand how this distribution results in the area dependence of the observed Ne,τeN_e,\tau_e, and discuss junction quality metrics based on the properties of ρ(τ(d))\rho(\tau(d)).

Moment-Generating Functions

Here we derive expressions for the moment-generating function for both normal and log-normal distributions. These functions are useful for evaluating the role of rare tunneling events, which is critical for analyzing subgap idci_{\mathrm{dc}}–vdcv_{\mathrm{dc}} data.

The main text defines the analysis parameters τe,Ne\tau_e,N_e in terms of the experimentally measured I1,I2I_1,I_2. These can also be expanded in terms of the moments of ρ(τ)\rho(\tau), MkM_k, as follows:

Mk=1Nch∑i=1Nchτik=⟨τk⟩=∫0∞τ(x)kρ(x) dx.(6)M_k = \frac{1}{N_{\mathrm{ch}}}\sum_{i=1}^{N_{\mathrm{ch}}}\tau_i^k = \langle\tau^k \rangle= \int_{0}^{\infty}\tau(x)^k\rho(x)\,dx. \tag*{(6)}

τe,Ne\tau_e,N_e can then be written in terms of M1,M2M_1,M_2:

NchM1=Nch⟨τ⟩=∑i=1Nchτi=Gnh2e2(7)N_{\mathrm{ch}}M_1 = N_{\mathrm{ch}}\langle\tau\rangle= \sum_{i=1}^{N_{\mathrm{ch}}}\tau_i = G_n\frac{h}{2e^2} \tag*{(7)}
NchM2=Nch⟨τ2⟩=∑i=1Nchτi2=Gsgh2e2(8)N_{\mathrm{ch}}M_2 = N_{\mathrm{ch}}\langle\tau^2\rangle= \sum_{i=1}^{N_{\mathrm{ch}}}\tau_i^2 = G_{\mathrm{sg}}\frac{h}{2e^2} \tag*{(8)}
τe=M2M1(9)\tau_e = \frac{M_2}{M_1} \tag*{(9)}
Ne=NchM12M2(10)N_e = N_{\mathrm{ch}}\frac{M_1^2}{M_2} \tag*{(10)}
Normal distribution

The normal distribution has the form:

ρN(d)=1σ2πe−(d−d0)22σ2(11)\rho_N(d) = \frac{1}{\sigma\sqrt{2\pi}}e^{-\frac{(d-d_0)^2}{2\sigma^2}} \tag*{(11)}

from which we can calculate the nthn^{\mathrm{th}} moment (using τ(d)n=(C/d)ne−2nκd\tau(d)^n = (C/d)^n e^{-2n\kappa d}, where CC is the prefactor of the Simmons transparency, Eq. (5) of the main text):

Mn=Cn∫−∞∞d−ne−12σ2[d2−2(d0−2nκσ2)d+d02]ddσ2π(12)M_n = C^n \int_{-\infty}^{\infty} d^{-n}e^{-\frac{1}{2\sigma^2}\left[d^2-2(d_0-2n\kappa\sigma^2)d+d_0^2\right]}\frac{dd}{\sigma\sqrt{2\pi}} \tag*{(12)}

Defining a shifted mean μn∗=d0−2nκσ2\mu_n^* = d_0 - 2n\kappa\sigma^2, MnM_n simplifies to:

Mn=Cne−2nκd0+2n2κ2σ2⟨d−n⟩μn∗(13)M_n = C^n e^{-2n\kappa d_0+2n^2\kappa^2\sigma^2}\left\langle d^{-n}\right\rangle_{\mu_n^*} \tag*{(13)}

The experimental data are limited to the first two moments of ρ(τ)\rho(\tau):

M1≈Cμ1∗e−2κd0+2κ2σ2,μ1∗=d0−2κσ2(14)M_1 \approx\frac{C}{\mu_1^*}e^{-2\kappa d_0+2\kappa^2\sigma^2},\quad\mu_1^* = d_0 - 2\kappa\sigma^2 \tag*{(14)}
M2≈C2(μ2∗)2e−4κd0+8κ2σ2,μ2∗=d0−4κσ2(15)M_2 \approx\frac{C^2}{(\mu_2^*)^2}e^{-4\kappa d_0+8\kappa^2\sigma^2},\quad\mu_2^* = d_0 - 4\kappa\sigma^2 \tag*{(15)}
Defining an ω=0.5\omega= 0.5 criterion, and the validity of Eqs. S9–S10.

In the equations above, τ(d)n\tau(d)^n weights the thickness distribution by e−2nκde^{-2n\kappa d}, and the channels that carry MnM_n are not those near d0d_0 but those near μn∗\mu_n^* which are displaced into the thin tail by

d0−μn∗σ=2nω,μn∗−μn+1∗σ=2ω,(16)\frac{d_0-\mu_n^*}{\sigma}=2n\omega,\qquad\frac{\mu_n^*-\mu_{n+1}^*}{\sigma}=2\omega, \tag*{(16)}

where the displacement is expressed in units of the disorder σ\sigma itself. Equivalently, the standard deviation of ln⁡τ≈2κσ=2ω\ln\tau\approx2\kappa\sigma= 2\omega. Both expressions are functions of ω\omega only: they contain no reference to d0d_0, θ\theta, or NchN_{\mathrm{ch}}. The condition 1−2ω=01-2\omega=0, i.e. ω=0.5\omega=0.5, is therefore the point at which each successive moment is sourced by a subpopulation displaced a full standard deviation from the mean relative to the previous one. Below it, all accessible moments are set by channels within ∼σ\sim\sigma of d0d_0 and the transparency statistics self-average; above it, M2M_2 is controlled by a rare thin-channel tail whose sampling depends on NchN_{\mathrm{ch}}. This is the defining property of the fat-tail regime and the reason ω\omega can be used as a figure of merit for the disorder. The result is self-consistent with the size-dependent estimate in the main text also. For θ=1.15\theta=1.15 eV (κ=5.49\kappa=5.49 nm−1^{-1}), ω=0.5\omega=0.5 corresponds to σ=0.5/κ=0.091\sigma=0.5/\kappa=0.091 nm. ω=0.5\omega=0.5 is therefore the lower limit for any size dependence, although the effect is gradual. Below it, τe\tau_e is size-independent at every NchN_{\mathrm{ch}}. The threshold quoted in the main text, κσ≳0.8\kappa\sigma\gtrsim0.8, lies above this limit because the experimental window 104<Nch<2×10610^4<N_{\mathrm{ch}}<2\times10^6 is finite. Between the two, the tail exists but is not sampled strongly enough to show over our range of junction sizes.

The same logic sets the limits of Eqs. (14) and (15). They are second-order Laplace approximations about μn∗\mu_n^*. Checked against direct numerical integration of Eq. (12) at d0=1.269d_0 = 1.269 nm, the closed form for M2M_2 is accurate to 2% at ω=0.5\omega= 0.5 and 6% at ω=0.7\omega= 0.7, but the expression fails entirely by ω≃1.32\omega\simeq1.32, where μ2∗=d0−4κσ2\mu_2^* = d_0 - 4\kappa\sigma^2 passes through zero and the 1/(μ2∗)21/(\mu_2^*)^2 prefactor diverges. At the fitted Region-I disorder ωI≃1.5\omega_{\mathrm{I}} \simeq1.5 (Table S2) μ2∗\mu_2^* is negative, so Eqs. (14)–(15) cannot be evaluated at the parameters of these devices and are retained only to expose the μn∗\mu_n^* structure. Every moment, ξ\xi, τe\tau_e, and NeN_e value quoted in this work is instead computed from either the order-statistics sum (Eq. (23)), from Monte Carlo sampling, or from direct numerical integration of Eq. (12). All of these calculations apply the same two cutoffs, a numerical floor d≥0.05d \ge0.05 nm on the thickness and τ≤1\tau\le1 on the transparency. At every fitted (d0,σ)(d_0,\sigma) and every measured NchN_{\mathrm{ch}}, the smallest order statistic lies well above one monolayer, so no fitted parameter depends on either cutoff. The far-field illustrations (Figure S5 and Figure S6) extrapolate to channel counts and disorders at which the thin tail reaches τ=1\tau= 1. There the capped and perfectly transmitting sub-population carries the higher moments, with the region where it dominates marked in Figure S6.

GeometrydId_{\mathrm{I}} (nm)σI\sigma_{\mathrm{I}} (nm)σI/dI\sigma_{\mathrm{I}}/d_{\mathrm{I}}ωI\omega_{\mathrm{I}}
Square Set A (120 Torr·min)1.4980.2850.1901.56
HAR A+B (l=200l = 200 nm, pooled)1.3020.2260.1731.24

Table S2. Small-junction (pure Region-I) best-fit parameters for Figure S8, with barrier height fixed at θI=1.15\theta_{\mathrm{I}} = 1.15 eV (κ=5.49\kappa= 5.49 nm−1^{-1}).

Figure S5. Tail enhancement factor ξ=M2/τ‾2\xi= M_2/\overline{\tau}^2 comparing order-statistics (OS) summation to the exact Nch→∞N_{\rm ch} \to\infty integral, for normal (solid) and log-normal (dashed) thickness distributions with d0=1.269d_0 = 1.269 nm and θ=1.15\theta= 1.15 eV, using the numerical floor d≥0.05d \ge0.05 nm and the unitarity cap τ≤1\tau\le1 of all calculations in this work. Shaded bands denote the 99% confidence interval (0.5–99.5 percentile) of the Monte Carlo simulations. Without spatial correlations, the variation around E[d(k)]\mathbb{E}[d_{(k)}] results in overestimated tail amplification for small NchN_{\rm ch}. a, Convergence with channel count NchN_{\rm ch} at fixed ω=0.5\omega= 0.5, 1.05, and 1.55; thin horizontal lines mark the Nch→∞N_{\rm ch} \to\infty limit. At the onset value ω=0.5\omega= 0.5 the curves are flat, i.e. ξ\xi is already converged at Nch=103N_{\rm ch} = 10^3. At ω=1.55\omega= 1.55 the normal-distribution curve saturates early because its thinnest channels reach the cap τ=1\tau= 1; the capped sub-population then carries M2M_2. b, Dependence on disorder strength ω\omega at fixed NchN_{\rm ch}, with 1000 Monte Carlo realizations drawn from the corresponding distribution. The dotted vertical line marks the fat-tail onset ω=0.5\omega= 0.5. The dimensionless parameter ω=κσ\omega= \kappa\sigma is the natural scale for the normal distribution; log-normal curves are plotted at matched variance (i.e. the same physical σ\sigma), with shape parameter s=ln⁡(1+σ2/d02)s = \sqrt{\ln(1+\sigma^2/d_0^2)}. At large ω\omega the two distributions diverge because the log-normal tail vanishes as d→0d \to0, suppressing the thin-barrier enhancement that dominates M2M_2.

Figure S6. Required number of channels Nch=Nchar(n)/ϵ2N_{\mathrm{ch}} = N_{\mathrm{char}}^{(n)}/\epsilon^2 for the partial moments Ma,bpM^p_{a,b} and Mc,dpM^p_{c,d} to converge to within a fractional tolerance ϵ=0.05\epsilon= 0.05 of their asymptotic values, shown as a function of (σ,d0)(\sigma, d_0) for normal a,c and log-normal b,d thickness distributions. The red contour marks Nch=106N_{\mathrm{ch}} = 10^6, i.e. a 1 μm21~\mu\mathrm{m}^2 junction at nch=106 μm−2n_{\mathrm{ch}} = 10^6~\mu\mathrm{m}^{-2}. M1M_1 converges according to its variance, Nchar(1)=M2/M12−1N_{\mathrm{char}}^{(1)} = M_2/M_1^2 - 1; M2M_2 convergence is set by Nchar(2)=M4/M22−1N_{\mathrm{char}}^{(2)} = M_4/M_2^2 - 1. For the log-normal distribution, typical moments are converged for all relevant AjjA_{jj}; M2M_2 is converged for devices with Ajjnch≳104A_{jj} n_{\mathrm{ch}} \gtrsim10^4–10510^5. For the normal distribution, M1M_1 depends on σ\sigma and d0d_0, with convergence occurring across a large range Ajjnch≈105A_{jj} n_{\mathrm{ch}} \approx10^5–10710^7. For disordered oxides (σI>0.25\sigma_I > 0.25 nm), small areas may therefore be expected to have Gn/AjjG_n/A_{jj} depend on AjjA_{jj}. M2M_2 effectively never converges for these devices, consistent with experimental results where Ne,τeN_e, \tau_e vary across the entire range of junction areas. The dashed line in the normal-distribution panels bounds the corner of parameter space in which more than 0.1% of all channels sit at the unitarity cap τ=1\tau= 1; there the capped sub-population dominates the moments, and the rapid convergence reflects the cap rather than the sampling of the normal tail. All fitted devices lie far outside this corner.

Log-normal distribution

For the log-normal distribution, we have:

ρLN(d)=1ds2πexp⁡(−(ln⁡d−m)22s2)(17)\rho_{\mathrm{LN}}(d) = \frac{1}{d s\sqrt{2\pi}} \exp\left(-\frac{(\ln d-m)^2}{2s^2}\right) \tag*{(17)}

For the log-normal to have the same mean and variance as the normal distribution, em+s2/2=d0e^{m+s^2/2}=d_0 and (es2−1)e2m+s2=σ2(e^{s^2}-1)e^{2m+s^2}=\sigma^2, giving m≈ln⁡d0m \approx\ln d_0 and s≈σ/d0s \approx\sigma/d_0 when ss is small. The nthn^{\mathrm{th}} moment is

Mn=Cns2π∫−∞∞efn(u) du(18)M_n = \frac{C^n}{s\sqrt{2\pi}}\int_{-\infty}^{\infty}e^{f_n(u)}\,du \tag*{(18)}

with

fn(u)=−nu−2nκeu−(u−m)22s2(19)f_n(u) = -nu - 2n\kappa e^u - \frac{(u-m)^2}{2s^2} \tag*{(19)}

Order Statistics and Partial Moments

The expectation value (first moment) of the ithi^{\mathrm{th}} order statistic (OS, or ordered thickness) taken from a set of NchN_{\mathrm{ch}} oxide thicknesses with given thickness distribution ρ(d)\rho(d) is given by:

E[d(i)]=∫tox=0tox=∞tox⋅fd(i)(tox) dtox(20)\mathbb{E}[d_{(i)}] = \int_{t_{\rm ox}=0}^{t_{\rm ox}=\infty} t_{\rm ox}\cdot f_{d_{(i)}}(t_{\rm ox})\,\mathrm{d}t_{\rm ox} \tag*{(20)}

where tox=dt_{\rm ox}=d is the oxide thickness, 1≤i≤Nch1 \le i \le N_{\rm ch}, and fd(i)f_{d_{(i)}} is the PDF of the iith order statistic. For a normal distribution this is estimated using the quantile (probit) approximation [44]:

E[d(i)]≈d0+σQ(iNch+1)(21)\mathbb{E}[d_{(i)}] \approx d_0+\sigma Q\left(\frac{i}{N_{\rm ch}+1}\right) \tag*{(21)}

where Q(⋅)Q(\cdot) denotes the standard-normal quantile (probit) function. For a log-normal distribution, E[d(i)]\mathbb{E}[d_{(i)}] is then given by:

E[d(i)]≈em+sQ(iNch+1)s=ln⁡(1+σ2d02)m=ln⁡d0−s2/2.(22)\begin{aligned} \mathbb{E}[d_{(i)}] &\approx e^{m+sQ\left(\frac{i}{N_{\rm ch}+1}\right)} \\ s &= \sqrt{\ln\left(1+\frac{\sigma^2}{d_0^2}\right)} \\ m &= \ln d_0-s^2/2. \tag*{(22)} \end{aligned}

Mean values of the partial moments can then be calculated for NchN_{\mathrm{ch}} (AjjA_{\mathrm{jj}}) using order statistics:

Mnp=∑k=1NchτnE[d(k)](23)M_n^p = \sum_{k=1}^{N_{\mathrm{ch}}} \tau^n \mathbb{E}[d_{(k)}] \tag*{(23)}

Tail Enhancement

At the heart of the non-self-averaging behavior and experimental trends for τe\tau_e, NeN_e is the fat tail of the ρ(τ)\rho(\tau) distribution. To quantify the rare-event statistics resulting from the tail, we use the above formalism and define an nthn^{\mathrm{th}} order tail-enhancement factor ξ\xi as the ratio of the nn-th moment to the value expected if all channels had the mean thickness d0d_0:

ξ=Mnτ(d0)n=∫0∞(d0x)ne−2nκ(x−d0)ρ(x) dx(24)\xi= \frac{M_n}{\tau(d_0)^n} = \int_0^\infty\left(\frac{d_0}{x}\right)^n e^{-2n\kappa(x-d_0)}\rho(x)\,dx \tag*{(24)}

Since MnM_n is defined per channel (Eq. (23) normalized by NchN_{\mathrm{ch}}), ξ\xi carries no explicit NchN_{\mathrm{ch}} factor; its NchN_{\mathrm{ch}} dependence enters only through the sampling of the tail. Figure S5 shows numerical results from Eqs. (12),(18) compared with results from 1000 Monte Carlo (MC) trials and OS calculations. The amount of oxide thickness disorder is represented by the dimensionless ω=κσ\omega= \kappa\sigma, which increases with σ\sigma. On one extreme, ω→0\omega\to0 represents a perfectly uniformly thick oxide layer. In such a case, nearly all channels have the same τ≈τe\tau\approx\tau_e and Ne≈NchN_e \approx N_{\mathrm{ch}}. In the other extreme, ω\omega grows monotonically as σ\sigma becomes of order d0d_0. This is further shown in the τe\tau_e - NchN_{\mathrm{ch}} simulation in Figure 4, where d0d_0 is fixed but σ\sigma is swept through representative values for a junction oxide.

For a normal distribution of dd, the tunneling transmission τ(d)=(C/d)e−2κd\tau(d) = (C/d)e^{-2\kappa d} maps a symmetric oxide thickness distribution into a heavily right-skewed transparency distribution. Both the exponential decay and the 1/d1/d prefactor amplify the contribution of thin-barrier channels. Even modest oxide roughness σ\sigma produces a fat-tailed ρ(τ)\rho(\tau) whose higher moments are dominated by rare thin channels. The resulting ξ\xi grows as disorder increases. Conversely, for the log-normal thickness distribution, the saddle point (Eq. (19)) never shifts far into the thin-barrier tail (d→0d \to0, i.e. u→−∞u \to-\infty), and ξ\xi saturates at a fixed value in junctions with Ajj>0.01 μm2A_{\mathrm{jj}} > 0.01\ \mu\mathrm{m}^2. It is this tail enhancement or fat-tail of the distribution that is largely responsible for the area dependence observed in τe\tau_e, NeN_e.

The expectation value of probability distributions with fat tails depends on the number of samples. For small devices, the expectation value is channel-number dependent and the value reflects the typical result, because few extreme outliers are sampled. Increasingly, more extreme values (thin channels) occur which dominate MkM_k, k≥2k \ge2, and push the typical solution toward the average. These rare thin-channel transparencies are large enough to rival the summed transparency of all remaining channels. With a large enough device, or large enough number of samples, the typical value of M1M_1, M2M_2 converges to the fully sampled average. The number of channels required to transition from the typical to the average depends on the relative level of oxide disorder, given by ω\omega, as in Figure S5. The fully sampled average regime is well described by the AjjA_{\mathrm{jj}}-independent values given numerically by Eqs. (12),(18).

Figure S5 exemplifies the second-moment tail enhancement M2/τ02M_2/\tau_0^2 and its AjjA_{jj}-dependence for normal distributions with ω>1\omega> 1. The log-normal distribution, however, has orders-of-magnitude smaller ξ\xi and requires only 10310^3–10410^4 channels for M2M_2 to reach the fully-sampled average. Normal-distributed oxide thicknesses have ω\omega-dependent convergence of the OS simulations to the fully-sampled numerical result. Figure S5 indicates that at ω≈1\omega\approx1, Nch∼107N_{\mathrm{ch}} \sim10^7 is required for convergence, at least 5 times as large as the biggest samples here. At ω=1.55\omega= 1.55 the earlier saturation is set by the cap τ≤1\tau\le1 rather than by sampling.

Numerical Comparison of log-normal and normal MnM_n convergence using order statistics

The convergence of M1p,M2pM_1^p,M_2^p partial moments to their average determines the area dependence of Ne,τeN_e,\tau_e. When partial moments are converged, τe\tau_e will not depend on AjjA_{jj}. While the saturation of M2/τ02M_2/\tau_0^2 matches convergence of M1p,M2pM_1^p,M_2^p (see Figure S5), here we quantitatively map the σ,d0\sigma,d_0-dependent convergence using the coefficient of variation CV⁡(τn)=SE⁡(M^n)NchMn\operatorname{CV}(\tau^n) = \frac{\operatorname{SE}(\hat{M}_n)\sqrt{N_{\mathrm{ch}}}}{M_n}. If we require MnpM_n^p to lie within ϵ\epsilon of MnM_n with probability 0.68, we have:

N(n,ϵ)=CV⁡(τn)2ϵ2=Nchar(n)ϵ2(25)N(n,\epsilon)=\frac{\operatorname{CV}(\tau^n)^2}{\epsilon^2}=\frac{N_{\mathrm{char}}^{(n)}}{\epsilon^2} \tag*{(25)}

where Nchar(1)=M2/M12−1=Nch/Ne−1N_{\mathrm{char}}^{(1)}=M_2/M_1^2-1=N_{\mathrm{ch}}/N_e-1 defines M1M_1 convergence, and Nchar(2)=M4/M22−1N_{\mathrm{char}}^{(2)}=M_4/M_2^2-1 defines M2M_2 convergence and hence NeN_e linearity in NchN_{\mathrm{ch}}, both of which are signatures of convergence to the thermodynamic regime. Figure S6 shows the required Nch(σ,d0)N_{\mathrm{ch}}(\sigma,d_0) for different ϵ\epsilon. If σ,d0\sigma,d_0 are small, normal and log-normal convergence of M1,M2M_1,M_2 occur in the same order of magnitude. However, as σ,d0\sigma,d_0 increase together, the log-normal distribution results in convergence orders of magnitude faster for M1M_1 (M2M_2). The areas probed experimentally all fall into the simulation range that has converged M1,M2M_1,M_2 for ρ(d)∼LN(d0,σ)\rho(d) \sim\mathcal{LN}(d_0,\sigma). The order of magnitude of NchN_{\mathrm{ch}} is not consistent with experimental results, which show no convergence of τe\tau_e.

Large Junction τe\tau_e, NeN_e

For large junctions, M1,M2M_1,M_2 converge to their Nch→∞N_{\mathrm{ch}} \to\infty average values given by numerical integration of Eq. (18) and Eq. (12). From Eq. (13):

Ne=NchM12M2=Nch⟨d−1⟩μ∗2⟨d−2⟩μ∗e−4κ2σ2=βNch(26)N_e=N_{\mathrm{ch}}\frac{M_1^2}{M_2}=N_{\mathrm{ch}}\frac{\langle d^{-1}\rangle_{\mu_*}^2}{\langle d^{-2}\rangle_{\mu_*}}e^{-4\kappa^2\sigma^2}=\beta N_{\mathrm{ch}} \tag*{(26)}

where β\beta is a constant. Similarly, τe=M2/M1\tau_e=M_2/M_1 converges to a constant value:

τe=⟨d−2⟩μ∗⟨d−1⟩μ∗Ce−2κd0+6κ2σ2(27)\tau_e=\frac{\langle d^{-2}\rangle_{\mu_*}}{\langle d^{-1}\rangle_{\mu_*}}C e^{-2\kappa d_0+6\kappa^2\sigma^2} \tag*{(27)}

both of which follow from Eq. (13) and inherit its validity range 2nω≲12n\omega\lesssim1; as noted above, they are not evaluable at the fitted ωI≈1.5\omega_I \approx1.5 and the quoted values come from the exact OS/MC/numerical calculations.

The model reproduces the experimental NeN_e area dependence. The constant-β\beta limit of Eq. (26) is the asymptote reached only once M2M_2 has converged, which for these disorder parameters requires Nch≳107N_{\rm ch} \gtrsim10^7 (Figure S6); over the measured range β\beta instead grows by roughly an order of magnitude. The experimental larger-area trend is dominated by the transition to less-disordered oxide in larger devices.

Small-area τe\tau_e, NeN_e: Why is there an area dependence?

We now derive the expected area dependence of τe\tau_e using order statistics. From Eq. (23), the partial moments Mnp(Nch)M_n^p(N_{\rm ch}) are deterministic functions of NchN_{\rm ch}. When the channel count increases Nch→Nch+1N_{\rm ch} \to N_{\rm ch}+1, a new thinnest channel always appears with expected thickness:

E[d(1)]=d0+σQ(1Nch+2)(28)\mathbb{E}[d_{(1)}] = d_0 + \sigma Q\left(\frac{1}{N_{\rm ch}+2}\right) \tag*{(28)}

for the normal distribution, and

E[d(1)]=em+sQ(1Nch+2)(29)\mathbb{E}[d_{(1)}] = e^{m+sQ\left(\frac{1}{N_{\rm ch}+2}\right)} \tag*{(29)}

for log-normal. Simultaneously, all existing quantiles shift from k/(Nch+1)→k/(Nch+2)k/(N_{\rm ch}+1) \to k/(N_{\rm ch}+2). In the small-junction regime the dominant contribution to δτe\delta\tau_e comes from the new extreme channel; the quantile compression of existing channels contributes at O(1/Nch2)\mathcal{O}(1/N_{\rm ch}^2). Defining the transparency of this new extreme channel to be τ1≡τ(E[d(1)])\tau_1 \equiv\tau(\mathbb{E}[d_{(1)}]), the updated effective transparency τe′\tau_e' is:

τe′=M2p+τ12M1p+τ1(30)\tau_e' = \frac{M_2^p+\tau_1^2}{M_1^p+\tau_1} \tag*{(30)}

and the expected change in τe\tau_e per added channel is:

E[δτe]=τe′−τe=τ1(τ1−τe)M1p+τ1≈τ1(τ1−τe)M1p.(31)\mathbb{E}[\delta\tau_e] = \tau_e' - \tau_e = \frac{\tau_1(\tau_1-\tau_e)}{M_1^p+\tau_1} \approx\frac{\tau_1(\tau_1-\tau_e)}{M_1^p}. \tag*{(31)}

We can make the approximation τ1≪M1p\tau_1 \ll M_1^p since even the most transparent single channel carries negligible I1I_1 compared to the sum. τe\tau_e increases with area whenever the new thinnest channel is more transparent than the current weighted average, i.e., τ1>τe\tau_1 > \tau_e.

For the normal distribution, the expected minimum thickness from Eq. (28) gives:

τ1=τ(E[d(1)])=CE[d(1)]e−2κE[d(1)](32)\tau_1 = \tau(\mathbb{E}[d_{(1)}]) = \frac{C}{\mathbb{E}[d_{(1)}]}e^{-2\kappa\mathbb{E}[d_{(1)}]} \tag*{(32)}

For large NchN_{\rm ch}, Q(1/(Nch+2))≈−2ln⁡NchQ(1/(N_{\rm ch}+2)) \approx-\sqrt{2\ln N_{\rm ch}}, so E[d(1)]≈d0−σ2ln⁡Nch\mathbb{E}[d_{(1)}] \approx d_0-\sigma\sqrt{2\ln N_{\rm ch}} and:

τ1∼e−2κd0+2κσ2ln⁡Nch(33)\tau_1 \sim e^{-2\kappa d_0+2\kappa\sigma\sqrt{2\ln N_{\rm ch}}} \tag*{(33)}

This function grows sub-polynomially but without bound in NchN_{\mathrm{ch}} (as e2κσ2ln⁡Nche^{2\kappa\sigma\sqrt{2\ln N_{\mathrm{ch}}}}). In contrast, for the log-normal distribution:

E[d(1)]=em−s2ln⁡Nch(34)\mathbb{E}[d_{(1)}] = e^{m-s\sqrt{2\ln N_{\mathrm{ch}}}} \tag*{(34)}

with m=ln⁡d0−s2/2m=\ln d_0-s^2/2 and s=ln⁡(1+σ2/d02)s=\sqrt{\ln(1+\sigma^2/d_0^2)}, giving:

τ1=CE[d(1)]e−2κem−s2ln⁡Nch(35)\tau_1=\frac{C}{\mathbb{E}[d_{(1)}]}e^{-2\kappa e^{m-s\sqrt{2\ln N_{\mathrm{ch}}}}} \tag*{(35)}

The rate dτe/dNchd\tau_e/dN_{\mathrm{ch}}, which is an experimentally measurable quantity, trends to zero. Noting that E[d(1)LN]∝e−ln⁡Nch\mathbb{E}[d_{(1)}^{\mathrm{LN}}]\propto e^{-\sqrt{\ln N_{\mathrm{ch}}}} converges to zero much more slowly than E[d(1)N]=d0−σ2ln⁡Nch\mathbb{E}[d_{(1)}^{\mathrm{N}}]=d_0-\sigma\sqrt{2\ln N_{\mathrm{ch}}}, τ1LN\tau_1^{\mathrm{LN}} saturates at modest NchN_{\mathrm{ch}} while τ1N\tau_1^{\mathrm{N}} continues to grow. This causes the log-normal τe\tau_e to converge to its true average in all but the smallest of junctions, while the normal-distribution τe\tau_e continues to increase. Based on the areal trends predicted by the model, only a normal distribution of oxide thickness matches the experimental AjjA_{\mathrm{jj}} dependence observed in Ne,τeN_e,\tau_e.

Validating the existence of a small-junction solution space

Although the area-dependent trend of Ne,τeN_e,\tau_e is directly predicted for ρ(d)\rho(d) normal, the question arises as to whether any physically reasonable combination of σ\sigma, d0d_0, θ\theta satisfies the actual experimental τe(Ajj)\tau_e(A_{\mathrm{jj}}), Ne(Ajj)N_e(A_{\mathrm{jj}}) trends. As a demonstration, we consider the three smallest junction sizes (Ajj≤0.153 μm2A_{\mathrm{jj}}\leq0.153\ \mu\mathrm{m}^2) because τe\tau_e increases monotonically in this region. OS is then used to calculate the deterministic solution space of normal and log-normal distributions for d0∈[0.8 nm,2 nm]d_0\in[0.8\ \mathrm{nm},2\ \mathrm{nm}], θ∈[0.5 eV,2.5 eV]\theta\in[0.5\ \mathrm{eV},2.5\ \mathrm{eV}], and σ∈[0.1 nm,0.6 nm]\sigma\in[0.1\ \mathrm{nm},0.6\ \mathrm{nm}]. For a given solution, if ∣Ne,sim−Ne,exp∣/Ne,exp<0.1|N_{e,\mathrm{sim}}-N_{e,\mathrm{exp}}|/N_{e,\mathrm{exp}}<0.1 and ∣τe,sim−τe,exp∣/τe,exp<0.1|\tau_{e,\mathrm{sim}}-\tau_{e,\mathrm{exp}}|/\tau_{e,\mathrm{exp}}<0.1, that parameter set scores +1. For every die where this holds we add another point; the 16 dies at the three smallest sizes (Ajj=0.054A_{\mathrm{jj}}=0.054, 0.080 and 0.153 μm2\mu\mathrm{m}^2) are scored. Results for both normal and log-normal ρ(d)\rho(d) are shown in Figure S7, with the solution spaces resulting from different (d0,σ,θd_0,\sigma,\theta) values. Curves sit along constant σ/d0∼1/5\sigma/d_0\sim1/5 (normal) and σ/d0∼1/3\sigma/d_0\sim1/3 (log-normal). The collinearity of the points indicates that θ\theta acts only as a scaling parameter within the model, and can be chosen based on TEM data or literature values for the oxide and contacts. As Figure S7,Figure S7 shows, no combinations from the log-normal distribution fit more than a single experimental (Ne,τeN_e,\tau_e) data point. The rapid convergence of M1,M2M_1,M_2 results in (i) a linear NeN_e vs. AjjA_{\mathrm{jj}} relationship instead of the sublinear dependence observed experimentally for small areas, and (ii) τe≈const\tau_e\approx\mathrm{const}, instead of the linear increase. These results are consistent with the convergence analysis presented in Figure S5 and Figure S6, but shown explicitly to hold also for the experimental data. We therefore conclude that the underlying ρ(τ)\rho(\tau) is described best by a normal distribution (Eq. (11)). We cannot rule out the possibility that the normal distribution skews right, but this upper tail is irrelevant to the tunneling transparencies and adds unnecessary complexity. This result reflects the properties of the lower tail of the oxide thickness distribution, where the majority of tunnel current flows.

Figure S7. Comparing the area dependence of junction observables assuming normal ρ(d)∼N(dI,σI)\rho(d) \sim\mathcal{N}(d_I,\sigma_I) or log-normal ρ(d)∼LN(dI,σI)\rho(d) \sim\mathcal{LN}(d_I,\sigma_I) oxide thickness distributions. a, Colored (σ,d0)(\sigma,d_0) points denote where, for 0.5<θ<2.50.5 < \theta< 2.5 eV, a solution curve passes within 10% of any experimental (Ne,τe)(N_e,\tau_e) pair. Each curve is scored against the 16 dies at the three smallest junction sizes (Ajj=0.054,0.080(A_{jj} = 0.054, 0.080 and 0.153 μm2)0.153\ \mu\mathrm{m}^2). From a total simulation pool of 288,684 valid (σ,d0,θ)(\sigma,d_0,\theta) combinations, for ρ(d)∼LN(dI,σI)\rho(d) \sim\mathcal{LN}(d_I,\sigma_I), 329 curves matched a single die and none matched more than one. For ρ(d)∼N(dI,σI)\rho(d) \sim\mathcal{N}(d_I,\sigma_I), 749 curves matched, up to five dies at once, and these curves captured the area dependence much more accurately for Ne,τeN_e,\tau_e, as shown in b and c. The dashed lines in c represent the expected Ne∝NchN_e \propto N_{\mathrm{ch}} when M1,M2M_1,M_2 reach convergence. d, Example histogram comparing optimum fit curves from (b,c). The normal distribution results in a much wider fat-tailed ρ(τ)\rho(\tau). e, The linear dependence in (a) shows that θ\theta is simply a rescaling parameter for the (d0,σ)(d_0,\sigma) solution space. The true metric determining Ne,τeN_e,\tau_e is σ/d0\sigma/d_0. Likewise, every solution has ω=κσ>1\omega= \kappa\sigma> 1, well above the fat-tail onset ω=0.5\omega= 0.5 and the size-dependence threshold ω≳0.8\omega\gtrsim0.8, so the entire solution space is in the non-self-averaging regime.

For a normal distribution, convergence of M2M_2 to the true average requires Nch≫106N_{\mathrm{ch}} \gg10^6 or Ajj>1 μm2A_{jj} > 1\ \mu\mathrm{m}^2, which is larger than standard junction sizes in quantum computing circuits. The histograms of the τ\tau distributions from (b),(c) are shown in Figure S7 on a log scale. As predicted above, fewer than ∼10\sim10 channels in the fat tail account for the differences between normal and log-normal. As Figure S7 shows, the solution space collapses onto lines of constant σ/d0\sigma/d_0: because θ\theta enters only through κ\kappa, NeN_e and τe\tau_e depend on (d0,σ,θ)(d_0,\sigma,\theta) through κd0\kappa d_0 and ω=κσ\omega= \kappa\sigma alone, and σ/d0=ω/(κd0)\sigma/d_0 = \omega/(\kappa d_0) is the θ\theta-independent invariant of the fit. Every solution has ω>1\omega> 1, which places the entire space beyond both thresholds established above: the fat-tail onset ω=0.5\omega= 0.5 (Eq. (16)), above which M2M_2 is sourced by a rare thin-channel tail rather than by channels near d0d_0, and the size-dependence threshold κσ≳0.8\kappa\sigma\gtrsim0.8 quoted in the main text, above which the sampled τe\tau_e varies over the experimental window $10^4 < N_{\mathrm{ch}} < 2 \times 10^6$ (strongly so for κσ≳1\kappa\sigma\gtrsim1). The Region-I fits of Table S2 (wI=1.2(w_I = 1.2 to 1.6)1.6) lie in the same regime. Below the onset, by contrast, all accessible moments are carried by channels within ∼σ\sim\sigma of d0d_0 and the transparency statistics self-average, so τe\tau_e is size-independent and Ne∝NchN_e \propto N_{\mathrm{ch}} (main text, Figure 4b,Figure 4c); Region II, with $\omega_I \approx 0.55 (§ 2.6), sits below the size-dependence threshold and self-averages over the measured range. ω\omega is therefore the figure of merit for junction quality: it sets both the convergence of MkpM_k^p with NchN_{\mathrm{ch}} and the tail enhancement ξ\xi (Figure S5), while σ/d0\sigma/d_0 fixes it for a given barrier height.

The remainder of the analysis will proceed considering only a normal distribution for ρ(d)\rho(d).

Model Fitting

Small-Junction Fitting

We fit the small-junction (pure Region-I) regime using:

L(dI,σI)=∑i[(log⁡τe,imodel−log⁡τe,iexperiment)2+(log⁡Ne,imodel−log⁡Ne,iexperiment)2+(log⁡(GnAjj)imodel−log⁡(GnAjj)iexperiment)2](36)\mathcal{L}(d_{\mathrm{I}},\sigma_{\mathrm{I}})=\sum_i\left[(\log\tau_{e,i}^{\mathrm{model}}-\log\tau_{e,i}^{\mathrm{experiment}})^2+\left(\log N_{e,i}^{\mathrm{model}}-\log N_{e,i}^{\mathrm{experiment}}\right)^2+\left(\log\left(\frac{G_n}{A_{\mathrm{jj}}}\right)_i^{\mathrm{model}}-\log\left(\frac{G_n}{A_{\mathrm{jj}}}\right)_i^{\mathrm{experiment}}\right)^2\right] \tag*{(36)}

where L(dI,σI)\mathcal{L}(d_{\mathrm{I}},\sigma_{\mathrm{I}}) is the cost function. As shown in Figure S7, the solution space for (dI,σI,θI)(d_{\mathrm{I}},\sigma_{\mathrm{I}},\theta_{\mathrm{I}}) is not unique, but rather scales with the choice of θ\theta. Further, σI/dI≈0.19\sigma_{\mathrm{I}}/d_{\mathrm{I}}\approx0.19 for all samples tested in this study, which limits the fitting so that optimization is constrained and a global minimum can be found.

Results are shown in Figure S8. The OS method of extracting parameters is also confirmed using the same dId_{\mathrm{I}}, σI\sigma_{\mathrm{I}}, θI\theta_{\mathrm{I}} in MC simulations (1000 runs). MC has the advantage of showing the entire solution space, compared to OS which gives the typical result. Unfortunately, in non-self-averaging systems the average of the entire solution space implies randomness 1–2 orders of magnitude larger than experimentally observed. For this reason, channel homogeneity is handled more deterministically with order statistics. As shown in Figure S8, the MC median value tracks with the OS value with the same input parameters for dId_{\mathrm{I}}, θI\theta_{\mathrm{I}}, σI\sigma_{\mathrm{I}}. Large statistics of devices (hundreds), which is not experimentally viable, would be needed to account for the rare extreme events.

Figure S8. Small-area fits: order statistics (OS) and Monte Carlo simulations. The OS fit is performed on the pooled l=200l = 200 nm HAR junctions (Sets A and B) and on the Square Set A 120 Torr·min junctions at the two smallest areas (Nch<105N_{\mathrm{ch}} < 10^{5}) using Eq. (36) with parameters listed in Table S2. The ±10%\pm10\% error band uses fixed σI\sigma_{\mathrm{I}}, dId_{\mathrm{I}} while varying θI\theta_{\mathrm{I}}. MC simulations show 80% CI using the same parameters extracted from the OS fitting. No correlations are included in the MC simulations, therefore due to the non-self-averaging behavior of τe\tau_{\mathrm{e}}, NeN_{\mathrm{e}} the resulting error bars are an order of magnitude larger than experimentally (or ‘typically’) observed. If we re-solve for σI\sigma_{\mathrm{I}}, dId_{\mathrm{I}} while varying θI\theta_{\mathrm{I}}, the resulting error bands are invisible behind the original line, further confirming that θI\theta_{\mathrm{I}} acts only as a scaling parameter for junctions with the same fabrication process. a, τe\tau_{\mathrm{e}} vs. NchN_{\mathrm{ch}}. b, NeN_{\mathrm{e}} vs. NchN_{\mathrm{ch}}. c, Gn/AjjG_{\mathrm{n}}/A_{\mathrm{jj}} vs. NchN_{\mathrm{ch}} trends are correctly reproduced. The AjjA_{\mathrm{jj}} dependence of Gn/AjjG_{\mathrm{n}}/A_{\mathrm{jj}} results from a large ωI=κσI\omega_{\mathrm{I}} = \kappa\sigma_{\mathrm{I}} value which prevents M1=Gnh/(2e2Nch)M_{1} = G_{\mathrm{n}}h/(2e^{2}N_{\mathrm{ch}}) from converging for Ajj<2A_{\mathrm{jj}} < 2 μm2^{2}. This is a consequence of a fat-tailed distribution, and can be reduced by fabricating junctions with smaller σI/dI\sigma_{\mathrm{I}}/d_{\mathrm{I}}. Varying θI=1.15±0.115\theta_{\mathrm{I}} = 1.15\pm0.115 eV results in Gn/AjjG_{\mathrm{n}}/A_{\mathrm{jj}} changes that are far greater than experimentally observed; therefore this disorder mechanism is not present in the junctions.

Region-II and Multi-device Fitting

θI=θII=1.15\theta_{\mathrm{I}}=\theta_{\mathrm{II}}=1.15 eV and the Region-II roughness σII=0.10\sigma_{\mathrm{II}}=0.10 nm are held fixed, with θ\theta serving as a scaling factor during combined cost minimization to find the remaining free parameters: dI(E)d_{\mathrm{I}}(\mathcal{E}), dII(E)d_{\mathrm{II}}(\mathcal{E}), and the threshold length lthl_{\mathrm{th}} that caps the Region-I area at Ath=lth2+2lthtbA_{\mathrm{th}}=l_{\mathrm{th}}^2+2l_{\mathrm{th}}t_b once min⁡(l,w)≥lth\min(l,w)\ge l_{\mathrm{th}} (main text, Eq. (4)). Results of the fitting are shown in Figure 4 in the main text. These results are deterministic and identical for all junctions of a given area AjjA_{\mathrm{jj}}.

Holding σII\sigma_{\mathrm{II}} fixed is a convention and not a fit result. The transport data bound σII\sigma_{\mathrm{II}} from above but the fit does not resolve it. Re-solving the joint fit with σII\sigma_{\mathrm{II}} pinned at values spanning 0.005–0.30 nm, refitting lthl_{\mathrm{th}} and dII(E)d_{\mathrm{II}}(\mathcal{E}) at each (Figure S9), leaves the cost flat to 1% up to σII=0.14\sigma_{\mathrm{II}}=0.14 nm, giving a 1σ\sigma upper bound σII≲0.14\sigma_{\mathrm{II}}\lesssim0.14 nm with no lower limit. Within this plateau σII\sigma_{\mathrm{II}} is degenerate with dIId_{\mathrm{II}}: both act only through the thin-barrier tail of τ(d)\tau(d), and a 28×\times change in σII\sigma_{\mathrm{II}} is absorbed by a ≲60\lesssim60 pm shift of dIId_{\mathrm{II}}. The upper bound is algebraic, since Region II must remain self-averaging over the measured range. Therefore ωII=κσII≲0.8\omega_{\mathrm{II}}=\kappa\sigma_{\mathrm{II}}\lesssim0.8 or σII≲0.15\sigma_{\mathrm{II}}\lesssim0.15 nm at θ=1.15\theta=1.15 eV. Beyond this threshold the Region-II tail becomes non-self-averaging and dominates I2I_2, erasing the small/large-junction contrast that defines lthl_{\mathrm{th}}. As σII\sigma_{\mathrm{II}} rises, the fit cost rises accordingly by 3×\times at σII=0.20\sigma_{\mathrm{II}}=0.20 nm (ωII=1.1\omega_{\mathrm{II}}=1.1) and 60×\times at 0.30 nm. The adopted

σII=0.10\sigma_{\mathrm{II}} = 0.10 nm (ωII=0.55\omega_{\mathrm{II}} = 0.55) sits inside the allowed plateau, well below the fitted Region-I roughness σI≈0.28\sigma_{\mathrm{I}} \approx0.28–0.290.29 nm; the residual conventional freedom enters the quoted dIId_{\mathrm{II}} only as a ±60\pm60 pm systematic.

References

  1. [1]Josephson, B. D. Possible new effects in superconductive tunnelling. Physics letters 1, 251–253 (1962).DOI
  2. [2]Josephson, B. D. The discovery of tunnelling superconductors. Reviews of Modern Physics 46, 251 (1974).
  3. [3]Google Quantum AI. Exponential suppression of bit or phase errors with cyclic error correction. Nature 595, 383–387 (2021).DOI
  4. [4]Sivak, V. V. et al. Real-time quantum error correction beyond break-even. Nature 616, 50–55 (2023).DOI
  5. [5]Krinner, S. et al. Realizing repeated quantum error correction in a distance-three surface code. Nature 605, 669–674 (2022).DOI
  6. [6]Putterman, H. et al. Hardware-efficient quantum error correction via concatenated bosonic qubits. Nature 638, 927–934 (2025).DOI
  7. [7]Roy, A. & Devoret, M. Introduction to parametric amplification of quantum signals with josephson circuits. Comptes Rendus Physique 17, 740–755 (2016).DOI
  8. [8]Albertinale, E. et al. Detecting spins by their fluorescence with a microwave photon counter. Nature 600, 434–438 (2021).DOI
  9. [9]Fagaly, R. L. Superconducting quantum interference device instruments and applications. Review of Scientific Instruments 77, 101101 (2006).DOI
  10. [10]Kim, Y. et al. Evidence for the utility of quantum computing before fault tolerance. Nature 618, 500–505 (2023).DOI
  11. [11]Morvan, A. et al. Phase transition in random circuit sampling. Nature 634, 328–333 (2024).DOI
  12. [12]Bravyi, S., Dial, O., Gambetta, J. M., Gil, D. & Nazario, Z. The future of quantum computing with superconducting qubits. Journal of Applied Physics 132, 160902 (2022).DOI
  13. [13]Place, A. P. et al. New material platform for superconducting transmon qubits with coherence times exceeding 0.3 milliseconds. Nature communications 12, 1779 (2021).DOI
  14. [14]Ganjam, S. et al. Surpassing millisecond coherence in on chip superconducting quantum memories by optimizing materials and circuit design. Nature Communications 15, 3687 (2024).DOI
  15. [15]Bland, M. P. et al. Millisecond lifetimes and coherence times in 2d transmon qubits. Nature 647, 343–348 (2025). URL http://dx.doi.org/10.1038/s41586-025-09687-4.
  16. [16]Willsch, D., Rieger, D. & Winkel, P. e. a. Observation of Josephson harmonics in tunnel junctions. Nature Physics 20, 815–821 (2024). URL https://doi.org/ 10.1038/s41567-024-02400-8.
  17. [17]Beenakker, C. W. J. Universal limit of critical-current fluctuations in mesoscopic Josephson junctions. Physical Review Letters 67, 3836–3839 (1991).pubmed.ncbi.nlm.nih.gov/10044838
  18. [18]Golubov, A. A., Kupriyanov, M. Y. & Il’ichev, E. The current-phase relation in josephson junctions. Rev. Mod. Phys. 76, 411–469 (2004). URL https://link.aps.org/doi/10.1103/RevModPhys.76.411.DOI
  19. [19]Féchant, M. et al. Offset charge dependence of measurement-induced transitions in transmons. Phys. Rev. Lett. 135, 180603 (2025). URL https://link.aps.org/doi/10.1103/yLJV-b4kj.DOI
  20. [20]Catelani, G., Nigg, S. E., Girvin, S. M., Schoelkopf, R. J. & Glazman, L. I. Decoherence of superconducting qubits caused by quasiparticle tunneling. Phys. Rev. B 86, 184514 (2012). URL https://link.aps.org/doi/10.1103/PhysRevB.86.184514.DOI
  21. [21]Kim, J. et al. Emergent harmonics in josephson tunnel junctions due to series inductance (2025). URL https://arxiv.org/abs/2507.08171. arXiv:2507.08171.
  22. [22]Refael, G. & Fisher, D. S. Energy correlations in random transverse field ising spin chains. Phys. Rev. B 70, 064409 (2004). URL https://link.aps.org/doi/10.1103/PhysRevB.70.064409.DOI
  23. [23]Shumeiko, V. S., Bratus, E. N. & Wendin, G. Scattering theory of superconductive tunneling in quantum junctions. Low Temperature Physics 23, 181–198 (1997). URL https://doi.org/10.1063/1.593475.
  24. [24]Lahiri, A., Choi, S.-J. & Trauzettel, B. Origin of the subharmonic gap structure of dc current-biased josephson junctions. Phys. Rev. B 111, L220504 (2025). URL https://link.aps.org/doi/10.1103/zbd5-1cc2.
  25. [25]van der Post, N., Peters, E. T., Yanson, I. K. & van Ruitenbeek, J. M. Subgap structure as function of the barrier in atom-size superconducting tunnel junctions. Phys. Rev. Lett. 73, 2611–2613 (1994). URL https://link.aps.org/doi/10.1103/PhysRevLett.73.2611.DOI
  26. [26]Octavio, M., Tinkham, M., Blonder, G. E. & Klapwijk, T. M. Subharmonic energy-gap structure in superconducting constrictions. Phys. Rev. B 27, 6739–6746 (1983). URL https://link.aps.org/doi/10.1103/PhysRevB.27.6739.DOI
  27. [27]Cuevas, J. C., Martín-Rodero, A. & Levy Yeyati, A. Hamiltonian approach to the transport properties of superconducting quantum point contacts. Physical Review B 54, 7366–7379 (1996).DOI
  28. [28]Flensberg, K., Bindslev Hansen, J. & Octavio, M. Subharmonic energy-gap structure in superconducting weak links. Physical Review B 38, 8707–8711 (1988).pubmed.ncbi.nlm.nih.gov/9945647
  29. [29]Räisänen, I. & Maasilta, I. Improvement of the simmons model for tunnel junctions (2025). URL https://arxiv.org/abs/2511.16337. arXiv:2511.16337.
  30. [30]Simmons, J. G. Electric Tunnel Effect between Dissimilar Electrodes Separated by a Thin Insulating Film. Journal of Applied Physics 34, 2581–2590 (1963).DOI
  31. [31]Oh, J.-S. et al. Correlating aluminum layer deposition rates, josephson junction microstructure, and superconducting qubits’ performance. Acta Materialia 284, 120631 (2025). URL https://www.sciencedirect.com/science/article/pii/S1359645424009790.DOI
  32. [32]Cyster, M. J., Smith, J. S., Vogt, N. et al. Simulating the fabrication of aluminium oxide tunnel junctions. npj Quantum Information 7, 12 (2021). URL https://doi.org/10.1038/s41534-020-00360-4.
  33. [33]Kennedy, O. W., Cole, J. H. & Shelly, C. D. Josephson junction tuning described by depinning physics. Phys. Rev. Lett. 135, 196202 (2025). URL https://link.aps.org/doi/10.1103/xkqs-grnd.DOI
  34. [34]Wolff, O. F. et al. Structural control of two-level defect density revealed by high-throughput correlative measurements of josephson junctions (2026). URL https://arxiv.org/abs/2602.11469. arXiv:2602.11469.
  35. [35]Zhu, Y., Beaudoin, F. & Guo, H. Device variability of josephson junctions induced by interface roughness (2026). URL https://arxiv.org/abs/2602.03037. arXiv:2602.03037.
  36. [36]Zeng, L. J. et al. Direct observation of the thickness distribution of ultra thin alox barriers in al/alox/al josephson junctions. Journal of Physics D: Applied Physics 48, 395308 (2015). URL https://doi.org/10.1088/0022-3727/48/39/395308.
  37. [37]Arute, F. et al. Quantum supremacy using a programmable superconducting processor. Nature 574, 505–510 (2019).DOI
  38. [38]Gambetta, J. M., Chow, J. M. & Steffen, M. Building logical qubits in a superconducting quantum computing system. npj Quantum Information 3, 2 (2017).DOI
  39. [39]Barends, R. et al. Coherent Josephson qubit suitable for scalable quantum integrated circuits. Physical Review Letters 111, 080502 (2013).DOI
  40. [40]Krantz, P. et al. A quantum engineer’s guide to superconducting qubits. Applied Physics Reviews 6, 021318 (2019).DOI
  41. [41]Baran, J. D., Grönbeck, H. & Hellman, A. Mechanism for limiting thickness of thin oxide films on aluminum. Phys. Rev. Lett. 112, 146103 (2014). URL https://link.aps.org/doi/10.1103/PhysRevLett.112.146103.
  42. [42]Cabrera, N. & Mott, N. F. Theory of the oxidation of metals. Reports on Progress in Physics 12, 163–184 (1949).
  43. [43]Aref, T. et al. Characterization of aluminum oxide tunnel barriers by combining transport measurements and transmission electron microscopy imaging. Journal of Applied Physics 116, 073702 (2014). URL https://doi.org/10.1063/1.4893473.
  44. [44]David, H. A. & Nagaraja, H. N. Order Statistics 3rd edn. Wiley Series in Probability and Statistics (Wiley-Interscience, 2003).

Paper details

vlnc:W_mkgqntjw
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY-SA 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents