betaFeedback

WorksPhysics and Astronomy

Atomic-scale Imaging of Time Reversal Breaking Supercurrents in Bi2Sr2CaCu2O8\mathrm{Bi_2Sr_2CaCu_2O_8}

Abstract

A persistent mystery of CuO2\mathrm{CuO_2}-based high temperature superconductivity has been the mechanism and consequences of time reversal (T\mathcal{T})-symmetry breaking. Although theoretical hypotheses focus upon persistent electronic currents as a potential cause, these have been invisible to direct experimental probes. That impasse has recently been removed by development of techniques for atomic-scale supercurrent visualization. Here, by imaging the energy-resolved density of electronic states N(r,E)N(r,E) in Bi2Sr2CaCu2O8\mathrm{Bi_2Sr_2CaCu_2O_8}, we search for the specific signatures of flowing supercurrents in a disordered dd-wave superconductor. Employing the strategy of artificial neural networks (ANN) for complex quantum matter visualization, we designed a dual-ANN (VMAP) to efficiently implement dd-wave supercurrent visualization, achieving imaging of supercurrent speeds V(r)V(r) at magnetic fields ∣B∣≤|B| \leq 50 μμT in Bi2Sr2CaCu2O8\mathrm{Bi_2Sr_2CaCu_2O_8}. Correspondence between specific characteristics of N(r,E)N(r,E) around each supercurrent loop and a quantum "Rosetta Stone" which translates N(r,E)N(r,E) to V(r)V(r), emphasizes the validity of V(r)V(r) detection and visualization. Then, for hole-densities 0.17 ≤p≤\leq p \leq 0.22 we observe statistically repeatable networks of supercurrent flows. Crucially, these supercurrents satisfy Kirchhoff's law at junctions within each network. Further, theoretical predictions for supercurrent speeds VV in disordered dd-wave superconductors are consistent with measured 1250 m/s <V<< V < 4000 m/s, and with reported magnetic field amplitudes ∣B∣≈|B| \approx 0.1 mT. Thus, detection and visualization of these supercurrent loop networks provides unique perspectives on T\mathcal{T}-symmetry breaking in cuprates. It also emphasizes the accelerating utility of ANN techniques for physics discovery from high data volume quantum matter research.

these supercurrent loop networks provides unique perspectives on T\mathcal{T}-symmetry breaking in cuprates. It also emphasizes the accelerating utility of ANN techniques for physics discovery from high data volume quantum matter research[25].

Hole doping antiferromagnetic insulating cuprates, converts the CuO2\mathrm{CuO}_2-plane into an intensely correlated metal[26, 27] whose antiferromagnetic spin-spin interactions apparently subtend high-temperature superconductivity. Yet its Hamiltonian sustains many possible orders, including T\mathcal{T}-symmetry breaking states[28]. Cuprate superconductivity exhibits an order-parameter Δ(r)=Δdeiθ(r)\Delta(\mathbf{r}) = \Delta_d e^{i\theta(\mathbf{r})}, where θ(r)\theta(\mathbf{r}) is the macroscopic quantum phase and Δd\Delta_d exhibits dx2−y2d_{x^2-y^2} symmetry. In these materials, T\mathcal{T}-symmetry breaking has long been reported. For example, zero-field muon spin relaxation (μ\muSR) and neutron diffraction experiments in YBa2Cu3O6+x\mathrm{YBa}_2\mathrm{Cu}_3\mathrm{O}_{6+x} report spontaneous static magnetic fields[2, 5] and in Bi2Sr2CaCu2O8\mathrm{Bi}_2\mathrm{Sr}_2\mathrm{CaCu}_2\mathrm{O}_8 quasistatic magnetic fields[3], into the superconducting phase. Circularly polarized angle resolved photoemission (ARPES) studies, detect spontaneous dichroism in Bi2Sr2CaCu2O8\mathrm{Bi}_2\mathrm{Sr}_2\mathrm{CaCu}_2\mathrm{O}_8 implying underlying T\mathcal{T}-symmetry breaking[4]. Neutron scattering reveals intra-unit-cell magnetic order in YBa2Cu3O6+x\mathrm{YBa}_2\mathrm{Cu}_3\mathrm{O}_{6+x} and HgBa2CuO4\mathrm{HgBa}_2\mathrm{CuO}_4, hypothesized due to T\mathcal{T}-symmetry breaking loop currents[6, 7] into the superconducting phase. Polar Kerr rotation in YBa2Cu3O6+x\mathrm{YBa}_2\mathrm{Cu}_3\mathrm{O}_{6+x} represents a non-reciprocity forbidden without T\mathcal{T}-symmetry breaking[8]. Second harmonic generation and birefringence also indicate breaking of T\mathcal{T}-symmetry[9]. Finally, the Josephson diode effect in Bi2Sr2CaCu2O8\mathrm{Bi}_2\mathrm{Sr}_2\mathrm{CaCu}_2\mathrm{O}_8 pristine/twist junctions requires T\mathcal{T}-symmetry breaking superconductivity[10, 11]. Because the long-range magnetic orders are absent in these systems, persistent atomic-scale supercurrents are hypothesized[12] as the cause of T\mathcal{T}-symmetry breaking. But no direct evidence of such supercurrents has ever been identified.

Theoretical hypotheses for the mechanism of cuprate T\mathcal{T}-symmetry breaking are legion[12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23]. Crystalline disorder, at surfaces or grain boundaries[14, 15, 23], is predicted to break T\mathcal{T}-symmetry. However, such crystal defects are not observed in our studies. Another possibility is spatially homogeneous T\mathcal{T}-symmetry breaking due to a mixed singlet-triplet pairing state with d+ipd+ip symmetry[16]. Alternatively, there are predictions that intra-unit-cell persistent loop currents in the pseudogap phase, not only break T\mathcal{T}-symmetry[17, 18], but also generate currents flowing along the edges of domains of differing intra-unit-cell ordering[18]. Figure 1a is a schematic of current trajectories within a disordered intra-unit-cell loop-current system[18]. More recently, studies of highly disordered dd-wave superconductors predict T\mathcal{T}-symmetry breaking[19, 20, 21, 22]. Here, misorientation of the quantum phase difference between adjacent superconductive domains generates supercurrent loops due to random-in-sign Josephson couplings[21] (Fig. 1b). Alternatively, concentrations of intense scattering by impurity or dopant atoms may generate d+isd+is superconductivity with Δdeiϕd+Δseiϕs\Delta_d e^{i\phi_d}+\Delta_s e^{i\phi_s} symmetry surrounding impurity sites, and thus supercurrents breaking T\mathcal{T}-symmetry[22, 23] (Fig. 1c). Evaluation by direct visualization of any of these concepts has been thwarted by the absence of techniques for atomic scale visualization of supercurrents.

Recently, however, scanning tunneling microscopy techniques for atomic-scale supercurrent visualization[24] have been developed. Here, the energy spectrum EkE_{\mathbf{k}} of Bogoliubov quasiparticle states ∣k⟩|\mathbf{k}\rangle in a superconductor is Ek=±E2(k)+ΔS2E_{\mathbf{k}}=\pm\sqrt{E^2(k)+\Delta_S^2}, where E(k)E(k) is the normal-state band structure referenced to the Fermi energy and ΔS\Delta_S is the electron-pairing order parameter. In a laboratory/crystal frame relative to which supercurrent flows at velocity V(r)≠0\mathbf{V}(\mathbf{r})\ne0, each EkE_{\mathbf{k}} receives a Galilean energy-boost[24, 29, 30, 31] yielding a new spectrum:

Ek′=±E2(k)+ΔS2+ℏk⋅V(r)(1)E_{\mathbf{k}}^{\prime}=\pm\sqrt{E^2(k)+\Delta_S^2+\hbar\mathbf{k}\cdot\mathbf{V}(\mathbf{r})} \tag*{(1)}

Figure 1d represents schematic consequences of Eqn. (1) for a dd-wave superconductor. Importantly, the specifics of each consequent N(E)N(E) spectrum depend on the vectorial direction of the supercurrent flow. Considering a dx2−y2d_{x^2-y^2} symmetry superconductor (Fig. 1e) the resultant N(E)N(E) spectra for supercurrents at angles 00, π/8\pi/8, π/4\pi/4 and 3π/83\pi/8 relative to the order parameter and for different supercurrent speeds V\mathbf{V}, are shown in Figs. 1f-i. This is a quantum “Rosetta Stone” translating theoretic spectral shapes N(r,E)N(\mathbf{r}, E) into supercurrent speeds V(r)\mathbf{V}(\mathbf{r}). Although based upon semiclassical theory, during development of our technique within that approximation[24] we demonstrated quantitatively correct supercurrent speeds V\mathbf{V} at R≳2ξR \gtrsim2\xi, where ξ\xi is the superconductor’s coherence length and RR is the scale associated with supercurrent variations[24]. For Bi2Sr2CaCu2O8\mathrm{Bi}_{2}\mathrm{Sr}_{2}\mathrm{Ca}\mathrm{Cu}_{2}\mathrm{O}_{8}, direct visualization reveals a true vortex core (not a vortex halo) where the order parameter reaches zero with a coherence length ξ≈0.7 nm\xi\approx0.7\,\mathrm{nm}, while the typical region surrounded by a supercurrent (Figs. 2, Figure 3, 4 below) has diameter d≈3 nmd \approx3\,\mathrm{nm} so that the semiclassical approximation appears not invalid. Thus, supercurrent speeds V(r)≡∣V(r)∣\mathbf{V}(\mathbf{r}) \equiv\lvert\mathbf{V}(\mathbf{r})\rvert up to 4000 m/s may, in principle, be visualized via distinct characteristics of the measured differential conductance g(r,E)g(\mathbf{r}, E) spectrum translated to V(r)\mathbf{V}(\mathbf{r}) using the “Rosetta Stone” in Fig. 1. Here we exploit this new strategy to detect and visualize T\mathcal{T}-symmetry breaking supercurrents in the canonical CuO2\mathrm{CuO}_{2}-based superconductor, Bi2Sr2CaCu2O8\mathrm{Bi}_{2}\mathrm{Sr}_{2}\mathrm{Ca}\mathrm{Cu}_{2}\mathrm{O}_{8}.

Figure 3.

High-data-volume spectroscopic imaging scanning tunneling microscopy (SISTM) is used for visualization of the density-of-electronic-states N(r,E)N(\mathbf{r}, E). The Bi2Sr2CaCu2O8\mathrm{Bi}_{2}\mathrm{Sr}_{2}\mathrm{Ca}\mathrm{Cu}_{2}\mathrm{O}_{8} samples 0.17≲p≲0.220.17 \lesssim p \lesssim0.22 are inserted into a cryogenic SISTM, cleaved at T<10T<10 K, and inserted in the STM scan head (Methods A). The differential tunneling conductance spectrum dI/dV(r,E)≡g(r,E)∝N(r,E)dI/dV(\mathbf{r}, E) \equiv g(\mathbf{r}, E) \propto N(\mathbf{r}, E) allows atomic-scale imaging of N(r,E)N(\mathbf{r}, E), whose severe chaoticity is well known[32, 33, 34], rendering subtle signatures of circulating supercurrents (e.g. Figs. 1f-i) impossible to discern. Indeed, analysis of such spectra at all spatial locations r\mathbf{r} using a dx2−y2d_{x^{2}-y^{2}} version of Eqn. (1) and for all possible V(r)\mathbf{V}(\mathbf{r}), has long proven numerically intractable (Methods B).

Today, however, artificial neural network (ANN) analysis of variegated high-volume experimental data sets[25] represents a promising opportunity to detect and visualize cuprate supercurrents. Consider the sequence of spectra in a dd-wave superconductor in Figs. 1f-i: clearly, to detect such subtle phenomena one needs to evaluate all g(r,E)g(\mathbf{r},E) in terms of the energy gap Δ(r)\Delta(\mathbf{r}), the supercurrent speed V(r)V(\mathbf{r}) and the supercurrent trajectory L(r)L(\mathbf{r}). We have developed a dual-ANN system dubbed VMAP for that specific task (Methods C1). VMAP consists of two modules: MAIN and COGNATE, with MAIN being the initial analysis step and COGNATE being the final precision step. To train VMAP we use a Bi2Sr2CaCu2O8+x\mathrm{Bi}_{2}\mathrm{Sr}_{2}\mathrm{Ca}\mathrm{Cu}_{2}\mathrm{O}_{8+x} normal state band dispersion:

E(k)=μ−2t1(cos⁡kx+cos⁡ky)+4t2cos⁡kxcos⁡ky−2t3(cos⁡2kx+cos⁡2ky)(2)E(k) = \mu- 2t_{1}(\cos k_{x} + \cos k_{y}) + 4t_{2}\cos k_{x}\cos k_{y} - 2t_{3}(\cos2k_{x} + \cos2k_{y}) \tag*{(2)}

where μ\mu is the chemical potential, t1t_{1}, t2t_{2}, and t3t_{3} are electron hopping parameters. The superconducting gap function is:

Δ(k)=Δ2(cos⁡kx−cos⁡ky)(3)\Delta(k) = \frac{\Delta}{2}(\cos k_{x} - \cos k_{y}) \tag*{(3)}

where Δ\Delta is the maximum gap amplitude. For a supercurrent with velocity V=(Vx,Vy)\mathbf{V} = (V_{x},V_{y}), the Bogoliubov band dispersion is modified by the Galilean boost EG(k)=ℏ(Vxkx+Vyky)E_{G}(k) = \hbar(V_{x}k_{x} + V_{y}k_{y}) yielding a Bogoliubov-DeGennes Hamiltonian[35]:

HGB=[E(k)+EG(k)Δ(k)Δ(k)∗−E(−k)+EG(k)](4)H_{GB} = \begin{bmatrix} E(k) + E_{G}(k) & \Delta(k) \\ \Delta(k)^{*} & -E(-k) + E_{G}(k) \end{bmatrix} \tag*{(4)}

The Green’s function for eigenstates of HGBH_{GB} is:

G(ω,k,Δ,Vx,Vy)=[ωI−HGB]−1=A[ω+E(−k)−EG(k)Δ(k)Δ(k)∗ω−E(k)−EG(k)](5)G(\omega,k,\Delta,V_{x},V_{y}) = [\omega I - H_{GB}]^{-1} = A \begin{bmatrix} \omega+ E(-k) - E_{G}(k) & \Delta(k) \\ \Delta(k)^{*} & \omega- E(k) - E_{G}(k) \end{bmatrix} \tag*{(5)}

where

A=1(ω−E(k)−EG(k))(ω+E(−k)−EG(k))−Δ2(k)(6)A = \frac{1}{(\omega- E(k) - E_{G}(k))(\omega+ E(-k) - E_{G}(k)) - \Delta^{2}(k)} \tag*{(6)}

We employ an energy ℏω=E+iΓ\hbar\omega= E + i\Gamma with a quasiparticle lifetime parameter ℏ/Γ\hbar/\Gamma[34]. The density of states N(E)N(E) can then be derived from the imaginary part of the Green’s function as:

N(E,Δ,Vx,Vy,Γ)≡−12πIm⁡[∑kTr⁡(G(ω,k,Δ,Vx,Vy))](7)N(E,\Delta,V_{x},V_{y},\Gamma) \equiv-\frac{1}{2\pi} \operatorname{Im}\left[\sum_{k}\operatorname{Tr}(G(\omega,k,\Delta,V_{x},V_{y}))\right] \tag*{(7)}

Quantitively accurate simulations of the Galilean boost effects in Eqn. (7) for the dx2−y2d_{x^{2}-y^{2}} superconductivity of Bi2Sr2CaCu2O8\mathrm{Bi}_{2}\mathrm{Sr}_{2}\mathrm{Ca}\mathrm{Cu}_{2}\mathrm{O}_{8} are shown in Figs. 1f-i for various supercurrent velocities V(r)\mathbf{V}(\mathbf{r}) at angles 00, π/8\pi/8, π/4\pi/4, and 3π/83\pi/8 relative to the order parameter. But, from symmetry, Eqn. (7) results in 8 supercurrent velocity vectors (±Vx,±Vy)(\pm V_{x}, \pm V_{y}), (±Vy,±Vx)(\pm V_{y}, \pm V_{x}) all having the same N(E)N(E) (Methods C). Thus we introduce the quantities VH=Max⁡(∣Vx∣,∣Vy∣)V_{H} = \operatorname{Max}(|V_{x}|, |V_{y}|) and VL=Min⁡(∣Vx∣,∣Vy∣)V_{L} = \operatorname{Min}(|V_{x}|, |V_{y}|) being the maximal and minimal components of supercurrent velocity vector V\mathbf{V}.

For training the MAIN and COGNATE models, we generate 3.5 million N(E,Δ,VH,VL,Γ)N(E, \Delta, V_{H}, V_{L}, \Gamma) spectra using Eqn. (7) with energy in the range [−100 meV, 100 meV][-100\ \mathrm{meV},\ 100\ \mathrm{meV}], Δ\Delta in the range [0 meV, 80 meV][0\ \mathrm{meV},\ 80\ \mathrm{meV}], Γ\Gamma in the range [0 meV, 10 meV][0\ \mathrm{meV},\ 10\ \mathrm{meV}], VHV_{H} and VLV_{L} in the range [0, 5000 m/s][0,\ 5000\ \mathrm{m/s}] (Methods C). Tight-binding parameters for E(k)E(k) are from angle resolved photoemission studies[36]. This synthetic dataset of 3.5 million simulated N(E,Δ,VH,VL,Γ)N(E, \Delta, V_{H}, V_{L}, \Gamma) is used to train both the MAIN and COGNATE models of VMAP.

Experimentally, g(E)g(E) spectra are typically of the form g(E)=(a+bE)×N(E,Δ,VH,VL,Γ)g(E) = (a + bE) \times N(E, \Delta, V_{H}, V_{L}, \Gamma), where both aa and bb are small parameters and unavoidable electronic noise also exists. Thus, Gaussian noise and a linear-EE background were added as augmentation to each Eqn. (7) simulated N(E)N(E) when training MAIN, with the noise level consistent with that measured in experimental g(r,E)g(\mathbf{r}, E) data (Methods C3). In operation, the MAIN module determines four primary parameters Δ(r)\Delta(\mathbf{r}), VH(r)V_{H}(\mathbf{r}), VL(r)V_{L}(\mathbf{r}), and Γ(r)\Gamma(\mathbf{r}), with the consequent supercurrent speed being V(r)≡VH2(r)+VL2(r)V(\mathbf{r}) \equiv\sqrt{V_{H}^{2}(\mathbf{r}) + V_{L}^{2}(\mathbf{r})} (Methods C4). To establish best correspondence of VMAP determined N(r,E)N(\mathbf{r}, E) with each experimental spectrum g(r,E)g(\mathbf{r}, E), the parameters Δ(r)\Delta(\mathbf{r}), VH(r)V_{H}(\mathbf{r}), VL(r)V_{L}(\mathbf{r}), and Γ(r)\Gamma(\mathbf{r}) are used by the COGNATE module. This backpropagates gradients of the physical parameters efficiently, allowing precise evaluation of N(r,E)N(\mathbf{r}, E) (Methods C5). Ultimately, the quality of fit parameter R2(r)R^{2}(\mathbf{r}) is evaluated by comparing each measured g(r,E)g(\mathbf{r}, E) to its VMAP reported N(r,E)N(\mathbf{r}, E) spectrum. Our MAIN+COGNATE process increases the FOV-average to R2≳0.95R^{2} \gtrsim0.95 in general, as demonstrated directly using data videos in Supplementary. Each empirical data set consists of a Bi2Sr2CaCu2O8\mathrm{Bi}_{2}\mathrm{Sr}_{2}\mathrm{Ca}\mathrm{Cu}_{2}\mathrm{O}_{8} topographic image T(r)T(\mathbf{r}) with a simultaneous and atomically registered g(r,E)g(\mathbf{r}, E) (Methods G). VMAP analysis of g(r,E)g(\mathbf{r}, E) then yields images of the superconducting energy gap Δ(r)\Delta(\mathbf{r}) and of supercurrents V(r)≡VH2(r)+VL2(r)V(\mathbf{r}) \equiv\sqrt{V_H^2(\mathbf{r}) + V_L^2(\mathbf{r})}. The validity of determination of N(r,E)N(\mathbf{r}, E), and thus of Δ(r)\Delta(\mathbf{r}) and V(r)V(\mathbf{r}), is exemplified directly in the Supplementary using three-panel videos of simultaneous g(r,E):N(r,E):∣g(r,E)−N(r,E)∣g(\mathbf{r}, E):N(\mathbf{r}, E):|g(\mathbf{r}, E)-N(\mathbf{r}, E)| which demonstrate high-quality VMAP fits to g(r,E)g(\mathbf{r}, E), throughout. Using a second ANN named V0MAP (Methods D) for which all training spectra are set at V=0V=0, we demonstrate by comparing R2(r)R^2(\mathbf{r}) for all N(r,E):g(r,E)N(\mathbf{r}, E):g(\mathbf{r}, E) correspondences from both ANNs, that supercurrent speeds V≲1250V \lesssim1250 m/s are indistinguishable from zero when using VMAP (Methods E). Nevertheless, the V≳1250V \gtrsim1250 m/s supercurrent trajectories in V(r)V(\mathbf{r}) can be visualized directly using

L(x,y)=−[1+α(∂V∂x)2][α∂2V∂y2−2α2(∂V∂x)(∂V∂y)∂2V∂x∂y+[1+α(∂V∂y)2]α∂2V∂x2][1+α(∂V∂x)2+α(∂V∂y)2]3/2(8)L(x,y)=-\frac{\left[1+\alpha\left(\frac{\partial V}{\partial x}\right)^2\right]\left[\alpha\frac{\partial^2 V}{\partial y^2}-2\alpha^2\left(\frac{\partial V}{\partial x}\right)\left(\frac{\partial V}{\partial y}\right)\frac{\partial^2 V}{\partial x\partial y}+\left[1+\alpha\left(\frac{\partial V}{\partial y}\right)^2\right]\alpha\frac{\partial^2 V}{\partial x^2}\right]}{\left[1+\alpha\left(\frac{\partial V}{\partial x}\right)^2+\alpha\left(\frac{\partial V}{\partial y}\right)^2\right]^{3/2}} \tag*{(8)}

(Methods F) designed to identify the spatial locations where intense supercurrents appear. Finally, for p<0.17p<0.17, the VMAP system does not return valid results because the ANN training model (Eqns. (2)-(7)) is demonstrably untenable for underdoped cuprates.

Exploiting these ANN techniques, we studied the pp-dependence of Δ(r)\Delta(\mathbf{r}), VH(r)V_H(\mathbf{r}), VL(r)V_L(\mathbf{r}), and Γ(r)\Gamma(\mathbf{r}) phenomena. Focusing on 0.17≲p≲0.220.17 \lesssim p \lesssim0.22 with relevant T(r)T(\mathbf{r}) shown Methods G, the results of imaging Δ(r)\Delta(\mathbf{r}) by VMAP are shown in Figs. 2a-c. These are highly consistent with conventional heterogeneous Δ(r)\Delta(\mathbf{r}) studies[32, 33, 34]. VMAP imaging of V(r)V(\mathbf{r}) in the FOV associated with each Δ(r)\Delta(\mathbf{r}) at left, generates the results shown in Figs. 2d-f. These images are remarkable, revealing the direct signature of persistently flowing supercurrents at nanometer scale in a canonical dd-wave cuprate superconductor, with speeds reaching 4000 m/s. The statistical repeatability of V(r)V(\mathbf{r}) phenomena is manifest. Separately visualized V+(r)V^{+}(\mathbf{r}) using g(r,E>0)g(\mathbf{r}, E>0) and V−(r)V^{-}(\mathbf{r}) using g(r,E<0)g(\mathbf{r}, E<0) reveals them indistinguishable, consistent with the Hamiltonian in Eqn. (4) (Methods H) and demonstrating the insensitivity to noise. A network of intense supercurrents with patches of mean diameter 2.7±0.92.7\pm0.9 nm (Methods I) is revealed in all cases.

The spatial resolution is fixed at a typical 0.2 nm pixel size, which is far below the observed patch size. The supercurrents are most easily visualized in Figs. 2g-i, showing the network of supercurrent loop trajectories L(r)L(r) associated with each panel at left.

For Galilean boosts to N(E)N(E) (Fig. 1d) very specific characteristics should occur for different supercurrent directions relative to the dx2−y2d_{x^{2}-y^{2}} symmetry order parameter (Figs. 1f-i and Methods J). In Figure 3a we explore these characteristics for numerous supercurrent loop trajectories identified by L(r)L(r) imaging. The color of dots indicates different directions relative to the order parameter. Figure 3b shows the R2(r)R^{2}(r) evaluated by VMAP in the same FOV of Figure 3a. In Figure 3c we show explicitly the predicted sequence of spectra for a circular supercurrent loop of either chirality. Figure 3d-Figure 3i then report the sequence of measured g(r,E)g(r,E) spectra on typical loop trajectories labeled 1#-6# in Figure 3a. Their detailed correspondence with Figure 3c underlines the validation of supercurrent speed V(r)V(r) imaging (Figure 3d-Figure 3i) using VMAP. Methods K and Extended Data Figure 10 show the equivalent correspondence as a function of pp.

Examples of individual supercurrent loop spectra

Extended Data Figure 10. Examples of individual supercurrent loop spectra

Supercurrent loops ought to exhibit yet another key characteristic, that of satisfying a continuity equation. We thus consider junctions where three supercurrents meet, and examine if quantitively measured supercurrents entering/leaving the junction satisfy Kirchhoff’s law. We mask the V(r)V(r) images (e.g. Figs. 2d-f) by simultaneous images of loop trajectories L(r)L(r) (e.g. Figs. 2g-i respectively), to produce a new image LV(r)LV(r) focusing on the supercurrent speeds along each trajectory in the network. A typical example is shown in Fig. 4a, revealing many tri-junctions and the measured supercurrent speeds approaching and departing them. Amazingly, the sum of two supercurrents adds to the value of the third in all well-defined tri-junctions detected within LV(r)LV(r) for all samples. Typical examples of this significant phenomenon are shown in Figs. 4b-h. More generally, a global supercurrent continuity constraint is explored using Si≡∑jVjS_i \equiv\sum_j V_j, where jj are the four x/yx/y nearest neighbors of site ii and VjV_j values can be any of (±VH,±VL)(\pm V_H,\pm V_L). Evaluating continuity, the objective is to determine if the sum of supercurrents entering and leaving any given pixel can add to 0, requiring a search for the minimum value of SiS_i labeled S(r)S(r). (Methods L). The measured S(r)S(r) results are shown in Figs. 4i-k and comparable quantitively with the supercurrent speeds from which they are derived in Figs. 2d-f respectively. Here the mean value of S(r)S(r) is below 20 m/s although supercurrent speeds in Figs. 2d-f reach up to 4000 m/s. This implies that supercurrent continuity is maintained throughout the tortuously heterogeneous V(r)V(r) images (Figs. 2d-f) to within approximately 1%. Throughout Fig. 4 one emphasizes that, if the VMAP ANN were dominated by poor SNR, or incorrect evaluation of V(r)V(r) from N(r,E)N(r,E), or conflation of heterogeneities from another degree of freedom (e.g. Δ(r)\Delta(r) or Γ(r)\Gamma(r)) into V(r)V(r), there would be no reason whatsoever for V(r)V(r) to satisfy either a global continuity equation in S(r)S(r) (Figs. 4i-k) or the local Kirchhoff’s law constraint (Figs. 4b-h). Hence, these results sustain VMAP derived V(r)V(r) as images of persistent supercurrents.

The long-standing challenge has been to identify and understand T\mathcal{T}-symmetry breaking in cuprates[2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23]. With respect to disordered intra-unit-cell loop-current systems sustaining domain wall supercurrents[17, 18], our results are perhaps not relevant because we focus on overdoped cuprates. For such samples, a nanoscale network of supercurrents induced by disorder in dd-wave superconductors[12, 19, 20, 21, 22] seems more consistent with direct supercurrent visualization (Figs. 2-4). Revealed is a novel superconductive state heterogeneously breaking T\mathcal{T}-symmetry somewhat akin to a spin-glass phase, but consisting of a disordered network of supercurrent loops, whose chirality cannot yet be determined. These networks exhibit speeds of several thousand m/s. While subsurface magnetic impurity effects can be ruled out on the basis of their minimal induced supercurrents (Methods I), spontaneous loop currents on these scales were predicted by both Refs. 21 and 22, using mean-field simulations for a disordered dd-wave superconductive state. Within closed loops of approximate diameter 3 nm, the observed supercurrents satisfy the circulation quantization constraint. Reference 22 predicts correct supercurrent speeds, loop sizes, and formation of supercurrent ‘mesas’, consistent with observations. However, different microscopic origins are proposed for the disordered loop current state in Ref. 21 and Ref. 22. At present, while it is possible to sustain the physical validity of this general class of supercurrent T\mathcal{T}-symmetry breaking theories, it is not possible to sharply distinguish between them. Theoretical advances will be required to achieve that aim.

Nevertheless, within the context of Eqns. (4)-(7) the empirics in Figs. 2-4 provide a radically new perspective on the physics of T\mathcal{T}-symmetry breaking in highly doped superconducting cuprates. First, an hypothetical term EG(k)E_G(k) in Eqn. (4) that is consistent with the experiments (Figs. 2-4) has few if any other possible causes except a Galilean energy boost due to supercurrents. Second, when we separately visualize V+(r)V^{+}(r) using the positive bias g(r,E>0)g(r,E>0) data and V−(r)V^{-}(r) using the negative bias g(r,E<0)g(r,E<0) data, they are practically indistinguishable (Methods H) emphasizing both the particle-hole symmetry of V(r)V(r) and the adequacy of the experimental signal to noise ratio. Third, the correspondence between g(r,E)g(r,E) spectra measured along the trajectory of supercurrent loops (Figure 3) to the supercurrent “Rosetta Stone” (Figs. 1e-i) provides substantiation of VMap’s supercurrent speed V(r)V(r) visualization. Fourth, circulation is observably quantized around close supercurrent loops. Fifth, from theoretic models with Fermi velocity VF≈3.5×105V_F \approx3.5 \times10^{5} m/s, a typical disordered dd-wave supercurrent velocity is predicted[22] to be V≈0.01VFV \approx0.01V_F, in good agreement observations (Fig. 2). Sixth, approximate estimates of the absolute value of loop supercurrents II based on the measured V(r)V(r) and estimated electron-pair density, yield I∼10−7I\sim10^{-7} A, so that that for ∼1.5\sim1.5 nm radius domains the magnetic field at center is ∣B∣∼0.1|B|\sim0.1 mT (Methods I). This is in reasonable agreement with reported spontaneous fields of T\mathcal{T}-symmetry breaking in cuprates. Finally, the supercurrent images V(r)V(r) are consistent with global continuity constraints (Figs. 4i-k) and locally satisfy Kirchhoff’s law at all valid tri-junctions (Figs. 4b-h). Thus, Figs. 2-4 provide novel empirics on T\mathcal{T}-symmetry breaking in cuprates. They also emphasized the accelerating utility ANN techniques for physics discovery[25] in high data volume quantum matter research. Ultimately, this general capability to visualize supercurrent loops may be employed to resolve outstanding mysteries of T\mathcal{T}-symmetry breaking in a variety of other superconductor classes[12].

Acknowledgements: We acknowledge and thank C. Varma for key motivation and guidance for this project. We thank N. Kennedy and A. Samoilenka for key advice.

Funding: P.J.H acknowledges support from the Leverhulme Foundation and NSF grant NSF-DMR-2231821. Shuqiu W. acknowledges support from the Engineering and Physical Sciences Research Council (EPSRC) under Award EP/Z53660X/1 and the support from the Royal Academy of Engineering / Leverhulme Trust Research Fellowship. B.H. acknowledges support from Research Ireland under GOI Postdoctoral Fellowship GOIPD/2026/2219. K.Z. acknowledges EU Horizon Europe MSCA funding under grant agreement No. 101203931 (QUASAR). J.C.S.D. acknowledges support from the Moore Foundation’s EPiQS Initiative through Grant GBMF9457. B.H. and J.C.S.D. acknowledge support from the UK Royal Society under Award R64897. Siyuan. W., H.Z., C.D., B.H, J.W., and J.C.S.D. acknowledge support from Research Ireland under Awards SFI 17/RP/5445 and RI 25/RP-2TF/13559

Competing interest: The authors declare no competing interests.

Contact details for corresponding authors: [email protected] (Bin Hu), [email protected] (J.C. Séamus Davis), [email protected] (Peter J. Hirschfeld).

Figure 1 dd−wave Galilean boost to Bogoliubov density of states from supercurrents

Image not extracted

Figure 1. dd−wave Galilean boost to Bogoliubov density of states from supercurrents. a. Schematic diagram of domain wall supercurrents within disordered intra-unit-cell loop-current systems. b. Schematic diagram of nanoscale superconductive phase-heterogeneity induced supercurrents for dx2−y2d_{x^2-y^2} symmetry superconductors. c. Schematic diagram of strong impurity scattering induced d+isd+is supercurrents, for dx2−y2d_{x^2-y^2} symmetry superconductors. d. Schematic of the Galilean boost effect on Bogoliubov quasiparticle energy and consequent alteration of N(E)N(E), due to supercurrents with speed V⃗>0\vec{V}>0. e. Schematic of a supercurrent loop oriented relative to the order parameter of a dx2−y2d_{x^2-y^2}superconductor. The predicted N(E)N(E) is different for each supercurrent direction and speed as shown in panels f-i. This is the “Rosetta Stone” translating each N(E)N(E) spectrum to a specific speed V, and froms the training set for VMAP. f. N(E)N(E) spectra from Eqns. (2)-(7) for supercurrents with different speed VV at an angle 0 to dd−wave order parameter. g. N(E)N(E) spectra from Eqns. (2)-(7) for supercurrents with different speed VV at an angle π/8\pi/8 to dd−wave order parameter. h. N(E)N(E) spectra from Eqns. (2)-(7) for supercurrents with different speed VV at an angle π/4\pi/4 to dd−wave order parameter. i. N(E)N(E) spectra from Eqns. (2)-(7) for supercurrents with different speed VV at an angle 3π/83\pi/8 to dd−wave order parameter.

Figure 2 VMAP determination of Δ(r)\Delta(\mathbf{r}), V(r)V(\mathbf{r}) and L(r)L(\mathbf{r}) versus hole density pp

Figure 2.

a-c Imaging energy gaps Δ(r)\Delta(\mathbf{r}) versus hole-density pp using VMAP. Relevant topographies are shown in Methods G and Extended Data Fig. 6. Data demonstrating the overall quality of g(r,E)g(\mathbf{r}, E) by VMAP is presented using data videos in Supplementary.

d-f Imaging supercurrent speeds V(r)V(\mathbf{r}) in the same FOV as associated Δ(r)\Delta(\mathbf{r}) panel to left, versus hole-density pp using VMAP.

g-i Imaging the network of supercurrent loops L(r)L(\mathbf{r}) in the same FOV as associated Δ(r)\Delta(\mathbf{r}) panel at left, versus hole-density pp using VMAP, where σr=4\sigma_r=4 Å, α=0.30\alpha=0.30 (ns)2^2.

Figure 3 Individual supercurrent loops as identified by VMAP

a. Supercurrent speeds on designated current loops 1#1^{\#}–6#6^{\#} in L(r)L(\mathbf{r}). Color code is as in Fig. 1e, designating distinct directions of supercurrent. There is an ambiguity as to whether the supercurrents are flowing clockwise or anti clockwise.

b. Goodness of fit R2(r)R^2(\mathbf{r}) of all VMAP N(r,E)N(\mathbf{r}, E) to measured g(r,E)g(\mathbf{r}, E), simultaneously measured with a. The normalized cross correlation coefficient between a and b is 0.03 demonstrating the typical result that there is virtually no spatial correspondence between R2(r)R^2(\mathbf{r}) and L(r)L(\mathbf{r}) so that the rare examples of deficient fitting do not falsely create the images of supercurrent speeds.

c. Sequence of theoretical N(r,E)N(\mathbf{r}, E) spectra on the trajectory of a supercurrent loop. Color code is as in Fig. 1e, designating distinct directions of supercurrent.

d. Sequence of g(r,E)g(\mathbf{r}, E) spectra on the trajectory of loop 1#1^{\#}.

e. Sequence of g(r,E)g(\mathbf{r}, E) spectra on the trajectory of loop 2#2^{\#}.

f. Sequence of g(r,E)g(\mathbf{r}, E) spectra on the trajectory of loop 3#3^{\#}.

g. Sequence of g(r,E)g(\mathbf{r}, E) spectra on the trajectory of loop 4#4^{\#}.

h. Sequence of g(r,E)g(\mathbf{r}, E) spectra on the trajectory of loop 5#5^{\#}.

i. Sequence of g(r,E)g(\mathbf{r}, E) spectra on the trajectory of loop 6#6^{\#}. The spectra in d-i are vertically shifted for clarity and have the corresponding speed listed in m/s.

Figure 1 (PDF p. 16).

Panels a–k showing $L V(r)$ and $S(r)$ images with velocity scales

Figure 4 (PDF p. 19).

Methods

Materials and measurements

Single crystals of Bi2Sr2CaCu2O8\mathrm{Bi}_{2}\mathrm{Sr}_{2}\mathrm{Ca}\mathrm{Cu}_{2}\mathrm{O}_{8} were grown by the traveling solvent-floating-zone (TSFZ) method. The spectroscopic imaging scanning tunneling microscope (SISTM) measurements of Bi2Sr2CaCu2O8\mathrm{Bi}_{2}\mathrm{Sr}_{2}\mathrm{Ca}\mathrm{Cu}_{2}\mathrm{O}_{8} crystals were performed on a custom instrument at 4.2 K. All Bi2Sr2CaCu2O8\mathrm{Bi}_{2}\mathrm{Sr}_{2}\mathrm{Ca}\mathrm{Cu}_{2}\mathrm{O}_{8} samples were cleaved in situ in cryogenic ultra-high vacuum at T<10T<10 K and then immediately transferred into the STM head. All SISTM measurements were conducted using a W scan tip, well characterized on a Au single crystal, before conducting measurements on Bi2Sr2CaCu2O8\mathrm{Bi}_{2}\mathrm{Sr}_{2}\mathrm{Ca}\mathrm{Cu}_{2}\mathrm{O}_{8} samples.

Method on supercurrent velocity analysis

The Galilean-boost energy EG(k)=ℏ(Vxkx+Vyky)E_G(k)=\hbar(V_xk_x+V_yk_y) is kk-dependent, while the density of states N(E)N(E) integrates over the whole BZ in kk-space, meaning that 8 supercurrents result in the same N(E)N(E): (±Vx,±Vy)(\pm V_x,\pm V_y), (±Vy,±Vx)(\pm V_y,\pm V_x) adding difficulty to the analysis. Earlier, we demonstrated a supercurrent visualization method[24] using a reference dataset, where each of the experimental spectra g(E)g(E) is compared individually to all of the reference dataset N(E)N(E), and the maximal coefficient of determination R2R^2 produces the best matched parameters. However, with four-parameter dd-wave supercurrent fitting as in this project, such a reference dataset would contain ∼\sim1.6 billion spectra. Hence this approach is inadequate for high data volume and heterogeneous cuprate g(r,E)g(r,E) data sets of a disordered dd-wave superconductor.

VMAP ANN architecture

To tackle the eight possible N(E)N(E) for each supercurrent velocity V\mathbf{V}, we introduce two quantities, namely the higher (lower) vectorial components VHV_H (VLV_L) of V\mathbf{V} such that V=(VHx^,VLy^)\mathbf{V}=(V_H\hat{x},V_L\hat{y}) or V=(VLx^,VHy^)\mathbf{V}=(V_L\hat{x},V_H\hat{y}). Thus

VH=Max⁡(∣Vx∣,∣Vy∣)(9)V_H=\operatorname{Max}(|V_x|,|V_y|) \tag*{(9)}
VL=Min(∣Vx∣,∣Vy∣)(10)V_L = \mathit{Min}(\lvert V_x\rvert,\lvert V_y\rvert) \tag*{(10)}

where Max\mathit{Max} and Min\mathit{Min} are referred to as taking the maximal and minimal values, respectively.

For a given parameter set (Γ,Δ,VH,VL)(\Gamma,\Delta,V_H,V_L), the N(E)N(E) within the energy range [E1,E2][E_1,E_2] can then be sharply defined. We use this as the reference dataset randomly sampling these four parameters. Next, we introduced a dual-artificial-neural-network-based method, namely VMAP, comprising a MAIN and a COGNATE. In practice, 3.5 million references N(E)N(E) curves are sufficient to train VMAP.

For MAIN, we aim to take the measured g(E)g(E) spectra as input and the four parameters (Γ,Δ,VH,VL)(\Gamma,\Delta,V_H,V_L) as output. However, this is a “one-to-many” mapping task, especially when the signal-to-noise ratio of input curve is inadequate, in which multiple parameter sets (Γ,Δ,VH,VL)(\Gamma,\Delta,V_H,V_L) can produce similar N(E)N(E) curve. To improve the performance of MAIN, we design a COGNATE that identifies the N(E)N(E) spectra from the MAIN output parameter set that corresponds best to the input g(E)g(E).

The COGNATE adopts a transformer-like architecture37 (Extended Data Figure S1). Each of the four parameters (Γ,Δ,VH,VL)(\Gamma,\Delta,V_H,V_L) is projected into the hidden space by an independent fully connected (FC) layer, and the four resulting representations are then fed into the self-attention block (Extended Data Figure S1). The energies corresponding to the sample bias in the experimental g(E)g(E) spectra are projected into the hidden space via an FC layer as well with one energy per token and then used as input to self-attention. In the cross-attention section, the energy representations serve as queries, while the encoded parameter representations provide both keys and values. This decoder block (dashed rectangle in Extended Data Figure S1) is repeated four times. Each block comprises a self-attention block (SAB, SAB(x)=x+SA(norm⁡(x))SAB(x) = x + SA(\operatorname{norm}(x)), where SASA is computing the scaled dot-product attention) over the projected energy, cross-attention, and a feedforward neural-network (FFN). A final FC layer maps the decoded energy representations to the predicted density of states N(E)N(E).

Supplementary Fig. 1. $L(r)$ maps at different $\sigma_{r}$ and $\alpha$

Figure S1. Supplementary Fig. 1. L(r)L(r) maps at different σr\sigma_{r} and α\alpha (a-i) The L(r)L(r) maps for different σr\sigma_{r} from 2.0 Å to 5.0 Å and at different α\alpha from 0.05 (ns)2^{2} to 0.20 (ns)2^{2}.

The MAIN module is designed primarily to extract the four parameters (Γ,Δ,VH,VL)(\Gamma,\Delta,V_H,V_L) from a given g(E)g(E) spectrum where EE is given (Extended Data Figure S1). In reality, a linear background signal may be present in the experimental g(V)g(V) spectra which are its data source. Thus, we make the augmentation before the training dataset goes into MAIN, that is:

NA(E,Γ,Δ,VH,VL,a,b)=(a+bE)×N(E,Γ,Δ,VH,VL)(11)N_A(E,\Gamma,\Delta,V_H,V_L,a,b)=(a+bE)\times N(E,\Gamma,\Delta,V_H,V_L) \tag*{(11)}

For the feature recognition, we use a Convolution Neural Network (CNN). The CNN is consistent of nine CNN blocks, where each of the CNN blocks includes two branches (Extended Data Figure S1). In the CNN block, the first branch starts with the n×1n \times1 learnable convolution layer connected with two 1×11 \times1 learnable convolution layers and the second branch is one 1×11 \times1 learnable convolution layer. In this work, the non-one kernel sizes are n=3,5,7,9,15,21,15,7,5n = 3, 5, 7, 9, 15, 21, 15, 7, 5 for the nine CNN blocks in sequence. Yet CNN alone is insufficient for accurate four-parameter extraction. Therefore, we take a self-attention branch in parallel with the CNN branch, and the two branches are then concatenated. Ultimately, four FC heads output the four parameters: one each for Γ\Gamma, Δ\Delta, VHV_H and VLV_L separately. The number of model parameters in either the COGNATE or the MAIN is less than one million.

Synthetic training-dataset generation of VMAP

We generate 3.5 million N(E)N(E) spectra using Eqn. (7), with energy in the range [−100-100 meV, 100 meV], Δ\Delta in the range [0 meV, 80 meV], Γ\Gamma in the range [0 meV, 10 meV], VHV_H and VLV_L in the range [0 m/s, 5000 m/s]. The normal state tight-binding parameters are interpolated from Ref. 36 and presented in the Supplementary Table 1. To get a valid N(E)N(E) curve, we use a dense kk-mesh of size 13501×1350113501 \times13501 and mask the effective kk-points in the energy range [−200-200 meV, 200 meV]. The same synthetic dataset is used to train both the COGNATE and the MAIN, where 3.15 million (90%) synthetic N(E)N(E) curves were used for training and 0.35 million (10%) synthetic N(E)N(E) curves were used for validation. We generate the synthetic dataset on 6×NVIDIA L40S GPUs, which takes approximately four days. Here, we also generate independent synthetic datasets for Bi2Sr2CaCu2O8\mathrm{Bi}_{2}\mathrm{Sr}_{2}\mathrm{Ca}\mathrm{Cu}_{2}\mathrm{O}_{8} samples with different hole doping levels.

Synthetic dataset augmentation for training VMAP

To enhance the capability of the MAIN for interpreting experimental data which contains random noise, we implemented random noise augmentation on the synthetic dataset used by the MAIN for training. The noise fnoise(E)f_{\mathrm{noise}}(E) is generated from a normal distribution generator (with mean value 0 and variance 1) to add random numbers at each energy point. The noise-augmented spectrum is TA(E)=N(E)+χfnoise(E)×N(E)‾T_A(E) = N(E) + \chi f_{\mathrm{noise}}(E) \times\overline{N(E)}, where N(E)N(E) is the synthetic spectrum with its mean value N(E)‾\overline{N(E)}, and χ\chi is a scalar factor that represents the noise level. At each training epoch, an independent noise realization is used for each synthetic spectrum; the same spectrum experiences different noise augmentations across epochs. We limit ln⁡(χ)\ln(\chi) in a linear range from ln⁡(0.005)\ln(0.005) to ln⁡(0.15)\ln(0.15) with χ\chi varying from 0.5% to 15% but centered at ~2.7%. Extended Data Figure 2 shows the example synthetic spectra without noise (Extended Data Figure 2) and with noise from 1% to 4% (Extended Data Figure 2-Extended Data Figure 2). With noise augmentation applied, the MAIN converges slower than trained on the clean synthetic dataset but achieves a better performance when evaluating the physical parameters of the experimental spectra. Noise augmentation is implemented only for the MAIN in this work.

Examples of data augmentation and the training of VMAP

Extended Data Figure 2. Examples of data augmentation and the training of VMAP. (a) The example of clean spectrum without noise augmentation. (b-d) The examples of spectra with 1% (b), 2% (c), and 4% (d) noise augmentation. (e) The examples of validation loss and accuracy during training COGNATE. (f) The examples of validation loss and accuracy during training MAIN.

VMAP Training

We use the Gaussian Error Linear Unit (GELU)38 as the activation function and AdamW39 optimizer throughout the VMAP. We also incorporate the position encoding37 (PE) for the COGNATE and the self-attention branch of the MAIN. For the COGNATE, we use the mean squared error (MSE) loss function LMSEL_{MSE} and take the coefficient of determination R2R^2 to evaluate the accuracy (Extended Data Figure 2).

LMSE=1N∑i=1N(yi−fi)2(12)L_{MSE} = \frac{1}{N}\sum_{i=1}^{N}(y_i-f_i)^2 \tag*{(12)}

where yiy_i is the target spectrum and fif_i is the corresponding reconstructed curve at energy ii. The COGNATE was trained for 100 epochs. For the MAIN, we use the mean absolute error (MAE) loss function LMAEL_{MAE}:

LMAE=1N∑i=1N∣pi−li∣(13)L_{MAE} = \frac{1}{N}\sum_{i=1}^{N} \lvert p_i-l_i\rvert \tag*{(13)}

where pip_i, (pi=Δ,Γ,VH,VLp_i = \Delta, \Gamma, V_H, V_L), and lil_i are the predicted parameters and the corresponding true values, respectively. The accuracy of each parameter is defined as:

Accuracy(pi)=1−∑s∣pi,s−li,s∣∑s∣li,s∣(14)Accuracy(p_i) = 1 - \frac{\sum_s \lvert p_{i,s}-l_{i,s}\rvert}{\sum_s \lvert l_{i,s}\rvert} \tag*{(14)}

where the index ss runs over all samples in the validation dataset. The COGNATE converges quicker than the MAIN since the COGNATE aims to reconstruct the spectrum from the given physical parameters and was trained on the clean dataset. In practice, we train the COGNATE first, where the COGNATE shows excellent performance (Extended Data Figure 2). In addition, we introduce a Physics-Informed Neural Network (PINN) loss LPINNL_{PINN} computed through the COGNATE to accelerate the convergence of the MAIN. The LPINNL_{PINN} is the same as LMSEL_{MSE} as in the COGNATE. The MAIN loss function is then the combined LMSEL_{MSE} and LPINNL_{PINN}:

LMAIN=(1−w)LMAE+wLPINN(15)L_{MAIN} = (1-w)L_{MAE} + wL_{PINN} \tag*{(15)}

where ww is a relative ratio and we take w=0.1w = 0.1 when training the MAIN in this work. The MAIN is trained for 300 epochs, with the PINN loss active during epochs 20 to 150. When the PINN loss is active, the COGNATE only performs forward computation with frozen weights and bias. We do not use Eqn. (7) to compute the PINN loss directly, as the numerical N(E)N(E) calculations are computationally extensive. The validation loss and accuracy of the physical parameters of MAIN are shown in Extended Data Figure 2.

VMAP Parameter Determination

The experimental data g(E)g(E) first is processed by MAIN, which outputs the four parameters (Γ,Δ,VH,VL)(\Gamma, \Delta, V_H, V_L), but this is not the final output. Instead, we use the (Γ,Δ,VH,VL)(\Gamma, \Delta, V_H, V_L) as input of the COGNATE, which generates the reconstructed spectrum NCOG(E)N_{\mathrm{COG}}(E). By fitting the ratio g(E)NCOG(E)\frac{g(E)}{N_{\mathrm{COG}}(E)} to a linear function a+bEa+bE, where g(E)g(E) is the target experimental spectrum, aa and bb can be obtained. We then use the MSE loss in the refinement procedure:

LFit=1N∑i=1N(yi(a+bEi)−fi)2(16)L_{\mathrm{Fit}}=\frac{1}{N}\sum_{i=1}^{N}\left(y_i(a+bE_i)-f_i\right)^2 \tag*{(16)}

In this scenario, the four physical parameters (Γ,Δ,VH,VL)(\Gamma,\Delta,V_H,V_L) receive gradients through the frozen COGNATE and are iteratively updated via the AdamW optimizer to minimize LFitL_{\mathrm{Fit}}. In this work, we perform 2000 optimization steps for each of the experimental spectrum and track the best parameters with the lowest LFitL_{\mathrm{Fit}}. The schematic of the refinement process is shown in Extended Data Figure 3 and 3b.

Schematic workflow of VMAP and evaluation on refinement

Extended Data Figure 3. The schematic workflow of VMAP and evaluation on Refinement

(a) The schematic of VMAP refinement. The MAIN module derives the parameters from the experimental g(r,E)g(r, E) spectra. Then those parameters, as the initial step, are input to the COGNATE model, where the background signal a+bEa + bE is solved via linear fitting. The parameters are updated through the backpropagation mechanism. Each experimental spectrum is evaluated over 2000 refinement steps. (b) The schematic diagram of VMAP workflow. (c) A typical example of the spatial distribution of the coefficient of determination

The final six parameters are used to recompute the spectra based on Eqn. (7), and the fitting quality is defined by the coefficient of validity R2R^2:

R2(r)=1−∑i=1m[g(r,Ei)−N(r,Ei)]2∑i=1m[g(r,Ei)−g‾(r)]2(17)R^2(r)=1-\frac{\sum_{i=1}^{m}[g(r,E_i)-N(r,E_i)]^2}{\sum_{i=1}^{m}[g(r,E_i)-\overline{g}(r)]^2} \tag*{(17)}

in which g(r,Ei)g(r,E_i) is the measured differential conductance spectrum and N(r,Ei)N(r,E_i) is the numerical recomputed spectrum from Eqn. (7) based on the given six parameters, g‾(r)\overline{g}(r) is the corresponding mean value of g(r,Ei)g(r,E_i). Extended Data Figure 3-Extended Data Figure 3 shows an example of the R2R^2 before and after the refinement procedure, from which the mean value of R2R^2 increased from 0.87 to 0.96.

V0MAP

In this work, the VMAP ANN has a lower limit on detectable supercurrent speeds. To quantify these limits statistically, we fix the VH=VL=0V_H=V_L=0 in a different ANN named V0MAP. V0MAP has identical ANN architecture as VMAP (Extended Data Figure S1), but with the training supercurrents fixed at zero (VH=VL=0V_H=V_L=0). In terms of the Hamiltonian Eqn. (4) which generates the 3.5 million training spectra for V0MAP, the term EG≡ℏk⋅V=0E_G\equiv\hbar\mathbf{k}\cdot\mathbf{V}=0. In this case, the V0MAP ANN searches for best fits using only Δ(r)\Delta(r), Γ(r)\Gamma(r) and the background $a(r)+ b(r)Eb(r)E as degrees of freedom. The R2R^2 comparison obtained from VMAP and V0MAP is presented in Extended Data Figure 4.

Image not extracted

Figure 4. a. To identify both the supercurrent speeds and trajectories we masked VMAP measured V(r)V(r) by the L(r)>0.5L(r)>0.5 s/m trajectory image with typical LV(r)LV(r) results shown here. The L(r)>0.5L(r)>0.5 s/m cut-off is strict and current can flow.

b. A typical tri-junction showing the VMAP quantified supercurrent speeds crossing the dashed line on approaching/departing all three channels. Two of these values add to the third as expected for Kirchhoff’s law.

c-g Typical examples of the key fact that this Kirchhoff’s law criterion is satisfied at all well-defined tri-junctions throughout all samples studied.

i-k The S(r)S(r) images obtained from V(r)V(r) in Fig. 2. A minimal boxcar averaging was applied to V(r)V(r) before calculating the LV(r)LV(r) and S(r)S(r).

Indistinguishability from zero of small supercurrent velocities

We compute the R2R^2 without supercurrent velocity and label it as R2(V=0)R^2(V = 0), then the quantity δR2\delta R^2 is defined as δR2=R2(V)−R2(V=0)\delta R^2 = R^2(V) - R^2(V = 0), where R2(V)R^2(V) is computed with finite supercurrent velocity. Extended Data fig:4a-c show the supercurrent velocity-dependent δR2\delta R^2, which is averaged at each velocity, indicating that the Galilean-boosted N(E)N(E) has a better description on the g(E)g(E) in general. Combined with the supercurrent velocity-dependent δR2\delta R^2 at different doping levels, the conservative indistinguishability of small supercurrent speed is ∼1250\sim1250 (±250\pm250) m/s in this work. It should be noted that the possible finer supercurrent structures for V<1250 m/s will be the focus of future research.

On the other hand, for clarity, we apply a cutoff L(r)>1L(r) > 1 mask onto the Δ(r)\Delta(r) (Extended Data fig:5a-c), where the domains of low supercurrent are shown clearly. In Extended Data fig:5d-f, we plot the histograms of the masked superconductive gap Δ\Delta within those supercurrent-velocity indistinguishable regions, where the results are shown in Extended Data fig:5a-c, the average superconductive gap from those domains ⟨Δ⟩=44(±7)\langle\Delta\rangle= 44(\pm7) meV for hole doping p=0.17p=0.17, and it decreases as the hole density increases, consistent with previous reportsand expectations.

Visualizing Supercurrent Loop Network

As an ancilla to VMAP for obtaining clearer images of the supercurrent loop network in rr-space, we adopt the two-dimensional curvature method on the supercurrent velocity maps V(r)V(r)40. For the square grid, where the pixel distances are equal along the xx and yy axes with the same units, the two-dimensional curvature image L(x,y)L(x,y) of the target image F=f(x,y)F = f(x,y) can be defined as (also see in Eqn. (8)):

L(x,y)=−[1+α(∂F∂x)2][α∂2F∂y2−2α2(∂F∂x)(∂F∂y)∂2F∂x∂y+[1+α(∂F∂y)2]α∂2F∂x2][1+α(∂F∂x)2+α(∂F∂y)2]3/2(18)L(x,y)=-\frac{\left[1+\alpha\left(\frac{\partial F}{\partial x}\right)^2\right]\left[\alpha\frac{\partial^2F}{\partial y^2}-2\alpha^2\left(\frac{\partial F}{\partial x}\right)\left(\frac{\partial F}{\partial y}\right)\frac{\partial^2F}{\partial x\partial y}+\left[1+\alpha\left(\frac{\partial F}{\partial y}\right)^2\right]\alpha\frac{\partial^2F}{\partial x^2}\right]}{\left[1+\alpha\left(\frac{\partial F}{\partial x}\right)^2+\alpha\left(\frac{\partial F}{\partial y}\right)^2\right]^{3/2}} \tag*{(18)}

where α\alpha is a positive free parameter. For simplicity, we take α\alpha in units of (ns)2^2, FF in units of m/s, and xx and yy in units of nm; then the curvature image L(x,y)L(x,y) when calculated on the function F(r)=V(r)F(r)=V(r) has units of s/m in this work. In the limit of α→0\alpha\to0, L(x,y)L(x,y) approaches the Laplacian with a negative sign, and with α→∞\alpha\to\infty, L(x,y)L(x,y) will diverge at the extrema of FF, giving L(x,y)L(x,y) a good choice to trace the extrema position. When tracing the supercurrent network, we mainly focus on the maximal velocity rather than the minimal velocity; therefore, we apply a cutoff L(x,y)≥0L(x,y) \ge0 when applying this negative second derivative method to the supercurrent velocity image V(x,y)V(x,y). We also employ a two-dimension Gaussian kernel with standard deviation σr\sigma_r to smooth the V(x,y)V(x,y) before computing the curvature map L(x,y)L(x,y). Supplementary Figure 1 shows a series of L(x,y)L(x,y) maps with different σr\sigma_r and α\alpha.

Topographies of the Bi2Sr2CaCu2O8\mathrm{Bi}_2\mathrm{Sr}_2\mathrm{CaCu}_2\mathrm{O}_8 samples for different hole doping

Extended Data Figure 6 shows the topography images of the Bi2Sr2CaCu2O8\mathrm{Bi}_2\mathrm{Sr}_2\mathrm{CaCu}_2\mathrm{O}_8 in the same field of view as in Figure 2, with hole doping from p=0.17p=0.17 to 0.22, from which the one-dimensional super-modulation gradually disappears with increasing the doping level. At the hole doping p=0.22p=0.22, the unidirectional super-modulation is almost invisible.

Topography images of the Bi$_2$Sr$_2$CaCu$_2$O$_8$ with different hole doping

Extended Data Figure 6. Topography images of the Bi2_2Sr2_2CaCu2_2O8_8 with different hole doping

(a) The topography image of optimal Bi2_2Sr2_2CaCu2_2O8_8 at doping p=0.17p=0.17 in the same FOV as Figs. 2a, d, g (VBias=−90V_{\mathrm{Bias}}=-90 mV, It=120I_{\mathrm{t}}=120 pA). (b) The topography image of over doped Bi2_2Sr2_2CaCu2_2O8_8 at doping p=0.20p=0.20 in the same FOV as Figs. 2b, e, h (VBias=60V_{\mathrm{Bias}}=60 mV, It=60I_{\mathrm{t}}=60 pA). (c) The topography image of over doped Bi2_2Sr2_2CaCu2_2O8_8 at doping p=0.22p=0.22 in the same FOV as Figs. 2c, f, i (VBias=40V_{\mathrm{Bias}}=40 mV, It=60I_{\mathrm{t}}=60 pA).

Separate imaging of supercurrent speeds for filled/empty states

In theory for the superconducting density of states, the Galilean-boosted N(E)N(E) is symmetric for the positive and negative energies when linear background augmentation is unimplemented, as it was constrained by the Bogoliubov-DeGennes Hamiltonian. To have a deeper inspection of the supercurrent speed identified in the positive and negative energy regimes, we separately symmetrize g(E)g(E) from the positive and negative energy regimes and perform the VMAP analysis independently. Here it is important to note that the random variations in g(r,E>0)g(r, E > 0) and g(r,E<0)g(r, E < 0) which are generated by finite SNR are completely uncorrelated. Effectively, these are two independent experiments.

The resulting supercurrent velocity V+V^{+} and V−V^{-} are responses to the positive energy g(E)>0g(E) > 0 and negative energy g(E)<0g(E) < 0, respectively. In this work, the experimental spectra are always collected in an energy-symmetric window [−E,+E][-E, +E], thus we can directly evaluate V±V^{\pm} from the same procedure of the MAIN, and refinement on the COGNATE with 2000 steps on each of the spectra.

Extended Data Fig. 7 shows the V+(r)V^{+}(r) and V−(r)V^{-}(r) in the same FOV as in Fig. 2 with the hole doping p=0.17p = 0.17, 0.200.20, and 0.220.22, from which the V+(r)V^{+}(r) and V−(r)V^{-}(r) are virtually identical. Although these are effectively two separate independent experiments, the V−(r)V^{-}(r) images derived by VMAP from g(r,E<0)g(r, E < 0) are virtually identical to the V+(r)V^{+}(r) images derived from g(r,E>0)g(r, E > 0) in all studies, providing additional strong evidence of the validity of VMAP.

Estimation of the effective magnetic field and comparison to reports

In the simplest scenario, when the electron pairs move to form a closed circle loop, the current II is:

I=nqV2πr(19)I = \frac{nqV}{2\pi r} \tag*{(19)}

where q=2eq = 2e is the effective charge in the supercurrent case and ee is the electron charge, nn is the number of flowing pairs, rr is the radius of the circle, and VV is the current velocity. Because the supercurrent circuits are only approximately 0.1 nm wide, no Meissner effect exists. Thus, the loop-current induced magnetic field from the Biot-Savart law is:

B=μ0I2r(20)B = \frac{\mu_0 I}{2r} \tag*{(20)}

Where μ0\mu_0 is the vacuum magnetic permeability. In Extended Data Figure 5-Extended Data Figure 5, we show the histograms of the supercurrent velocity of V(V>0)V(V > 0), where the averaged supercurrent velocity is ⟨V⟩∼2100\langle V\rangle\sim2100 m/s throughout the hole doping p=0.17p = 0.17 to p=0.22p = 0.22. To estimate the radius rr of the supercurrent loop, we compute the autocorrelation of L(r)L(r) and the results are shown in Extended Data Fig. 8, from which the sharp peaks that close to the image center indicate the value of the diameter (2r2r) of the loops. The corresponding average radius of the supercurrent loop is ⟨r⟩=1.35(±0.43)\langle r\rangle= 1.35(\pm0.43) nm. By fitting the line cuts in the autocorrelation of L(r)L(r) with an exponential function

Superconducting gap patterns and supercurrent velocities

Extended Data Figure 5. Superconducting gap patterns and supercurrent velocities. (a-c) The spatial distribution of the masked superconductive gaps Δ(r)\Delta(r) with hole doping p=0.17p=0.17 (a), p=0.20p=0.20 (b), and p=0.22p=0.22 (c), showing the patterns of the superconductive gap domains. (d-e) The histogram of superconductive gap from (a-c) with hole doping p=0.17p=0.17 (d), p=0.20p=0.20 (e) and p=0.22p=0.22 (f), showing the averaged ⟨Δ⟩=44(±7)\langle\Delta\rangle= 44(\pm7) meV (d), ⟨Δ⟩=32(±5)\langle\Delta\rangle= 32(\pm5) meV (e), and ⟨Δ⟩=27(±6)\langle\Delta\rangle= 27(\pm6) meV (f), respectively. (g-i) The histogram of supercurrent velocity with hole doping p=0.17p=0.17 (g), p=0.20p=0.20 (h) and p=0.22p=0.22 (i), showing the averaged ⟨V⟩=2171(±885)\langle V\rangle= 2171(\pm885) m/s (g) ⟨V⟩=2023(±768)\langle V\rangle= 2023(\pm768) m/s (h) and ⟨V⟩=2174(±919)\langle V\rangle= 2174(\pm919) m/s (i), respectively.

y=Ae−ddL+c(21)y = A e^{-\frac{d}{d_L}} + c \tag*{(21)}

where AA and cc are the free parameters and dLd_L is the exponential decay length. The decay length dLd_L from the center of autocorrelation of L(r)L(r) is computed, from which the averaged dL=0.14d_L=0.14 nm (Extended Data Fig. 8), giving the width of the supercurrent loops. Meanwhile, from the normal-state band structure Eqn. (2), we estimate the Fermi wavevector kF∼4k_F \sim4 nm−1^{-1}. The electron density in an uncorrelated two-dimension electron gas is:

ne=kF22π(22)n_e = \frac{k_F^2}{2\pi} \tag*{(22)}

On the loop with radius rr and loop width dLd_L, the corresponding electron population is 2πrdLne2\pi r d_L n_e, resulting in the Cooper pair population within the loop width

nL=kF2rdL2(23)n_L = \frac{k_F^2 r d_L}{2} \tag*{(23)}

Taking the values of kFk_F, rr, and dLd_L into Eqn. (23), the estimated number of flowing pairs on a supercurrent loop is nL∼1.5n_L \sim1.5, the corresponding supercurrent intensity I=1×10−7I = 1 \times10^{-7} A, and the corresponding effective magnetic field ∣B∣≈0.1|B| \approx0.1 mT. This estimated effective magnetic field is expected and consistent with the μ\muSR-reported value[2, 3].

On the other hand, magnetic dipoles beneath the surface of the sample might lead to current loops due to Meissner screening. However, a static moment mm at depth dd below the CuO2_2 plane produces an in-plane vector potential whose maximum is Amax⁡≈0.38μ0m/(4πd2)A_{\max} \approx0.38 \mu_0 m/(4\pi d^2)41. The London response of the condensate to it would be a supercurrent of speed V≈(e/me)AV \approx(e/m_e)A42. For m=5μBm = 5\mu_B at d=0.5d = 0.5 nm this gives V≈1V \approx1 m/s, and even for a cluster with m=100μBm = 100\mu_B only V≈25V \approx25 m/s. Hence, this magnetic impurity phenomenon could only produce effects that are two to three orders of magnitude weaker than what we report in this work.

“Rosetta Stone” of dd-wave spectra in presence of a supercurrent

For a uniform superfluid velocity VV the condensate acquires a center-of-mass momentum ps=2mVp_s = 2mV, which Galilean boosts the quasiparticle spectrum Ek=±ξk2+Δk2E_k = \pm\sqrt{\xi_k^2 + \Delta_k^2} (ξk\xi_k is the normal-state band dispersion) by

EG(k)=VF(k)⋅ps(24)E_G(\mathbf{k}) = \mathbf{V}_F(\mathbf{k}) \cdot\mathbf{p}_s \tag*{(24)}

The local density of states is then obtained from the imaginary part of the current-carrying Green’s function,

N(ω)=−(1π)∑kIm⁡[ω+EG(k)+ξk(ω+ϵG(k)+i0+)2−ξk2−Δk2](25)N(\omega) = -\left(\frac{1}{\pi}\right) \sum_k \operatorname{Im}\left[\frac{\omega+ E_G(k) + \xi_k}{(\omega+ \epsilon_G(k) + i0^+)^2 - \xi_k^2 - \Delta_k^2}\right] \tag*{(25)}

with Δk=Δ0cos⁡(2φ)\Delta_k = \Delta_0\cos(2\varphi) for a dx2−y2d_{x^2-y^2} superconductor over a circular Fermi surface. If the superfluid velocity makes an angle θ\theta with the crystal xx-axis, Fermi surface ϵG(φ)=VFpscos⁡(φ−θ)\epsilon_G(\varphi) = V_Fp_s\cos(\varphi- \theta), so the four nodal quasiparticles are shifted unequally according to their projection along the superfluid velocity. Consequently, the zero-bias density of states, which involves an integral over all kF(φ)k_F(\varphi) on the Fermi surface, acquires a finite contribution proportional to the Galilean boost at the nodes.

On a nearly cylindrical Fermi surface, the gap takes the form Δk=Δ0cos⁡(2φ)\Delta_k = \Delta_0\cos(2\varphi), so the excitation energy along the Fermi surface is E(φ)=∣Δ0cos⁡(2φ)∣E(\varphi) = \lvert\Delta_0\cos(2\varphi)\rvert. The Bogoliubov dispersion then possesses four symmetry-related saddle points at the antinodes, kAN=(±kF,0)k_{AN} = (\pm k_F, 0) and (0,±kF)(0, \pm k_F), where ξk=0\xi_k = 0 and ∣Δk∣=Δ0\lvert\Delta_k\rvert= \Delta_0. These saddle points give rise to van Hove singularities in the quasiparticle density of states at ω=±Δ0\omega= \pm\Delta_0, producing the familiar BCS coherence peaks. A finite superfluid velocity lifts the degeneracy of the four antinodal saddle points through the Galilean boost EG(k)=VFpsE_G(k) = V_Fp_s.

For superfluid velocity directed along an antinodal axis (θ=0\theta= 0), the Fermi velocities at the antinodes (0,±kF)(0, \pm k_F) are perpendicular to the flow and therefore experience no Galilean boost, while those at (±kF,0)(\pm k_F, 0) are shifted by ±VFps\pm V_F p_s. Consequently, each coherence peak evolves into a characteristic three-peak structure consisting of an unsplit central peak at ω=±Δ0\omega= \pm\Delta_0, arising from the two unaffected antinodes, together with two side peaks shifted upward and downward by the Galilean boost energy (Extended Data Figure 9).

Dependence of local spectra on superfluid velocity

Extended Data Figure 9. Dependence of local spectra on superfluid velocity. (a-b) Velocity along antinode (a) and corresponding spectrum (b). (c-d) Velocity along node (c) and corresponding spectrum (d).

In contrast, when the superfluid velocity is directed along a nodal axis (θ=π/4)(\theta= \pi/4), all four antinodes acquire Galilean boost of equal magnitude VFps/2V_F p_s/\sqrt{2}, with two shifted upward and two downward, so the central peak disappears and only a pair of split coherence peaks remains (Extended Data Figure 9). Thus, the evolution from a three-peak spectrum for antinodal flow to a two-peak spectrum for nodal flow provides a direct spectroscopic signature of the orientation of the superfluid velocity relative to the dx2−y2d_{x^2-y^2} order parameter.

Supercurrent loop trajectories for different hole densities

Extended Data Fig. 10 shows the supercurrent loop trajectories, L(r)L(r) across the whole FOV, with hole dopings from p=0.17p=0.17 to 0.220.22. For each FOV, we identify examples of individual supercurrent loops identified by the VMAP ANN. Colored dots indicate different directions relative to the crystal axes using the same concept as Figs. 1f-i and are vertically shifted to add clarity. The correspondence between Extended Data Fig. 10, Figure 3 and Figs. 1f-i provides additional confirmation of supercurrent flow in a dd-wave superconductor across different doping and FOVs.

Evaluation of the superfluid current conservation

The supercurrent loop trajectories sometimes form tri-junctions throughout all FOV’s studied in this paper. At each of these junctions, Kirchhoff’s law can be evaluated to estimate if the superfluid current is conserved. The function LV(r)LV(r) is calculated by masking the V(r)V(r) with the L(r)>0.5L(r) > 0.5 s/m map to get a speed-resolved network image. This is shown in Fig. 4a for p=0.17p=0.17.

Due to the network pattern of LV(r)LV(r), a tri-junction can be distorted by nearby junctions, so a well-defined junction is selected with each branch being un-interfered with by other crossing branches, as shown in Figs. 4b-h. A linecut is then taken across each of the 3 branches, a little distance away from the junction and the average speed along the lines are extracted. In all cases where there are well-defined junctions, the speed of 1 branch is equal to the sum of the other two branches satisfying Kirchhoff’s Law.

Separately, to evaluate the superfluid current conservation globally, we assume that the superfluid density does not vary strongly locally and can be taken as a constant value in the minimal patch unit. Then, Kirchhoff’s law in the superfluid current velocity version can be described as evaluating the value ∑jV(j)\sum_j V(j) for each site ii, where the sum of observed supercurrent speeds through a site ii is derived from its four neighbor sites jj. VMAP determines the larger and smaller of the two orthogonal velocity components, VH(j)V_H(j) and VL(j)V_L(j), at each pixel jj. Therefore, the supercurrent speed along the horizontal axis at pixels j=1,3j=1,3 is one of four values ±VH(j)\pm V_H(j); ±VL(j)\pm V_L(j), while the supercurrent speed along the vertical axis at pixels j=2,4j=2,4 is also one of four values ±VH(j)\pm V_H(j); ±VL(j)\pm V_L(j). This means that for each of the 4 sites jj surrounding pixel ii there are 4 possible supercurrents. Thus there are 44=2564^4=256 possible combinations of ±VH(j)\pm V_H(j), ±VL(j)\pm V_L(j) for all four sites jj surrounding site ii. For each pixel ii we then evaluate all possible values of:

S(i)=VRx−VLx+VDy−VUy(26)S(i)=V_R^x-V_L^x+V_D^y-V_U^y \tag*{(26)}

where RR, LL, DD and UU denote the right, left, lower and upper neighbors spanning jj. At site ii, the relevant components (VRxV_R^x, VLxV_L^x for RR and LL; VDyV_D^y, VUyV_U^y for DD and $U) at each neighbor jj then take any of the four measured values ±VH(j)\pm V_H(j), ±VL(j)\pm V_L(j). S(ri)S(r_i) is then defined as the minimum value of ∣S(i)∣\lvert S(i)\rvert over all 256 possible combinations. In practice, we implement a minimal boxcar averaging on the superfluid current map V(r)V(r) to enhance the signal-to-noise ratio before computing S(r)S(r). The results of S(r)S(r) are shown in the Fig. 4 and Supplementary Extended Data Figure 2, where the mean value of S(r)S(r) is between 15 m/s and 20 m/s throughout all pp, the corresponding ratio Sˉ/Vˉ\bar{S}/\bar{V} is about 1% with and 2% in all dopings studied in this work, providing evidence for universal supercurrent conservation throughout.

Cross-correlation among different parameters from VMAP

In Supplementary Extended Data Figure 2, we show all parameters in rr-space obtained from the VMAP associated with the Δ(r)\Delta(r) and V(r)V(r) in a typical FOV with hole doping p=0.20p = 0.20, and the corresponding normalized cross-correlation coefficients XX among different parameters are shown in Supplementary Table 2. The normalized cross-correlation coefficient XX is:

X(M,N)=∫M(r)N(r) dr∫∣M(r)∣2 dr∫∣N(r)∣2 dr(27)X(M,N) = \frac{\int M(r)N(r)\,\mathrm{d}r}{\sqrt{\int|M(r)|^2\,\mathrm{d}r\int|N(r)|^2\,\mathrm{d}r}} \tag*{(27)}

where M(r)M(r) and N(r)N(r) are the target images after the mean-value subtracted. When we measure the normalized cross correlation X(M,N)X(M,N) between M(r)M(r) and N(r)N(r) (where M(r)M(r) and N(r)N(r) are the output from VMAP), it is typically X<0.5X < 0.5 for all samples studied (Supplementary Table 2).

Methods-only References

VMAP architecture

Figure Extended Data Fig. 1. Extended Data Fig. 1. VMAP architecture. (a) The architecture of the COGNATE, which takes the energy variable EE, Δ\Delta, Γ\Gamma, VHV_{\mathrm{H}}, and VLV_{\mathrm{L}} as input and outputs N(E)N(E). (b) The architecture of the MAIN, which takes the energy and N(E)N(E) as input, and outputs the four parameters: Δ\Delta, Γ\Gamma, VHV_{\mathrm{H}}, and VLV_{\mathrm{L}}. (c) The architecture of the Self-Attention block as in (a) and (b). (d) The architecture of Convolution-Neural-Network (CNN) block as in (b).

R2(r)R^2(r) without the COGNATE refinement. (d) The histogram of R2(r)R^2(r) as in (c), showing the mean value ⟨R2⟩=0.87\langle R^2\rangle=0.87. (e) The spatial distribution of the coefficient of determination R2(r)R^2(r) after the COGNATE refinement in the same FOV as (c). (f) The histogram of R2(r)R^2(r) as in (e), showing the mean value ⟨R2⟩=0.96\langle R^2\rangle=0.96.

Indistinguishable small supercurrents for three hole-doped samples

Extended Data Figure 4. Indistinguishable small supercurrents V<1250V<1250 m/s from V=0V=0. (a) The supercurrent velocity-dependent δR2=R2(V)−R2(V=0)\delta R^2 = R^2(V) - R^2(V=0) in the sample with hole doping p=0.17p=0.17, where the positive correlation indicates that the Galilean-boosted N(E)N(E) provides a better description than N(E)N(E) without supercurrent velocity, especially in the high supercurrent regime. (b-c) The same as in (a) but in the sample with hole doping p=0.20p=0.20 (b) and p=0.22p=0.22 (c). A boxcar average, with supercurrent velocity of 250 m/s, was applied to (a-c).

Extended Data Fig. 7: $V^{+}(r)$ and $V^{-}(r)$ identified by the VMAP

Figure 7. V+(r)V^{+}(r) and V−(r)V^{-}(r) identified by the VMAP

(a–c) Example spectra at p=0.17p=0.17 (a), p=0.20p=0.20 (b) and p=0.22p=0.22 (c), fitted separately for E<0E < 0 (giving V−V^{-}) and for E>0E > 0 (giving V+V^{+}). (d–f) The V+(r)V^{+}(r) identified by the VMAP with different hole doping p=0.17p=0.17 (d), p=0.20p=0.20 (e), and p=0.22p=0.22 (f). (g–i) The V−(r)V^{-}(r) identified by the VMAP with different hole doping p=0.17p=0.17 (g), p=0.20p=0.20 (h), and p=0.22p=0.22 (i) in the same FOV as in (d–f).

Figure Extended Data Fig. 8. Estimation of the supercurrent loop radius. (a) The auto-correlation map of L(r)L(r) with hole doping p=0.17p=0.17. (b) The auto-correlation map of L(r)L(r) with hole doping p=0.20p=0.20. (c) The auto-correlation map of L(r)L(r) with hole doping p=0.22p=0.22. (d-i) The line cuts of L(r)L(r) marked in (a-c), indicating the averaged supercurrent loop

radius r=1.4r = 1.4 nm (half of the dimeter), and the exponential function fitting (the solid orange curves) give the averaged loop with dL=0.14d_L = 0.14 nm.

(a) The L(r)L(r) as in the Supplementary Figure S1-Extended Data Figure 2, where the hole doping p=0.20p=0.20. (b-d) The examples of individual supercurrent loops identified by the VMAP, the loops 1#-3# are marked in (a) (VBias=60V_{\mathrm{Bias}}=60 mV, It=60I_t=60 pA, Vmod=3.0V_{\mathrm{mod}}=3.0 mV). The spectra in (b-d) are vertically shifted for clarity. (e) The L(r)L(r) as in the Fig. 2, where the hole doping p=0.20p=0.20. (f-h) The examples of individual supercurrent loops identified by the VMAP, the loops 1#-3# are marked in (e) (VBias=60V_{\mathrm{Bias}}=60 mV, It=60I_t=60 pA, Vmod=2.5V_{\mathrm{mod}}=2.5 mV). The spectra in (f-h) are vertically shifted for clarity. (i) The L(r)L(r) as in the Fig. 2, where the hole doping p=0.22p=0.22. (j-l) The examples of individual supercurrent loops identified by the VMAP, the loops 1#-3# are marked in (i) (VBias=40V_{\mathrm{Bias}}=40 mV, It=60I_t=60 pA, Vmod=2.5V_{\mathrm{mod}}=2.5 mV). The spectra in (j-l) are vertically shifted for clarity.

Supplementary Information

ppμ\mu (meV)t1t_{1} (meV)t2t_{2} (meV)t3t_{3} (meV)
0.1740736010836
0.2041436010836
0.2242736010836

Table 1. The tight binding parameters for Bi2Sr2CaCu2O8\mathrm{Bi}_{2}\mathrm{Sr}_{2}\mathrm{CaCu}_{2}\mathrm{O}_{8} with hole doping p=0.17p=0.17, 0.20, and 0.22 studied in this work.

XXp=0.17p = 0.17p=0.20p = 0.20p=0.22p = 0.22p=0.20p = 0.20 (Small FOV)
Δ−V\Delta- V0.490.440.560.40
Δ−Γ\Delta- \Gamma0.330.250.280.39
Δ−a\Delta- a0.550.04-0.17-0.06
Δ−b\Delta- b0.550.520.350.50
V−ΓV - \Gamma0.280.090.190.13
V−aV - a0.20-0.14-0.40-0.26
V−bV - b0.330.350.310.33
Γ−a\Gamma- a0.24-0.01-0.220.02
Γ−b\Gamma- b0.210.160.100.24
a−ba - b0.77-0.52-0.54-0.52

Table 2. Supplementary Table 2. The cross-coefficient among different parameters obtained from VMAP for different hole doping pp.

Simultaneous determination of different parameters in a typical FOV

Figure 2 (PDF p. 54). Simultaneous determination of different parameters in a typical FOV

The VMAP analysis with hole doping p=0.20p=0.20 in the same FOV of Supplementary Figure S1. (a) The topography image. (b-h) The parameters obtained from VMAP in the same field FOV as in (a). (i-p) A series of example spectra obtained from VMAP, where the positions are marked in (b-h) (VBias=60V_{\mathit{Bias}}=60 mV, It=60I_t=60 pA, Vmod=3V_{\mathit{mod}}=3 mV).

Supplementary Videos

To demonstrate validity of determination of Δ(r)\Delta(r) and V(r)V(r) by the VMAP ANN we have prepared three-panel Video of simultaneous g(r,E):N(r,E):∣g(r,E)−N(r,E)∣g(r, E):N(r, E):\lvert g(r, E) - N(r, E)\rvert: Supplementary Video 1 for p=0.17p=0.17, Supplementary Video 2 for p=0.20p=0.20, and Supplementary Video 3 for p=0.22p=0.22. Because the ∣g(r,E)−N(r,E)∣\lvert g(r, E) - N(r, E)\rvert manifestly yields values that are virtually zero for all EE, this demonstrates directly and visually the excellent quality of the fits to g(r,E)g(r, E), which are generated by VMAP assignment of N(r,E)N(r, E), throughout this work.

References

  1. [2]Sonier, J. E. et al. Anomalous Weak Magnetism in Superconducting YBa₂Cu₃O₆₊ₓ. Science 292, 1692–1695 (2001).pubmed.ncbi.nlm.nih.gov/11387468
  2. [3]Pal, A. et al. Quasistatic internal magnetic field detected in the pseudogap phase of Bi₂₊ₓSr₂₋ₓCaCu₂O₈₊δ by muon spin relaxation. Phys. Rev. B 97, 060502(R) (2018).DOI
  3. [4]Kaminski, A. et al. Spontaneous breaking of time-reversal symmetry in the pseudogap state of a high-Tc superconductor. Nature 416, 610–613 (2002).pubmed.ncbi.nlm.nih.gov/11948344
  4. [5]Fauqué, B. et al. Magnetic Order in the Pseudogap Phase of High-Tc Superconductors. Phys. Rev. Lett. 96, 197001 (2006).pubmed.ncbi.nlm.nih.gov/16803131
  5. [6]Li, Y. et al. Unusual magnetic order in the pseudogap region of the superconductor HgBa₂CuO₄₊δ. Nature 455, 372–375 (2008).DOI
  6. [7]Bourges, P., Bounoua, D. & Sidis, Y. Loop currents in quantum matter. C. R. Physique 22, 7–31 (2021).DOI
  7. [8]Xia, J. et al. Polar Kerr-Effect Measurements of the High-Temperature YBa₂Cu₃O₆₊ₓ Superconductor: Evidence for Broken Symmetry near the Pseudogap Temperature. Phys. Rev. Lett. 100, 127002 (2008).pubmed.ncbi.nlm.nih.gov/18517903
  8. [9]Zhang, J. et al. Discovery of slow magnetic fluctuations and critical slowing down in the pseudogap phase of YBa₂Cu₃Oᵧ. Sci. Adv. 4, eaao5235 (2018).DOI
  9. [10]Zhao, S. Y. F. et al. Time-reversal symmetry breaking superconductivity between twisted cuprate superconductors. Science 382, 1422–1427 (2023).DOI
  10. [11]Qi, S. et al. High-temperature field-free superconducting diode effect in high-Tc cuprates. Nat. Commun. 16, 531 (2025).DOI
  11. [12]Andersen, B. M., Kreisel, A. & Hirschfeld, P. J. Spontaneous time-reversal symmetry breaking by disorder in superconductors. Front. Phys. 12, 1353425 (2024).DOI
  12. [13]Fogelström, M. & Yip, S.-K. Time-reversal symmetry-breaking states near grain boundaries between d-wave superconductors. Phys. Rev. B 57, R14060–R14063 (1998).
  13. [14]Håkansson, M., Löfwander, T. & Fogelström, M. Spontaneously broken time-reversal symmetry in high-temperature superconductors. Nat. Phys. 11, 755–760 (2015).DOI
  14. [15]Wennerdal, N. W., Ask, A., Holmvall, P., Löfwander, T. & Fogelström, M. Breaking time-reversal and translational symmetry at edges of d-wave superconductors: microscopic theory and comparison with quasiclassical theory. Phys. Rev. Res. 2, 043198 (2020).DOI
  15. [16]Kaur, R. P. & Agterberg, D. F. Broken time reversal symmetry and superconducting states in the cuprates. Phys. Rev. B 68, 100506 (2003).DOI
  16. [17]Varma, C. M. Theory of the pseudogap state of the cuprates. Phys. Rev. B 73, 155113 (2006).DOI
  17. [18]Varma, C. M. Pseudogap in cuprates in the loop-current ordered state. J. Phys. Condens. Matter 26, 505701 (2014).DOI
  18. [19]Spivak, B., Oreto, P. & Kivelson, S. A. A Theory of quantum metal to superconductor transitions in highly conducting systems. Phys. Rev. B 77, 214523 (2008).DOI
  19. [20]Kivelson, S. A. & Spivak, B. Macroscopic character of composite high-temperature superconducting wires. Phys. Rev. B 92, 184502 (2015).DOI
  20. [21]Li, Z.-X., Kivelson, S. A. & Lee, D.-H. Superconductor-to-metal transition in overdoped cuprates. npj Quantum Mater. 6, 36 (2021).arxiv.org/abs/2010.06091
  21. [22]Breið, C. N., Hirschfeld, P. J. & Andersen, B. M. Supercurrents and spontaneous time-reversal symmetry breaking by nonmagnetic disorder in unconventional superconductors. Phys. Rev. B 105, 014504 (2022).DOI
  22. [23]Breið, C. N., Kreisel, A., Roig, M., Hirschfeld, P. J. & Andersen, B. M. Time-reversal symmetry breaking from lattice dislocations in superconductors. Phys. Rev. B 109, 014505 (2024).DOI
  23. [24]Liu, X., Chong, Y. X., Sharma, R. & Davis, J. C. S. Atomic-scale visualization of electronic fluid flow. Nat. Mater. 20, 1480–1484 (2021).DOI
  24. [25]Zhang, Y. et al. Machine learning in electronic-quantum-matter imaging experiments. Nature 570, 484–490 (2019).
  25. [26]Arovas, D. P., Berg, E., Kivelson, S. A. & Raghu, S. The Hubbard Model. Annu. Rev. Condens. Matter Phys. 13, 239–274 (2022).
  26. [27]Qin, M., Schäfer, T., Andergassen, S., Corboz, P. & Gull, E. The Hubbard Model: a computational perspective. Annu. Rev. Condens. Matter Phys. 13, 275–302 (2022).DOI
  27. [28]Davis, J. C. S. & Lee, D.-H. Concepts relating magnetic interactions, intertwined electronic orders, and strongly correlated superconductivity. Proc. Natl. Acad. Sci. 110, 17623–17630 (2013).DOI
  28. [29]Yip, S.K. & Sauls, J. A. Nonlinear Meissner effect in CuO superconductors. Phys. Rev. Lett. 69, 2264 (1992).DOI
  29. [30]Volovik, G.E. Superconductivity with lines of GAP nodes: density of states in the vortex. Sov. Phys. JETP Lett. 58, 469 (1993).
  30. [31]Wu, H. & Sauls, J. A. Majorana excitations, spin and mass currents on the surface of topological superfluid ³He-B. Phys. Rev. B 88, 184506 (2013).DOI
  31. [32]Lang, K. M. et al. Imaging the granular structure of high-Tc superconductivity in underdoped Bi₂Sr₂CaCu₂O₈+δ. Nature 415, 412–416 (2002).DOI
  32. [33]McElroy, K. et al. Atomic-Scale Sources and Mechanism of Nanoscale Electronic Disorder in Bi₂Sr₂CaCu₂O₈+δ. Science 309, 1048–1052 (2005).pubmed.ncbi.nlm.nih.gov/16099978
  33. [34]Alldredge, J. W. et al. Evolution of the electronic excitation spectrum with strongly diminishing hole density in superconducting Bi₂Sr₂CaCu₂O₈+δ. Nat. Phys. 4, 319–326 (2008).DOI
  34. [35]Kübert, C. & Hirschfeld, P. J. QUASIPARTICLES IN THE VORTEX STATE OF DIRTY d-WAVE SUPERCONDUCTORS. J. Phys. Chem. Solids 59, 1835–1837 (1998).DOI
  35. [36]Drozdov, I. K. et al. Phase diagram of Bi₂Sr₂CaCu₂O₈+δ revisited. Nat. Commun. 9, 5210 (2018).DOI

Paper details

vlnc:W_ch332zgh
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 2

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