WorksPhysics and Astronomy
Atomic-scale Imaging of Time Reversal Breaking Supercurrents in
Abstract
A persistent mystery of -based high temperature superconductivity has been the mechanism and consequences of time reversal ()-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 in , we search for the specific signatures of flowing supercurrents in a disordered -wave superconductor. Employing the strategy of artificial neural networks (ANN) for complex quantum matter visualization, we designed a dual-ANN (VMAP) to efficiently implement -wave supercurrent visualization, achieving imaging of supercurrent speeds at magnetic fields 50 T in . Correspondence between specific characteristics of around each supercurrent loop and a quantum "Rosetta Stone" which translates to , emphasizes the validity of detection and visualization. Then, for hole-densities 0.17 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 in disordered -wave superconductors are consistent with measured 1250 m/s 4000 m/s, and with reported magnetic field amplitudes 0.1 mT. Thus, detection and visualization of these supercurrent loop networks provides unique perspectives on -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 -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 -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 -symmetry breaking states[28]. Cuprate superconductivity exhibits an order-parameter , where is the macroscopic quantum phase and exhibits symmetry. In these materials, -symmetry breaking has long been reported. For example, zero-field muon spin relaxation (SR) and neutron diffraction experiments in report spontaneous static magnetic fields[2, 5] and in quasistatic magnetic fields[3], into the superconducting phase. Circularly polarized angle resolved photoemission (ARPES) studies, detect spontaneous dichroism in implying underlying -symmetry breaking[4]. Neutron scattering reveals intra-unit-cell magnetic order in and , hypothesized due to -symmetry breaking loop currents[6, 7] into the superconducting phase. Polar Kerr rotation in represents a non-reciprocity forbidden without -symmetry breaking[8]. Second harmonic generation and birefringence also indicate breaking of -symmetry[9]. Finally, the Josephson diode effect in pristine/twist junctions requires -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 -symmetry breaking. But no direct evidence of such supercurrents has ever been identified.
Theoretical hypotheses for the mechanism of cuprate -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 -symmetry. However, such crystal defects are not observed in our studies. Another possibility is spatially homogeneous -symmetry breaking due to a mixed singlet-triplet pairing state with symmetry[16]. Alternatively, there are predictions that intra-unit-cell persistent loop currents in the pseudogap phase, not only break -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 -wave superconductors predict -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 superconductivity with symmetry surrounding impurity sites, and thus supercurrents breaking -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 of Bogoliubov quasiparticle states in a superconductor is , where is the normal-state band structure referenced to the Fermi energy and is the electron-pairing order parameter. In a laboratory/crystal frame relative to which supercurrent flows at velocity , each receives a Galilean energy-boost[24, 29, 30, 31] yielding a new spectrum:
Figure 1d represents schematic consequences of Eqn. (1) for a -wave superconductor. Importantly, the specifics of each consequent spectrum depend on the vectorial direction of the supercurrent flow. Considering a symmetry superconductor (Fig. 1e) the resultant spectra for supercurrents at angles , , and relative to the order parameter and for different supercurrent speeds , are shown in Figs. 1f-i. This is a quantum “Rosetta Stone” translating theoretic spectral shapes into supercurrent speeds . Although based upon semiclassical theory, during development of our technique within that approximation[24] we demonstrated quantitatively correct supercurrent speeds at , where is the superconductor’s coherence length and is the scale associated with supercurrent variations[24]. For , direct visualization reveals a true vortex core (not a vortex halo) where the order parameter reaches zero with a coherence length , while the typical region surrounded by a supercurrent (Figs. 2, Figure 3, 4 below) has diameter so that the semiclassical approximation appears not invalid. Thus, supercurrent speeds up to 4000 m/s may, in principle, be visualized via distinct characteristics of the measured differential conductance spectrum translated to using the “Rosetta Stone” in Fig. 1. Here we exploit this new strategy to detect and visualize -symmetry breaking supercurrents in the canonical -based superconductor, .
Figure 3.
High-data-volume spectroscopic imaging scanning tunneling microscopy (SISTM) is used for visualization of the density-of-electronic-states . The samples are inserted into a cryogenic SISTM, cleaved at K, and inserted in the STM scan head (Methods A). The differential tunneling conductance spectrum allows atomic-scale imaging of , 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 using a version of Eqn. (1) and for all possible , 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 -wave superconductor in Figs. 1f-i: clearly, to detect such subtle phenomena one needs to evaluate all in terms of the energy gap , the supercurrent speed and the supercurrent trajectory . 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 normal state band dispersion:
where is the chemical potential, , , and are electron hopping parameters. The superconducting gap function is:
where is the maximum gap amplitude. For a supercurrent with velocity , the Bogoliubov band dispersion is modified by the Galilean boost yielding a Bogoliubov-DeGennes Hamiltonian[35]:
The Green’s function for eigenstates of is:
where
We employ an energy with a quasiparticle lifetime parameter [34]. The density of states can then be derived from the imaginary part of the Green’s function as:
Quantitively accurate simulations of the Galilean boost effects in Eqn. (7) for the superconductivity of are shown in Figs. 1f-i for various supercurrent velocities at angles , , , and relative to the order parameter. But, from symmetry, Eqn. (7) results in 8 supercurrent velocity vectors , all having the same (Methods C). Thus we introduce the quantities and being the maximal and minimal components of supercurrent velocity vector .
For training the MAIN and COGNATE models, we generate 3.5 million spectra using Eqn. (7) with energy in the range , in the range , in the range , and in the range (Methods C). Tight-binding parameters for are from angle resolved photoemission studies[36]. This synthetic dataset of 3.5 million simulated is used to train both the MAIN and COGNATE models of VMAP.
Experimentally, spectra are typically of the form , where both and are small parameters and unavoidable electronic noise also exists. Thus, Gaussian noise and a linear- background were added as augmentation to each Eqn. (7) simulated when training MAIN, with the noise level consistent with that measured in experimental data (Methods C3). In operation, the MAIN module determines four primary parameters , , , and , with the consequent supercurrent speed being (Methods C4). To establish best correspondence of VMAP determined with each experimental spectrum , the parameters , , , and are used by the COGNATE module. This backpropagates gradients of the physical parameters efficiently, allowing precise evaluation of (Methods C5). Ultimately, the quality of fit parameter is evaluated by comparing each measured to its VMAP reported spectrum. Our MAIN+COGNATE process increases the FOV-average to in general, as demonstrated directly using data videos in Supplementary. Each empirical data set consists of a topographic image with a simultaneous and atomically registered (Methods G). VMAP analysis of then yields images of the superconducting energy gap and of supercurrents . The validity of determination of , and thus of and , is exemplified directly in the Supplementary using three-panel videos of simultaneous which demonstrate high-quality VMAP fits to , throughout. Using a second ANN named V0MAP (Methods D) for which all training spectra are set at , we demonstrate by comparing for all correspondences from both ANNs, that supercurrent speeds m/s are indistinguishable from zero when using VMAP (Methods E). Nevertheless, the m/s supercurrent trajectories in can be visualized directly using
(Methods F) designed to identify the spatial locations where intense supercurrents appear. Finally, for , 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 -dependence of , , , and phenomena. Focusing on with relevant shown Methods G, the results of imaging by VMAP are shown in Figs. 2a-c. These are highly consistent with conventional heterogeneous studies[32, 33, 34]. VMAP imaging of in the FOV associated with each 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 -wave cuprate superconductor, with speeds reaching 4000 m/s. The statistical repeatability of phenomena is manifest. Separately visualized using and using 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 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 associated with each panel at left.
For Galilean boosts to (Fig. 1d) very specific characteristics should occur for different supercurrent directions relative to the symmetry order parameter (Figs. 1f-i and Methods J). In Figure 3a we explore these characteristics for numerous supercurrent loop trajectories identified by imaging. The color of dots indicates different directions relative to the order parameter. Figure 3b shows the 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 spectra on typical loop trajectories labeled 1#-6# in Figure 3a. Their detailed correspondence with Figure 3c underlines the validation of supercurrent speed imaging (Figure 3d-Figure 3i) using VMAP. Methods K and Extended Data Figure 10 show the equivalent correspondence as a function of .

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 images (e.g. Figs. 2d-f) by simultaneous images of loop trajectories (e.g. Figs. 2g-i respectively), to produce a new image 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 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 , where are the four nearest neighbors of site and values can be any of . 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 labeled . (Methods L). The measured 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 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 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 from , or conflation of heterogeneities from another degree of freedom (e.g. or ) into , there would be no reason whatsoever for to satisfy either a global continuity equation in (Figs. 4i-k) or the local Kirchhoff’s law constraint (Figs. 4b-h). Hence, these results sustain VMAP derived as images of persistent supercurrents.
The long-standing challenge has been to identify and understand -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 -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 -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 -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 -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 -symmetry breaking in highly doped superconducting cuprates. First, an hypothetical term 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 using the positive bias data and using the negative bias data, they are practically indistinguishable (Methods H) emphasizing both the particle-hole symmetry of and the adequacy of the experimental signal to noise ratio. Third, the correspondence between 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 visualization. Fourth, circulation is observably quantized around close supercurrent loops. Fifth, from theoretic models with Fermi velocity m/s, a typical disordered -wave supercurrent velocity is predicted[22] to be , in good agreement observations (Fig. 2). Sixth, approximate estimates of the absolute value of loop supercurrents based on the measured and estimated electron-pair density, yield A, so that that for nm radius domains the magnetic field at center is mT (Methods I). This is in reasonable agreement with reported spontaneous fields of -symmetry breaking in cuprates. Finally, the supercurrent images 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 -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 -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 −wave Galilean boost to Bogoliubov density of states from supercurrents
Figure 1. −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 symmetry superconductors. c. Schematic diagram of strong impurity scattering induced supercurrents, for symmetry superconductors. d. Schematic of the Galilean boost effect on Bogoliubov quasiparticle energy and consequent alteration of , due to supercurrents with speed . e. Schematic of a supercurrent loop oriented relative to the order parameter of a superconductor. The predicted is different for each supercurrent direction and speed as shown in panels f-i. This is the “Rosetta Stone” translating each spectrum to a specific speed V, and froms the training set for VMAP. f. spectra from Eqns. (2)-(7) for supercurrents with different speed at an angle 0 to −wave order parameter. g. spectra from Eqns. (2)-(7) for supercurrents with different speed at an angle to −wave order parameter. h. spectra from Eqns. (2)-(7) for supercurrents with different speed at an angle to −wave order parameter. i. spectra from Eqns. (2)-(7) for supercurrents with different speed at an angle to −wave order parameter.
Figure 2 VMAP determination of , and versus hole density
Figure 2.
a-c Imaging energy gaps versus hole-density using VMAP. Relevant topographies are shown in Methods G and Extended Data Fig. 6. Data demonstrating the overall quality of by VMAP is presented using data videos in Supplementary.
d-f Imaging supercurrent speeds in the same FOV as associated panel to left, versus hole-density using VMAP.
g-i Imaging the network of supercurrent loops in the same FOV as associated panel at left, versus hole-density using VMAP, where Å, (ns).
Figure 3 Individual supercurrent loops as identified by VMAP
a. Supercurrent speeds on designated current loops – in . 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 of all VMAP to measured , 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 and so that the rare examples of deficient fitting do not falsely create the images of supercurrent speeds.
c. Sequence of theoretical spectra on the trajectory of a supercurrent loop. Color code is as in Fig. 1e, designating distinct directions of supercurrent.
d. Sequence of spectra on the trajectory of loop .
e. Sequence of spectra on the trajectory of loop .
f. Sequence of spectra on the trajectory of loop .
g. Sequence of spectra on the trajectory of loop .
h. Sequence of spectra on the trajectory of loop .
i. Sequence of spectra on the trajectory of loop . The spectra in d-i are vertically shifted for clarity and have the corresponding speed listed in m/s.
Figure 1 (PDF p. 16).

Figure 4 (PDF p. 19).
Methods
Materials and measurements
Single crystals of were grown by the traveling solvent-floating-zone (TSFZ) method. The spectroscopic imaging scanning tunneling microscope (SISTM) measurements of crystals were performed on a custom instrument at 4.2 K. All samples were cleaved in situ in cryogenic ultra-high vacuum at 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 samples.
Method on supercurrent velocity analysis
The Galilean-boost energy is -dependent, while the density of states integrates over the whole BZ in -space, meaning that 8 supercurrents result in the same : , adding difficulty to the analysis. Earlier, we demonstrated a supercurrent visualization method[24] using a reference dataset, where each of the experimental spectra is compared individually to all of the reference dataset , and the maximal coefficient of determination produces the best matched parameters. However, with four-parameter -wave supercurrent fitting as in this project, such a reference dataset would contain 1.6 billion spectra. Hence this approach is inadequate for high data volume and heterogeneous cuprate data sets of a disordered -wave superconductor.
VMAP ANN architecture
To tackle the eight possible for each supercurrent velocity , we introduce two quantities, namely the higher (lower) vectorial components () of such that or . Thus
where and are referred to as taking the maximal and minimal values, respectively.
For a given parameter set , the within the energy range 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 curves are sufficient to train VMAP.
For MAIN, we aim to take the measured spectra as input and the four parameters 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 can produce similar curve. To improve the performance of MAIN, we design a COGNATE that identifies the spectra from the MAIN output parameter set that corresponds best to the input .
The COGNATE adopts a transformer-like architecture37 (Extended Data Figure S1). Each of the four parameters 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 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, , where 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 .

Figure S1. Supplementary Fig. 1. maps at different and (a-i) The maps for different from 2.0 Å to 5.0 Å and at different from 0.05 (ns) to 0.20 (ns).
The MAIN module is designed primarily to extract the four parameters from a given spectrum where is given (Extended Data Figure S1). In reality, a linear background signal may be present in the experimental spectra which are its data source. Thus, we make the augmentation before the training dataset goes into MAIN, that is:
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 learnable convolution layer connected with two learnable convolution layers and the second branch is one learnable convolution layer. In this work, the non-one kernel sizes are 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 , , and 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 spectra using Eqn. (7), with energy in the range [ meV, 100 meV], in the range [0 meV, 80 meV], in the range [0 meV, 10 meV], and 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 curve, we use a dense -mesh of size and mask the effective -points in the energy range [ meV, 200 meV]. The same synthetic dataset is used to train both the COGNATE and the MAIN, where 3.15 million (90%) synthetic curves were used for training and 0.35 million (10%) synthetic 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 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 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 , where is the synthetic spectrum with its mean value , and 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 in a linear range from to with 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.

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 and take the coefficient of determination to evaluate the accuracy (Extended Data Figure 2).
where is the target spectrum and is the corresponding reconstructed curve at energy . The COGNATE was trained for 100 epochs. For the MAIN, we use the mean absolute error (MAE) loss function :
where , (), and are the predicted parameters and the corresponding true values, respectively. The accuracy of each parameter is defined as:
where the index 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 computed through the COGNATE to accelerate the convergence of the MAIN. The is the same as as in the COGNATE. The MAIN loss function is then the combined and :
where is a relative ratio and we take 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 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 first is processed by MAIN, which outputs the four parameters , but this is not the final output. Instead, we use the as input of the COGNATE, which generates the reconstructed spectrum . By fitting the ratio to a linear function , where is the target experimental spectrum, and can be obtained. We then use the MSE loss in the refinement procedure:
In this scenario, the four physical parameters receive gradients through the frozen COGNATE and are iteratively updated via the AdamW optimizer to minimize . In this work, we perform 2000 optimization steps for each of the experimental spectrum and track the best parameters with the lowest . The schematic of the refinement process is shown in Extended Data Figure 3 and 3b.

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 spectra. Then those parameters, as the initial step, are input to the COGNATE model, where the background signal 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 :
in which is the measured differential conductance spectrum and is the numerical recomputed spectrum from Eqn. (7) based on the given six parameters, is the corresponding mean value of . Extended Data Figure 3-Extended Data Figure 3 shows an example of the before and after the refinement procedure, from which the mean value of 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 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 (). In terms of the Hamiltonian Eqn. (4) which generates the 3.5 million training spectra for V0MAP, the term . In this case, the V0MAP ANN searches for best fits using only , and the background $a(r)+ as degrees of freedom. The comparison obtained from VMAP and V0MAP is presented in Extended Data Figure 4.
Figure 4. a. To identify both the supercurrent speeds and trajectories we masked VMAP measured by the s/m trajectory image with typical results shown here. The 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 images obtained from in Fig. 2. A minimal boxcar averaging was applied to before calculating the and .
Indistinguishability from zero of small supercurrent velocities
We compute the without supercurrent velocity and label it as , then the quantity is defined as , where is computed with finite supercurrent velocity. Extended Data fig:4a-c show the supercurrent velocity-dependent , which is averaged at each velocity, indicating that the Galilean-boosted has a better description on the in general. Combined with the supercurrent velocity-dependent at different doping levels, the conservative indistinguishability of small supercurrent speed is () 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 mask onto the (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 within those supercurrent-velocity indistinguishable regions, where the results are shown in Extended Data fig:5a-c, the average superconductive gap from those domains meV for hole doping , 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 -space, we adopt the two-dimensional curvature method on the supercurrent velocity maps 40. For the square grid, where the pixel distances are equal along the and axes with the same units, the two-dimensional curvature image of the target image can be defined as (also see in Eqn. (8)):
where is a positive free parameter. For simplicity, we take in units of (ns), in units of m/s, and and in units of nm; then the curvature image when calculated on the function has units of s/m in this work. In the limit of , approaches the Laplacian with a negative sign, and with , will diverge at the extrema of , giving 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 when applying this negative second derivative method to the supercurrent velocity image . We also employ a two-dimension Gaussian kernel with standard deviation to smooth the before computing the curvature map . Supplementary Figure 1 shows a series of maps with different and .
Topographies of the samples for different hole doping
Extended Data Figure 6 shows the topography images of the in the same field of view as in Figure 2, with hole doping from to 0.22, from which the one-dimensional super-modulation gradually disappears with increasing the doping level. At the hole doping , the unidirectional super-modulation is almost invisible.

Extended Data Figure 6. Topography images of the BiSrCaCuO with different hole doping
(a) The topography image of optimal BiSrCaCuO at doping in the same FOV as Figs. 2a, d, g ( mV, pA). (b) The topography image of over doped BiSrCaCuO at doping in the same FOV as Figs. 2b, e, h ( mV, pA). (c) The topography image of over doped BiSrCaCuO at doping in the same FOV as Figs. 2c, f, i ( mV, pA).
Separate imaging of supercurrent speeds for filled/empty states
In theory for the superconducting density of states, the Galilean-boosted 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 from the positive and negative energy regimes and perform the VMAP analysis independently. Here it is important to note that the random variations in and which are generated by finite SNR are completely uncorrelated. Effectively, these are two independent experiments.
The resulting supercurrent velocity and are responses to the positive energy and negative energy , respectively. In this work, the experimental spectra are always collected in an energy-symmetric window , thus we can directly evaluate 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 and in the same FOV as in Fig. 2 with the hole doping , , and , from which the and are virtually identical. Although these are effectively two separate independent experiments, the images derived by VMAP from are virtually identical to the images derived from 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 is:
where is the effective charge in the supercurrent case and is the electron charge, is the number of flowing pairs, is the radius of the circle, and 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:
Where is the vacuum magnetic permeability. In Extended Data Figure 5-Extended Data Figure 5, we show the histograms of the supercurrent velocity of , where the averaged supercurrent velocity is m/s throughout the hole doping to . To estimate the radius of the supercurrent loop, we compute the autocorrelation of 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 () of the loops. The corresponding average radius of the supercurrent loop is nm. By fitting the line cuts in the autocorrelation of with an exponential function

Extended Data Figure 5. Superconducting gap patterns and supercurrent velocities. (a-c) The spatial distribution of the masked superconductive gaps with hole doping (a), (b), and (c), showing the patterns of the superconductive gap domains. (d-e) The histogram of superconductive gap from (a-c) with hole doping (d), (e) and (f), showing the averaged meV (d), meV (e), and meV (f), respectively. (g-i) The histogram of supercurrent velocity with hole doping (g), (h) and (i), showing the averaged m/s (g) m/s (h) and m/s (i), respectively.
where and are the free parameters and is the exponential decay length. The decay length from the center of autocorrelation of is computed, from which the averaged 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 nm. The electron density in an uncorrelated two-dimension electron gas is:
On the loop with radius and loop width , the corresponding electron population is , resulting in the Cooper pair population within the loop width
Taking the values of , , and into Eqn. (23), the estimated number of flowing pairs on a supercurrent loop is , the corresponding supercurrent intensity A, and the corresponding effective magnetic field mT. This estimated effective magnetic field is expected and consistent with the SR-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 at depth below the CuO plane produces an in-plane vector potential whose maximum is 41. The London response of the condensate to it would be a supercurrent of speed 42. For at nm this gives m/s, and even for a cluster with only 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 -wave spectra in presence of a supercurrent
For a uniform superfluid velocity the condensate acquires a center-of-mass momentum , which Galilean boosts the quasiparticle spectrum ( is the normal-state band dispersion) by
The local density of states is then obtained from the imaginary part of the current-carrying Green’s function,
with for a superconductor over a circular Fermi surface. If the superfluid velocity makes an angle with the crystal -axis, Fermi surface , 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 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 , so the excitation energy along the Fermi surface is . The Bogoliubov dispersion then possesses four symmetry-related saddle points at the antinodes, and , where and . These saddle points give rise to van Hove singularities in the quasiparticle density of states at , producing the familiar BCS coherence peaks. A finite superfluid velocity lifts the degeneracy of the four antinodal saddle points through the Galilean boost .
For superfluid velocity directed along an antinodal axis (), the Fermi velocities at the antinodes are perpendicular to the flow and therefore experience no Galilean boost, while those at are shifted by . Consequently, each coherence peak evolves into a characteristic three-peak structure consisting of an unsplit central peak at , arising from the two unaffected antinodes, together with two side peaks shifted upward and downward by the Galilean boost energy (Extended Data Figure 9).

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 , all four antinodes acquire Galilean boost of equal magnitude , 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 order parameter.
Supercurrent loop trajectories for different hole densities
Extended Data Fig. 10 shows the supercurrent loop trajectories, across the whole FOV, with hole dopings from to . 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 -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 is calculated by masking the with the s/m map to get a speed-resolved network image. This is shown in Fig. 4a for .
Due to the network pattern of , 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 for each site , where the sum of observed supercurrent speeds through a site is derived from its four neighbor sites . VMAP determines the larger and smaller of the two orthogonal velocity components, and , at each pixel . Therefore, the supercurrent speed along the horizontal axis at pixels is one of four values ; , while the supercurrent speed along the vertical axis at pixels is also one of four values ; . This means that for each of the 4 sites surrounding pixel there are 4 possible supercurrents. Thus there are possible combinations of , for all four sites surrounding site . For each pixel we then evaluate all possible values of:
where , , and denote the right, left, lower and upper neighbors spanning . At site , the relevant components (, for and ; , for and $U) at each neighbor then take any of the four measured values , . is then defined as the minimum value of over all 256 possible combinations. In practice, we implement a minimal boxcar averaging on the superfluid current map to enhance the signal-to-noise ratio before computing . The results of are shown in the Fig. 4 and Supplementary Extended Data Figure 2, where the mean value of is between 15 m/s and 20 m/s throughout all , the corresponding ratio 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 -space obtained from the VMAP associated with the and in a typical FOV with hole doping , and the corresponding normalized cross-correlation coefficients among different parameters are shown in Supplementary Table 2. The normalized cross-correlation coefficient is:
where and are the target images after the mean-value subtracted. When we measure the normalized cross correlation between and (where and are the output from VMAP), it is typically for all samples studied (Supplementary Table 2).
Methods-only References

Figure Extended Data Fig. 1. Extended Data Fig. 1. VMAP architecture. (a) The architecture of the COGNATE, which takes the energy variable , , , , and as input and outputs . (b) The architecture of the MAIN, which takes the energy and as input, and outputs the four parameters: , , , and . (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).
without the COGNATE refinement. (d) The histogram of as in (c), showing the mean value . (e) The spatial distribution of the coefficient of determination after the COGNATE refinement in the same FOV as (c). (f) The histogram of as in (e), showing the mean value .

Extended Data Figure 4. Indistinguishable small supercurrents m/s from . (a) The supercurrent velocity-dependent in the sample with hole doping , where the positive correlation indicates that the Galilean-boosted provides a better description than without supercurrent velocity, especially in the high supercurrent regime. (b-c) The same as in (a) but in the sample with hole doping (b) and (c). A boxcar average, with supercurrent velocity of 250 m/s, was applied to (a-c).

Figure 7. and identified by the VMAP
(a–c) Example spectra at (a), (b) and (c), fitted separately for (giving ) and for (giving ). (d–f) The identified by the VMAP with different hole doping (d), (e), and (f). (g–i) The identified by the VMAP with different hole doping (g), (h), and (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 with hole doping . (b) The auto-correlation map of with hole doping . (c) The auto-correlation map of with hole doping . (d-i) The line cuts of marked in (a-c), indicating the averaged supercurrent loop
radius nm (half of the dimeter), and the exponential function fitting (the solid orange curves) give the averaged loop with nm.
(a) The as in the Supplementary Figure S1-Extended Data Figure 2, where the hole doping . (b-d) The examples of individual supercurrent loops identified by the VMAP, the loops 1#-3# are marked in (a) ( mV, pA, mV). The spectra in (b-d) are vertically shifted for clarity. (e) The as in the Fig. 2, where the hole doping . (f-h) The examples of individual supercurrent loops identified by the VMAP, the loops 1#-3# are marked in (e) ( mV, pA, mV). The spectra in (f-h) are vertically shifted for clarity. (i) The as in the Fig. 2, where the hole doping . (j-l) The examples of individual supercurrent loops identified by the VMAP, the loops 1#-3# are marked in (i) ( mV, pA, mV). The spectra in (j-l) are vertically shifted for clarity.
Supplementary Information
| (meV) | (meV) | (meV) | (meV) | |
| 0.17 | 407 | 360 | 108 | 36 |
| 0.20 | 414 | 360 | 108 | 36 |
| 0.22 | 427 | 360 | 108 | 36 |
Table 1. The tight binding parameters for with hole doping , 0.20, and 0.22 studied in this work.
| (Small FOV) | ||||
| 0.49 | 0.44 | 0.56 | 0.40 | |
| 0.33 | 0.25 | 0.28 | 0.39 | |
| 0.55 | 0.04 | -0.17 | -0.06 | |
| 0.55 | 0.52 | 0.35 | 0.50 | |
| 0.28 | 0.09 | 0.19 | 0.13 | |
| 0.20 | -0.14 | -0.40 | -0.26 | |
| 0.33 | 0.35 | 0.31 | 0.33 | |
| 0.24 | -0.01 | -0.22 | 0.02 | |
| 0.21 | 0.16 | 0.10 | 0.24 | |
| 0.77 | -0.52 | -0.54 | -0.52 |
Table 2. Supplementary Table 2. The cross-coefficient among different parameters obtained from VMAP for different hole doping .

Figure 2 (PDF p. 54). Simultaneous determination of different parameters in a typical FOV
The VMAP analysis with hole doping 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) ( mV, pA, mV).
Supplementary Videos
To demonstrate validity of determination of and by the VMAP ANN we have prepared three-panel Video of simultaneous : Supplementary Video 1 for , Supplementary Video 2 for , and Supplementary Video 3 for . Because the manifestly yields values that are virtually zero for all , this demonstrates directly and visually the excellent quality of the fits to , which are generated by VMAP assignment of , throughout this work.
References
- [2]Sonier, J. E. et al. Anomalous Weak Magnetism in Superconducting YBa₂Cu₃O₆₊ₓ. Science 292, 1692–1695 (2001).pubmed.ncbi.nlm.nih.gov/11387468
- [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
- [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
- [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
- [6]Li, Y. et al. Unusual magnetic order in the pseudogap region of the superconductor HgBa₂CuO₄₊δ. Nature 455, 372–375 (2008).DOI
- [7]Bourges, P., Bounoua, D. & Sidis, Y. Loop currents in quantum matter. C. R. Physique 22, 7–31 (2021).DOI
- [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
- [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
- [10]Zhao, S. Y. F. et al. Time-reversal symmetry breaking superconductivity between twisted cuprate superconductors. Science 382, 1422–1427 (2023).DOI
- [11]Qi, S. et al. High-temperature field-free superconducting diode effect in high-Tc cuprates. Nat. Commun. 16, 531 (2025).DOI
- [12]Andersen, B. M., Kreisel, A. & Hirschfeld, P. J. Spontaneous time-reversal symmetry breaking by disorder in superconductors. Front. Phys. 12, 1353425 (2024).DOI
- [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).
- [14]Håkansson, M., Löfwander, T. & Fogelström, M. Spontaneously broken time-reversal symmetry in high-temperature superconductors. Nat. Phys. 11, 755–760 (2015).DOI
- [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
- [16]Kaur, R. P. & Agterberg, D. F. Broken time reversal symmetry and superconducting states in the cuprates. Phys. Rev. B 68, 100506 (2003).DOI
- [17]Varma, C. M. Theory of the pseudogap state of the cuprates. Phys. Rev. B 73, 155113 (2006).DOI
- [18]Varma, C. M. Pseudogap in cuprates in the loop-current ordered state. J. Phys. Condens. Matter 26, 505701 (2014).DOI
- [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
- [20]Kivelson, S. A. & Spivak, B. Macroscopic character of composite high-temperature superconducting wires. Phys. Rev. B 92, 184502 (2015).DOI
- [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
- [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
- [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
- [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
- [25]Zhang, Y. et al. Machine learning in electronic-quantum-matter imaging experiments. Nature 570, 484–490 (2019).
- [26]Arovas, D. P., Berg, E., Kivelson, S. A. & Raghu, S. The Hubbard Model. Annu. Rev. Condens. Matter Phys. 13, 239–274 (2022).
- [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
- [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
- [29]Yip, S.K. & Sauls, J. A. Nonlinear Meissner effect in CuO superconductors. Phys. Rev. Lett. 69, 2264 (1992).DOI
- [30]Volovik, G.E. Superconductivity with lines of GAP nodes: density of states in the vortex. Sov. Phys. JETP Lett. 58, 469 (1993).
- [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
- [32]Lang, K. M. et al. Imaging the granular structure of high-Tc superconductivity in underdoped Bi₂Sr₂CaCu₂O₈+δ. Nature 415, 412–416 (2002).DOI
- [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
- [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
- [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
- [36]Drozdov, I. K. et al. Phase diagram of Bi₂Sr₂CaCu₂O₈+δ revisited. Nat. Commun. 9, 5210 (2018).DOI



