Paper deep dive
Fast holographic inversion of superconducting domes
Sejin Kim
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 10/11/2026, 3:35:19 AM
Summary
This paper introduces a computationally efficient method for inverting superconducting phase diagrams (domes) to determine the scalar mass function M(F^2) in holographic superconductor models. By utilizing an analytic Jacobian derived from Wronskian identities, the inversion time is reduced from days to minutes/hours. The study identifies a 'blind region' in the mass function that cannot be determined by the dome shape alone and proposes scanning the scalar mass at the boundary (M(0)) to resolve this ambiguity, finding that while the dome constrains the mass where horizons reach, fixing the dual operator dimension requires a second observable.
Entities (8)
Relation Signals (6)
Holographic Superconductor → hasparameter → Scalar Mass Function M(F^2)
confidence 98% · A holographic superconductor whose scalar mass depends on the gauge field strength, M(F^2), reproduces a superconducting dome
YBa2Cu3Oy → hasphasediagram → Superconducting Dome
confidence 97% · measured phase diagrams of YBa2Cu3Oy... are domes
Scalar Mass at Boundary M(0) → determines → Dual Operator Dimension
confidence 96% · M(0) is the scalar mass M(0) that fixes the dimension of the dual operator through Delta(Delta-3)=M(0)
Wronskian Identity → enables → Fast Inversion
confidence 95% · Here that condition is obtained... from two integrations... so the gradient needs no integration of its own... an inversion takes from about ten minutes to an hour
Levenberg-Marquardt → usedfor → Inversion
confidence 95% · the natural method is not gradient descent but Levenberg–Marquardt... an inversion then takes from about ten minutes to an hour
Blind Region → limits → Scalar Mass Function M(F^2)
confidence 94% · The dome thus constrains M where its horizons reach... on the interval between the value F^2 takes at the horizon... and F^2=0... the dome does not determine M
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:A holographic superconductor whose scalar mass depends on the gauge field strength, $M(\Fsq)$, reproduces a superconducting dome for a suitable $M$, and recovering that $M$ from a given dome has so far taken days for a single training run. We propose a new way of training this model, with which an inversion takes from about ten minutes to an hour. Training needs the gradient of the condition that fixes the critical temperature, which the earlier method obtains by finite differences, repeating the bulk integrations for every training parameter. Here that condition is obtained, without any fit, from two integrations started at the horizon and at the boundary, and its derivative with respect to $M$ is an integral over the same two solutions, so the gradient needs no integration of its own. We use the speed to study the part of $M$ that a dome cannot determine, on the interval between the value $\Fsq$ takes at the horizon for the lowest doping and $\Fsq=0$, at which $M$ is the scalar mass $M(0)$ that fixes the dimension of the dual operator. We hold the scalar mass at several values, which we call pinned masses, retrain everything else at each, and find that the reconstructions agree wherever the horizons of the dome reach, including the minima of $M$, and differ only on that interval. A rule that keeps the reconstruction with the simplest closed form recovers both the scalar mass and the mass function of a test dome. On Gaussian and double-Gaussian domes and on the measured phase diagrams of YBa$_{2}$Cu$_{3}$O$_{y}$ and 2M-WS$_{2}$, however, the pinned mass it keeps rests on ties or on narrow margins, so for these targets the scalar mass is left open. The dome thus constrains $M$ where its horizons reach, and fixing the dimension of the dual operator needs a second observable.
Tags
Links
- Source: https://arxiv.org/abs/2610.09938v1
- Canonical: https://arxiv.org/abs/2610.09938v1
Trouble viewing inline? Open PDF directly →
Full Text
77,783 characters extracted from source content.
Expand or collapse full text
Fast holographic inversion of superconducting domes Sejin Kim Affiliation: Center for AI and Natural Sciences, Korea Institute for Advanced Study, Seoul 02455, Korea Email: sejin@kias.re.kr Abstract A holographic superconductor whose scalar mass depends on the gauge field strength, M(F2)M(F^2), reproduces a superconducting dome for a suitable M, and recovering that M from a given dome has so far taken days for a single training run. We propose a new way of training this model, with which an inversion takes from about ten minutes to an hour. Training needs the gradient of the condition that fixes the critical temperature, which the earlier method obtains by finite differences, repeating the bulk integrations for every training parameter. Here that condition is obtained, without any fit, from two integrations started at the horizon and at the boundary, and its derivative with respect to M is an integral over the same two solutions, so the gradient needs no integration of its own. We use the speed to study the part of M that a dome cannot determine, on the interval between the value F2F^2 takes at the horizon for the lowest doping and F2=0F^2=0, at which M is the scalar mass M(0)M(0) that fixes the dimension of the dual operator. We hold the scalar mass at several values, which we call pinned masses, retrain everything else at each, and find that the reconstructions agree wherever the horizons of the dome reach, including the minima of M, and differ only on that interval. A rule that keeps the reconstruction with the simplest closed form recovers both the scalar mass and the mass function of a test dome. On Gaussian and double-Gaussian domes and on the measured phase diagrams of YBa2Cu3Oy and 2M-WS2, however, the pinned mass it keeps rests on ties or on narrow margins, so for these targets the scalar mass is left open. The dome thus constrains M where its horizons reach, and fixing the dimension of the dual operator needs a second observable. 1 Introduction Holographic duality Maldacena (1998); Gubser et al. (1998); Witten (1998) turns a strongly coupled phase transition into a boundary-value problem for classical fields in one more dimension, and the holographic superconductor Gubser (2008); Hartnoll et al. (2008a); Hartnoll et al. (2008b) is its standard example. Below a critical temperature TcT_c a charged black hole in anti-de Sitter (AdS) space acquires charged scalar hair, and the region in which such hairy black holes exist is the superconducting phase Hartnoll (2009); Herzog (2009); Horowitz (2011); Cai et al. (2015). TcT_c is the temperature at which the linearized scalar first admits a normalizable mode regular at the horizon, the condensation mode. Measured phase diagrams of many superconductors are domes, in which the critical temperature rises and then falls with doping Keimer et al. (2015). Holographic models produce domes from a modulated chemical potential Ganguli et al. (2013), from competing orders Kiritsis and Li (2016), also with broken translations Baggioli and Goykhman (2016), from a doping field that shifts the effective mass of the order parameter Chen et al. (2016a) and from a coupling between two charges Cai and Sin (2021), and a dome was obtained by letting the scalar mass depend on the local field strength, M(F2)M(F^2) Seo et al. (2024); Kim et al. (2025). The difficulty is finding such a holographic model. A bulk theory that reproduces an observed phenomenon has been found by trial and error, a coupling chosen by hand, the phase diagram computed and compared with the measured one, and the coupling adjusted, so the model is tuned to the known result. Solving the inverse problem with neural networks removes the tuning of the coupling by hand. Given a dome, the critical temperature against doping, find an M(F2)M(F^2) that produces it, so that the bulk interaction is fixed by reproducing the measurement rather than chosen by the modeler. Recovering bulk data from boundary observables has its own history. Metrics have been recovered from geodesic or two-point data Hammersley (2006); Bilson (2008), from extremal surfaces Hubeny (2012); Engelhardt and Horowitz (2017) and from entanglement entropy Spillane (2013). More recently the bulk has been represented by a neural network, a composition of affine maps and pointwise nonlinearities adjusted until the output matches data Jeong et al. (2025); Park et al. (2022), or produced directly from boundary data by a generative model trained on many geometries Park et al. (2026); Kim (2026). The radial evolution of a bulk field was interpreted as the layers of a deep network Hashimoto et al. (2018b); Hashimoto et al. (2018a), recast as a Boltzmann machine Hashimoto (2019) and as a neural ordinary differential equation Hashimoto et al. (2021), and used to learn spatial geometry from entanglement You et al. (2018), the metric and dilaton of a holographic model of quantum chromodynamics from meson spectra Akutagawa et al. (2020), black-hole metrics from the shear viscosity Yan et al. (2020) and a bulk spacetime from the frequency-dependent conductivity Ahn et al. (2024). Recently the geometry and gauge field of a charged black hole have been reconstructed from fermionic spectral functions Hashimoto et al. (2026); Xiao et al. (2026), the latter applied to the strange-metal phenomenology of a cuprate, the metric from entanglement entropy and Wilson loops Filev (2026) and the bulk scalar potential from the equation of state Bea et al. (2024); Jimenez et al. (2026), and each of these reconstructions meets a degeneracy or a region of the bulk that the boundary data do not probe. The dome itself has been inverted once Kim et al. (2025), with a physics-informed neural network (PINN) Raissi et al. (2019); Karniadakis et al. (2021), one whose objective is the residual of the governing equation rather than a fit to labeled examples. The mass function it finds is lowest at the field strengths that correspond to the top of the dome, since a more negative effective mass near the horizon raises the critical temperature. Its computational cost limits what it can be used for. The source J is extracted by fitting the near-boundary expansion of a solution integrated from the horizon, and the objective built from it is minimized over roughly 10310^3 iterations of Adam Kingma and Ba (2015), a first-order method with per-coordinate adaptive steps. Each iteration takes a finite-difference gradient, np+1n_p+1 bulk integrations per point of the dome for npn_p parameters, so a single reconstruction is already expensive, and any number training does not determine has to be chosen once and then kept. We change the forward computation first. The solution ψH _H regular at the horizon is integrated from the horizon and the normalizable solution ψB _B from the boundary, both are stopped at an interior point, and their Wronskian W is computed there. On shell W does not depend on where it is evaluated, and it vanishes exactly when the regular solution is normalizable, so W=0W=0 is the condition =0J=0, obtained without fitting the near-boundary expansion. Each point of the dome fixes a background, and W evaluated on it is that point’s residual, the number training drives to zero. The backward computation then reuses the two forward solutions. In the radial coordinate z, running from the boundary at z=0z=0 to the horizon at z=1z=1, a Green’s identity gives, for any parameter θj _j of M that leaves M(0)M(0) fixed, ∂W∂θj=∫01ψH(z)ψB(z)∂M∂θj(F2(z))dz4, ∂ W∂ _j= _0^1 _H(z)\, _B(z)\, ∂ M∂ _j (F^2(z) )\, dzz^4, (1) an integral of the product of the two forward solutions already computed, with ψH _H normalized by its value at the horizon and ψB _B by its leading coefficient at the boundary. The whole Jacobian, the matrix of derivatives of the residuals with respect to the parameters, is therefore obtained together with the residual at any number of parameters. With an analytic Jacobian, exact where the residuals vanish (Appendix A), the natural method is not gradient descent but Levenberg–Marquardt Levenberg (1944); Marquardt (1963), which for a sum of squares builds a local quadratic model from the residual sensitivities and solves for the step. At the hundred points per dome used here an inversion then takes from about ten minutes to an hour instead of days. A faster inversion does not remove the part of M that the dome leaves undetermined. Every dome starts at a finite doping, and between the value F2F^2 takes at the horizon for that doping and F2=0F^2=0 lies an interval that contains no horizon value of the dome and on which the residual depends only weakly. On that interval the dome does not determine M. What is undetermined there is a function and not a number, and its value at F2=0F^2=0 is the scalar mass M(0)M(0), the value of M at the AdS boundary, which fixes the dimension Δ of the dual operator through Δ(Δ−3)=M(0) ( -3)=M(0). Section 3 calls that interval the blind region. A single training run returns some function there, determined by the regularization, the basis and the starting point rather than by the dome. The faster inversion makes M(0)M(0) something one can scan rather than something one must choose. At this speed M(0)M(0) can be held at several values m0m_0, which we call pinned masses, everything else retrained at each, and the whole family inspected. One value in the scan is natural, m0=−3/4m_0=-3/4. From m0=−3/4m_0=-3/4 upward the scalar respects the stability bound of the near-horizon AdS2 and no critical temperature survives at zero doping, and we place the pinned masses there, at m0=−3/4m_0=-3/4, −1/2-1/2 and −1/4-1/4. Choosing among the members of the scan needs a rule of its own. Each reconstruction is fitted by closed forms from a library fixed in advance, and among those within 5%5\% of a parabola across the blind region the one reproduced with the fewest symbolic parameters, the numbers a closed form needs, is kept. This is a restricted form of symbolic regression, the search for a closed-form expression that reproduces a curve Schmidt and Lipson (2009); Udrescu and Tegmark (2020), with a description-length preference for the shortest Rissanen (1978). On every dome the reconstructions at the three pinned masses share the positions and depths of the minima of M and agree over the range its horizons reach, apart from the deepest end, while they differ in the blind region (Sec. 5.2), so the dome constrains M there but not its value at F2=0F^2=0. On a dome solved forward from a quadratic M with M(0)=−1/2M(0)=-1/2, made after the conventions of the rule had been fixed, the rule picks the reconstruction held at −1/2-1/2 and returns the quadratic itself (Sec. 5.1). On the Gaussian and double-Gaussian domes and on the measured phase diagrams of YBa2Cu3Oy (YBCO) and 2M-WS2 it keeps one pinned mass per target (Sec. 5.2), but each is decided by a tie, by one symbolic parameter or by the 5%5\% threshold, so the scalar mass of these four targets is left open. Fixing it needs a second observable, which Sec. 6 discusses. 2 Setup We work with a charged scalar on the Reissner–Nordström (RN) black hole in AdS4 with two linear axions Andrade and Withers (2014), the background of earlier holographic superconductors with momentum relaxation Andrade and Gentle (2015); Kim et al. (2015); Baggioli and Goykhman (2015) and the one on which the dome inversion was first run Kim et al. (2025). The scalar carries charge q, and its mass is a function of the invariant F2≡FμνFμνF^2≡ F_μνF^μν, so that M(F2)|ψ|2M(F^2)|ψ|^2 is a nonlinear interaction between the superconducting order and the charge density Seo et al. (2024); Kim et al. (2025). With Newton’s constant set by 16πGN=116π G_N=1 and unit AdS radius the action is S=∫d4x−g[R+6−14F2−12∑a=12(∂χa)2−|Dψ|2−M(F2)|ψ|2],S= d^4x -g\, [R+6- 14F^2- 12 _a=1^2(∂ _a)^2-|Dψ|^2-M(F^2)|ψ|^2 ], (2) with χa _a the axion fields, Dμ=∇μ−iqAμD_μ= _μ-iqA_μ and q=1q=1 throughout, so the M an inversion returns is the one that goes with this charge. The field equations of Eq. (2) are Rμν−12Rgμν−3gμν=12(1+4M′(F2)|ψ|2)FμρFνρ+12∑a=12∂μχa∂νχa+12[Dμψ(Dνψ)∗+Dνψ(Dμψ)∗]−12gμν[14F2+12∑a=12(∂χa)2+|Dψ|2+M(F2)|ψ|2],∇μ[(1+4M′(F2)|ψ|2)Fμν]=iq[ψ∗Dνψ−ψ(Dνψ)∗],∇μ∇μχa=0,DμDμψ−M(F2)ψ=0, gathered aligned R_μν- 12Rg_μν-3g_μν&= 12 (1+4M (F^2)|ψ|^2 )F_μρF_ν^ρ\\ & + 12 _a=1^2 _μ _a _ν _a+ 12 [D_μψ(D_νψ)^*+D_νψ(D_μψ)^* ]\\ & - 12g_μν [ 14F^2+ 12 _a=1^2(∂ _a)^2+|Dψ|^2+M(F^2)|ψ|^2 ],\\ _μ [ (1+4M (F^2)|ψ|^2 )F^μν ]&=iq [ψ^*D^νψ-ψ(D^νψ)^* ], aligned\\ _μ∇^μ _a=0, D_μD^μψ-M(F^2)\,ψ=0, gathered (3) with M′=dM/dF2M =dM/dF^2. The dome is determined in the regime |ψ|≪1|ψ| 1. At TcT_c the condensate is infinitesimal and its contributions to Eq. (3) drop out at linear order, so the background is the solution with ψ=0ψ=0. In the radial coordinate z, with the AdS boundary at z=0z=0 and the horizon at z=1z=1, it is the RN black hole with axions Andrade and Withers (2014), ds2=1z2[−hdt2+dz2h+dy12+dy22],A=Q(1−z)dt,χa=κya,h=(1−z)[1+z+(1−κ22)z2−Q24z3], gatheredds^2= 1z^2 [-h\,dt^2+ dz^2h+dy_1^2+dy_2^2 ], A=Q(1-z)\,dt, _a=κ y_a,\\ h=(1-z) [1+z+ (1- κ^22 )z^2- Q^24z^3 ], gathered (4) at temperature T=(3−κ2/2−Q2/4)/4πT=(3-κ^2/2-Q^2/4)/4π. The axions, linear in the boundary coordinates yay_a, break translations homogeneously and relax momentum at a rate set by κ Andrade and Withers (2014). Here they serve two purposes only, to supply the scale κ in which doping and temperature are measured below and, at zero charge, an extremal horizon at κ2=6κ^2=6 whose near-horizon AdS2 region sets the window of Sec. 3, and no transport property is used. Near the boundary the asymptotic expansion of each field defines a source and a response. The scalar behaves as ψ=z3−Δ+zΔ+…,Δ(Δ−3)=M(0),ψ=Jz^3- +Oz +…, ( -3)=M(0), (5) so that J is the source of the dual operator of dimension Δ and O is proportional to its expectation value. Only the mass at F2=0F^2=0 enters Δ , because F2F^2 vanishes at the boundary. A superconducting instability is a normalizable mode, =0J=0 with ≠0O≠ 0, and TcT_c is the highest temperature at which one exists. On this background the scalar obeys (hz2ψ′)′+[q2At2z2h−M(F2)z4]ψ=0. ( hz^2\,ψ ) + [ q^2A_t^2z^2h- M(F^2)z^4 ]ψ=0. (6) For two solutions ψ1 _1 and ψ2 _2 the Wronskian and its derivative, in which the terms in ψ1′ψ2′ _1 _2 cancel, are W=(hz2)(ψ1ψ2′−ψ1′ψ2), W= ( hz^2 ) ( _1 _2 - _1 _2 ), W′=ψ1(hψ2′z2)′−ψ2(hψ1′z2)′. W = _1 ( h _2 z^2 ) - _2 ( h _1 z^2 ) . (7) Equation (6) makes both terms of W′W equal to (M(F2)/z4−q2At2/(z2h))ψ1ψ2 (M(F^2)/z^4-q^2A_t^2/(z^2h) ) _1 _2, so W′=0W =0 and W is independent of z. Regularity at the horizon fixes ψ′ψ|z→1=−M(Fh2)4πT, ψ ψ |_z→ 1=- M(F^2_h)4π T, (8) with Fh2F^2_h the horizon value of F2F^2. The phase diagram is drawn in the dimensionless plane (x,τ)=(Q/κ2,T/κ)(x,τ)=(Q/κ^2,T/κ), charge density and temperature in units of the momentum-relaxation scale, with x playing the role of doping and a dome the curve τc(x)=Tc/κ _c(x)=T_c/κ. Because τ decreases monotonically in κ at fixed x, each point of a dome fixes (κ,Q)(κ,Q) uniquely and independently of M, before any training. The one property of Eq. (4) that the rest of the paper relies on is the range of F2F^2. On each background F2=−2Q2z4F^2=-2Q^2z^4 runs from 00 at the boundary to Fh2=−2Q2F^2_h=-2Q^2 at the horizon, so a point of the dome at doping x probes M on [Fh2,0][F^2_h,0] and nowhere else. At the low temperatures of a dome κ is close to its extremal value, at which Q=(1+12x2−1)/xQ=( 1+12x^2-1)/x grows with x, so |Fh2||F^2_h| grows with doping and points at higher doping probe M at more negative F2F^2. The horizon values of a dome fill the horizon range [Fmin2,Fmax2][F^2_ ,F^2_ ], with Fmax2=−2Qmin2F^2_ =-2Q_ ^2 from the smallest doping and Fmin2=−2Qmax2F^2_ =-2Q_ ^2 from the largest. No Fh2F^2_h falls in (Fmax2,0](F^2_ ,0], although on every background F2F^2 takes all values in that interval between the horizon and the boundary. The four domes inverted in Sec. 5.2 are, in order, a Gaussian and a double Gaussian specified directly in (x,τ)(x,τ) and the measured phase diagrams of YBCO and 2M-WS2. 3 The blind region Every dome we invert starts at a finite doping, and below its smallest doping there are no data. Since Fh2=−2x2κ4F^2_h=-2x^2κ^4 grows in magnitude with doping, the missing dopings are exactly the ones whose Fh2F^2_h would have fallen in (Fmax2,0](F^2_ ,0], and F2=0F^2=0 is the end point of that interval. Finer sampling within the measured range does not change this. The residual depends on M in that interval through only a few combinations of its values, for a reason other than the missing data. No Fh2F^2_h falls in it, so M there never enters the horizon condition Eq. (8) and affects the residual only through the bulk, where the kernel of Eq. (1) weights a change of M at radius z by ψHψB/z4 _H _B/z^4. When the condensation mode exists, ψH _H is proportional to ψB _B, which falls as zΔz at the boundary, so the weight falls as z2Δ−4z^2 -4 toward z→0z→ 0, the boundary, where F2=−2Q2z4F^2=-2Q^2z^4 vanishes on every background. Per unit F2F^2, the variable of the ansatz, the weight does the opposite, since dz∝|F2|−3/4dF2dz |F^2|^-3/4\,dF^2 turns z2Δ−4dz^2 -4dz into |F2|(2Δ−7)/4dF2|F^2|^(2 -7)/4\,dF^2, which grows toward F2=0F^2=0 with the same power on every background, so near F2=0F^2=0 the points of the dome differ mainly in a prefactor and fix only a few combinations of M. We call the interval (Fmax2,0](F^2_ ,0] the blind region, a name justified at the accuracy of the reconstructions of Sec. 5, which measure how weakly it is constrained. Weak sensitivity here does not mean that M(0)M(0) has no effect. The same number enters Eq. (5) and sets the exponent Δ with which the integration from the boundary starts, so changing M(0)M(0) changes the whole dome, and the rest of M has to be retrained to restore it. Retraining succeeds, because a change at F2=0F^2=0 is absorbed by coefficients in the blind region. The dome does not determine M(0)M(0). We therefore scan it, holding the scalar mass at each pinned mass m0m_0 and retraining everything else, and the family of reconstructions is the object of study. Holding M(0)M(0) exactly, rather than penalizing its distance from a value, is what makes that family well defined, and Sec. 4.1 builds the identity M(0)=m0M(0)=m_0 into the ansatz. Without a pinned mass, the scalar mass that training returns is set by something other than the data. In the Gaussian basis of Ref. Kim et al. (2025), which does not fix M(0)M(0) and in which every basis function is nonzero at F2=0F^2=0, training on the Gaussian target lowers M(0)M(0) until a one-sided penalty bounds it at −2.0001-2.0001. Without the penalty it reaches −2.16-2.16, and on a measured target the AdS4 Breitenlohner–Freedman bound −9/4-9/4 Breitenlohner and Freedman (1982), below which Δ ceases to be real, while training with 100100 narrow functions and no penalty returns −0.11-0.11, close to its starting value. Equation (5) turns the scan into a scan of the dimension of the dual operator, Δ=32+94+m0 = 32+ 94+m_0, which gives each pinned mass a meaning independent of the training. The three pinned masses used throughout, m0=−1/4m_0=-1/4, −1/2-1/2 and −3/4-3/4, correspond to Δ=2.914 =2.914, 2.8232.823 and 2.7252.725. What bounds the scan below is a statement about zero doping, which training never evaluates. As x→0x→ 0 the charge Q=xκ2Q=xκ^2 vanishes, and with it AtA_t and F2F^2 at every z, so Eq. (6) reduces to a neutral scalar with the single constant mass m0m_0. The critical temperature at x≪1x 1 is therefore a function of m0m_0 alone, up to corrections that vanish with x, and we write it τc(0) _c(0). Since τc(0) _c(0) cannot separate one target from another, it is a property of m0m_0 and not a test of the training. The near-horizon geometry determines where τc(0) _c(0) vanishes. At strictly zero doping the lowest available temperature is the extremal one, κ2=6κ^2=6, where h→3(1−z)2h→ 3(1-z)^2, the near-horizon geometry is AdS2×ℝ2_2×R^2 with radius L22=1/3L_2^2=1/3, and substituting ψ∼(1−z)αψ (1-z)^α in Eq. (6) gives 3α(α+1)=m0.3α(α+1)=m_0. (9) Above the AdS2 stability bound m0L22≥−1/4m_0L_2^2≥-1/4 Denef and Hartnoll (2009); Faulkner et al. (2011), that is for m0≥−3/4m_0≥-3/4 or Δ≥2.725 ≥ 2.725, the exponent α in Eq. (9) is real, no condensation mode exists as x→0x→ 0, and τc(0)=0 _c(0)=0. Below the bound α is complex and τc(0) _c(0) is not zero. It rises continuously and by orders of magnitude as m0m_0 falls, with the essential singularity expected where an AdS2 bound is crossed Jensen et al. (2010); Iqbal et al. (2015), τc(0)≃τ∗exp(−π/ν) _c(0) _* (-π/ν) with ν2=−14−13m0ν^2=- 14- 13m_0. Figure 1 shows τc(0) _c(0) against m0m_0. A critical temperature at zero doping is therefore absent only for m0≥−3/4m_0≥-3/4. We take the window −3/4≤m0<0-3/4≤ m_0<0, with the pinned masses −3/4-3/4, −1/2-1/2 and −1/4-1/4, so that no reconstruction has a critical temperature at zero doping. The upper end keeps Δ below the marginal value 33, which m0=0m_0=0 reaches, so that the order parameter stays a relevant operator. The dome of Sec. 5.1, the one whose generating M is known, has M(0)=−1/2M(0)=-1/2, the middle pinned mass. Figure 1: The critical temperature at zero doping, τc(0) _c(0), against the scalar mass m0m_0, with Δ on the top axis. Points are the calculation at x=10−3x=10^-3 with M held constant at m0m_0, which ends at m0=−0.9m_0=-0.9 where τc(0) _c(0) is already below 10−610^-6, and the dashed line is τ∗exp(−π/ν) _* (-π/ν) with τ∗=0.55 _*=0.55 fixed by the point at m0=−0.92m_0=-0.92, continued to the AdS2 bound m0=−3/4m_0=-3/4, above which τc(0) _c(0) vanishes. It also agrees with the points at m0=−1m_0=-1 and −0.9-0.9, which do not enter τ∗ _*. The shaded window −3/4≤m0<0-3/4≤ m_0<0 holds the three pinned masses (triangles), and the gray strip is the range of the dome peaks. 4 PINN with a Kolmogorov–Arnold network 4.1 The network The earlier inversion Kim et al. (2025) is a PINN in which M is expanded in Gaussian functions, the residual of each point of the dome is the source J, and the governing equation is imposed by integrating it rather than at collocation points, the sampled points at which a PINN ordinarily penalizes the residual of the equation. Ours keeps that structure and replaces two of its parts, the residual by the Wronskian of each point and the Gaussian expansion by a Kolmogorov–Arnold network (KAN) Liu et al. (2025), a network whose trainable functions are expansions in B-splines, which makes the model a physics-informed KAN Shukla et al. (2024). The forward pass, the evaluation of the output from the inputs, is two integrations of Eq. (6), the branches, one from the horizon with Eq. (8) and one from the boundary with ψ∼zΔψ z , which meet at an interior point. Each is a continuous-depth network whose depth coordinate is z and whose hidden state is the pair (ψ,ψ′)(ψ,ψ ), of width two because the equation is second order. It is a residual network, whose layers each add a correction to their input, with all layers drawing their weights from the same parameters and the step chosen by the integrator. The background (κi,Qi)( _i,Q_i) of the i-th point of the dome is a fixed input, determined in Sec. 2 and never trained. The output is produced by a head with no parameters, the final layer that maps the state to the output. Each branch is fixed only up to an overall scale, since Eq. (6) is linear and homogeneous, so a difference of the two would change under a rescaling of either. What is invariant is whether the phase-space vectors (ψH,ψH′)( _H, _H ) and (ψB,ψB′)( _B, _B ) are parallel, which they are exactly when a normalizable mode regular at the horizon exists. At the matching point zmidz_ mid the head forms W^=ψHψB′−ψH′ψB‖(ψH,ψH′)‖‖(ψB,ψB′)‖|zmid∈[−1,1], W= _H _B - _H _B \|( _H, _H ) \|\, \|( _B, _B ) \| |_z_ mid∈[-1,1], (10) the sine of the angle between the two terminal states. The normalization lets the points of a dome be compared although their raw amplitudes differ by many orders of magnitude, and since W is independent of z on shell, zmid=0.5z_ mid=0.5 is chosen only because it splits the work evenly. The ansatz for M is the smallest KAN, a single edge from F2F^2 to M Liu et al. (2025), whose trainable function is one linear layer acting on fixed features, M(F2)=m0+∑j=1100θj[Bj(F2)−Bj(0)],M(F^2)=m_0+ _j=1^100 _j [B_j(F^2)-B_j(0) ], (11) where the features BjB_j are cubic B-splines, piecewise cubics with two continuous derivatives, on knots uniform over [Fmin2,0][F^2_ ,0]. The features are not learned. The only trained quantities are the weights θ of that layer, and because they enter M linearly the residual stays close to linear in the parameters over a step, which is why Levenberg–Marquardt applies at all. Here, m0m_0 is the bias and the pinned mass. The subtraction makes M(0)=m0M(0)=m_0 an identity in the coefficients rather than a constraint on them, so no combination of the θj _j changes the scalar mass, and m0m_0 is carried as the first entry of the parameter vector, held fixed by a box of half-width 10−910^-9, one of 101101 parameters. The features have to be localized, and the reason is the ordering of Sec. 2. Each feature Bj−Bj(0)B_j-B_j(0) is nonzero on at most four knot intervals, with one exception, the last, B100−1B_100-1, which vanishes at F2=0F^2=0 and equals −1-1 below the final knot interval, so it shifts M on the whole horizon range relative to the pinned mass and is the column through which a change at F2=0F^2=0 affects the horizon range. Since a point of the dome probes M on [Fh2,0][F^2_h,0] and |Fh2||F^2_h| grows with doping, a feature placed at some F2F^2 affects only the points whose Fh2F^2_h is more negative than that value. The points at higher doping can therefore be matched by features the points at lower doping do not depend on, so matching them leaves the fit at lower doping unchanged. An ordinary neural network in place of Eq. (11) breaks this property. Its features, the hidden units, are affine functions of F2F^2 passed through a pointwise nonlinearity such as tanh , and each varies over the whole range of F2F^2. A change made to match a high doping therefore changes M everywhere, including near F2=0F^2=0 where the points at lower doping constrain it. The Green’s identity itself holds for any parameterization, so with global features this property is lost while the derivative is unaffected. The ansatz enters the branches in three places, and two of them are easy to miss. It sets the coefficient M(F2(z))/z4M(F^2(z))/z^4 that multiplies the state at depth z in Eq. (6), which is the layer weight. It fixes the horizon branch’s input through M(Fh2)/4πTM(F^2_h)/4π T in Eq. (8). And the boundary branch starts from zΔz with Δ=Δ(m0) = (m_0), so its input depends on the bias alone. The pinned mass of Sec. 3 and the input of one branch are the same quantity, a second reason to hold it exactly. 4.2 Training In deep learning the backward pass, the computation of the gradient of the output with respect to every parameter, has always been the expensive part, in memory more than in arithmetic, and a research subject of its own. Backpropagation Rumelhart et al. (1986), the reverse mode of automatic differentiation Baydin et al. (2018), gives the full gradient at a small multiple of the cost of the forward pass but has to retain or recompute every intermediate state, which has led to schemes that store a few states and recompute the rest Chen et al. (2016b). For a continuous-depth network the adjoint method Chen et al. (2018) avoids the storage by integrating an adjoint equation backward in depth, but requires a second integration whose accuracy has to be controlled. Here the backward pass needs no integration of its own. In network terms, Eq. (1) says that the gradient with respect to the weight at depth z, the forward activation there times the adjoint, the sensitivity of the output to the state at z, is ψH(z)ψB(z) _H(z) _B(z). Equation (6) is its own adjoint, so the adjoint of the boundary branch solves the same equation with the horizon branch’s values at the matching point as terminal data, which makes it the horizon branch itself continued past zmidz_ mid, and conversely. The adjoint of one branch is the forward solution of the other, and there is no separate backward integration because nothing is left to compute. Two qualifications apply. The first is that only M is trained, entering Eq. (6) as the mass of the scalar, while the branches are exact integrations of that equation for the current M and have no weights of their own. The Green’s identity holds because the branches solve that equation exactly and would fail for trained approximations of them, so we do not call the branches neural networks and use network terms for them only to compare their structure with the earlier PINN. The second concerns cost. The residual needs each branch only as far as zmidz_ mid, but the integral in Eq. (1) runs over the whole bulk, so a Jacobian evaluation doubles the integration. The gain is that the observable is obtained without a fit and that the Jacobian is available at any number of parameters, not that the forward pass is shorter. Because the ansatz is linear in θ, the Jacobian factorizes. With the product ψHψB _H _B computed once for the i-th point, the derivative ∂Wi∂θj=∫01ψHψB[Bj(F2(z))−Bj(0)]dz4 ∂ W_i∂ _j= _0^1 _H _B\, [B_j (F^2(z) )-B_j(0) ]\, dzz^4 (12) is one contraction of a design matrix, the features evaluated on the integration grid and independent of θ, against a per-point kernel. Table 1 compares the cost of one update in the two methods. The earlier one integrates once per point for each evaluation of its objective and, having no backward pass, needs np+1n_p+1 evaluations to finite-difference a gradient, so its cost grows with the parameter count, while ours needs two integrations per point at any count, since the residual and the whole Jacobian are computed from the same pair of solutions. parameters method integrations seconds 2525 earlier 728728 12.612.6 this 5656 1.271.27 5050 earlier 1,4281,428 25.625.6 this 5656 1.331.33 101101 earlier 2,8562,856 53.953.9 this 5656 1.351.35 Table 1: One update in each method, on the phase boundary of Sec. 5.1 at 2828 points, with the same basis and the same integration accuracy. The earlier method is one shooting solution per point, J from a fit of the near-boundary expansion, and a finite-difference gradient for Adam Kim et al. (2025). Ours is the two solutions, each integrated across the whole bulk, with W W evaluated at the matching point and the Jacobian of Eq. (1). The two methods also need different numbers of updates. From a constant M at 101101 parameters on the phase boundary of Table 1, Levenberg–Marquardt converges in nine updates and 3535 seconds, while Adam is run for about 10310^3 iterations Kim et al. (2025), which at the last row of the table is most of a day. Every result below uses 100100 points per dome, and there one reconstruction takes from about ten minutes to an hour, while 10310^3 Adam iterations at the cost per integration of Table 1 would take more than two days. That is what makes a scan of three pinned masses on five phase boundaries practical. The loss, minimized over the parameters θ, is ℒ(θ)=∑iW^i2+λ2‖D(2)θ‖2+ρ22‖Cθ‖2,L(θ)= _i W_i^2+λ^2 \|D^(2)θ \|^2+ρ^2N^2 \|Cθ \|^2, (13) with i running over the points of the dome. The first term is the misfit to the data. The second is a Tikhonov term, in which D(2)θD^(2)θ is the second difference of M on a uniform grid in F2F^2, (D(2)θ)a=M(ua+1)−2M(ua)+M(ua−1)(ua+1−ua)2, (D^(2)θ )_a= M(u_a+1)-2M(u_a)+M(u_a-1)(u_a+1-u_a)^2, (14) with uau_a, a=0,…,249a=0,…,249, evenly spaced over [Fmin2,0][F^2_ ,0]. Since M is linear in θ, D(2)D^(2) is a fixed matrix, and the term penalizes curvature everywhere. It fixes the directions the kernel leaves unconstrained. The third term acts on the blind region alone. That third term is the shape requirement. Since the dome constrains M only weakly in the blind region, we require of a reconstruction there not that it be flat but that it continue from the horizon range without additional structure, staying close to a parabola with a single curvature. It is imposed by a penalty rather than a constraint because making M exactly linear there increases the phase-boundary error by a factor of three to nine, so a nearly linear M fits the data and an exactly linear one does not. The matrix C takes the second divided difference of the coefficients in the blind region, each placed at the average of its three interior knots, (Cθ)j=1ξj+1−ξj−1(θj+1−θjξj+1−ξj−θj−θj−1ξj−ξj−1),ξj=13(tj+1+tj+2+tj+3),(Cθ)_j= 1 _j+1- _j-1 ( _j+1- _j _j+1- _j- _j- _j-1 _j- _j-1 ), _j= 13 (t_j+1+t_j+2+t_j+3 ), (15) for every three adjacent coefficients θj−1,θj,θj+1 _j-1, _j, _j+1 whose B-splines reach the blind region, with tjt_j and tj+4t_j+4 the end knots of BjB_j, so that Cθ=0Cθ=0 exactly when M is linear there and the penalty is minimized by a straight line, the zero-curvature limit of a parabola. On all four targets CθCθ matches M′/2M /2 there to three digits. The prefactor =ℓb2/(4δMbnb)N= _ b^2/(4\,δ M_ b n_ b), with ℓb _ b the width of the blind region, δMb=m0−M(Fmax2)δ M_ b=m_0-M(F^2_ ) its total rise and nbn_ b the number of rows of C, makes ρ the weight on the relative deviation from a straight line, since a curvature M′M displaces the midpoint of a segment of length ℓb _ b from the straight line by M′ℓb2/8M _ b^2/8. Without it the same ρ acts with strengths that differ by a factor of sixty across the four targets. On a given dome δMbδ M_ b grows with m0m_0, so the same ρ allows more absolute curvature at the less negative pinned masses. Both weights are set by shape and neither depends on the dome. Every reconstruction below minimizes Eq. (13) with the same optimizer and the same weights, λ=7×10−4λ=7× 10^-4 and ρ=1ρ=1, so the shape requirement is tuned to no reconstruction. The normalization needs M(Fmax2)<m0M(F^2_ )<m_0, since ∝1/δMbN 1/δ M_ b, and all fifteen reconstructions reported here satisfy it. 4.3 The rule Choosing a value inside the window needs a rule that the residual does not provide. The rule applies two criteria after the scan, and the first checks the shape requirement, which the loss imposes only as a penalty. For each reconstruction we compute its deviation from a parabola over the blind region, ϵ2=max|M−∑r=02βr(F2)r||m0−M(Fmax2)|,β=argmin∫Fmax20(M−∑r=02βr(F2)r)2dF2, _2= |M- _r=0^2 _r(F^2)^r ||m_0-M(F^2_ )|, β= _F^2_ ^0 (M- _r=0^2 _r(F^2)^r )^2dF^2, (16) with β=(β0,β1,β2)β=( _0, _1, _2) the coefficients of the least-squares parabola in F2F^2, and we keep those below a threshold of 5%5\%. Under the second criterion, closed forms are fitted to every reconstruction, to the curve the spline returned rather than to the phase boundary, and among those that pass the first criterion the pinned mass kept is the one whose reconstruction is reproduced by the closed form with the fewest symbolic parameters. The closed forms come from a library fixed in advance, M¯=m0+∑r=1nccr(F2)r+∑l=1nfdlΦ(sl,wl),Φ(s,w)=e−((F2−s)/w)2−e−(s/w)2, M=m_0+ _r=1^n_cc_r(F^2)^r+ _l=1^n_fd_l\, (s_l,w_l), (s,w)=e^-((F^2-s)/w)^2-e^-(s/w)^2, (17) a polynomial of degree ncn_c plus nfn_f Gaussian functions of amplitude dld_l, center sls_l and width wlw_l, with the subtraction in Φ keeping M¯(0)=m0 M(0)=m_0 identically. The degree runs over nc=1,2,3n_c=1,2,3 when nf≤1n_f≤ 1 and over nc=2,3n_c=2,3 when 2≤nf≤52≤ n_f≤ 5, and the number of symbolic parameters is k=nc+3nfk=n_c+3n_f, the pinned mass not being counted. Each point of the dome depends on M only between F2=0F^2=0 and its own horizon value Fh2F^2_h. The end of M on which only the last two points, the two highest dopings, depend is where the reconstruction is least reliable, so the closed form is fitted without it. Every point of the dome stays in the reconstruction. The coefficients and widths are bounded and neighboring Gaussian functions are kept apart.11 1 Amplitudes are bounded by |dl|≤20|d_l|≤ 20, widths by 0.02≤wl≤100.02≤ w_l≤ 10 and polynomial coefficients by |cr|≤100|c_r|≤ 100. The center of the first Gaussian function lies inside the comparison grid, and each further center lies at larger F2F^2 than the previous one, by 5%5\% to 100%100\% of the grid length, and may lie past F2=0F^2=0, where no background reaches. A closed form reproduces a reconstruction when its root-mean-square curve residual over the compared interval of F2F^2 is at most 0.6%0.6\% of the range of M on the whole curve, the accuracy of the reconstruction itself, and ties in k between pinned masses are broken by the mean error of the phase boundary each closed form regenerates. 5 Results 5.1 A phase boundary whose mass function is known Every target of Sec. 5.2 specifies a dome and leaves M unknown, so there the question can only be what the reconstruction returns and never whether it returns the M that generated the dome. Here we write one down, Mgen(F2)=−12+2.125F2+0.525(F2)2,M_ gen(F^2)=- 12+2.125\,F^2+0.525\,(F^2)^2, (18) solve Eq. (6) forward at 100100 values of x to obtain a phase boundary, and train on it by the same procedure, without using MgenM_ gen. The generating scalar mass is −1/2-1/2. Equation (18) is chosen to give a dome of about the right shape, not a mass function anything is known to have, so recovering it is a statement about the inversion and nothing else. Run at 101101 parameters, at the weights of Sec. 4.2 and on the three pinned masses the four targets use, the scan gives mean phase-boundary errors of 0.2590.259, 0.2470.247 and 0.254%0.254\% from m0=−1/4m_0=-1/4 down to −3/4-3/4, flat to 5%5\% over a range in which the scalar mass changes by 1/21/2. The smallest value happens to be at the generating scalar mass, but a difference of 5%5\% is a consequence of the weak sensitivity of Sec. 3, not a minimum. Without the shape requirement the smallest value is not even at the generating scalar mass. On this dome the two highest dopings are x=0.337x=0.337 and 0.3340.334, and the part left out of the comparison is F2<−4.719F^2<-4.719, the Fh2F^2_h of the third-highest doping, 1212 of the 250250 points on which the curve is compared. All three reconstructions pass the first criterion of Sec. 4.3, and the second selects the generating scalar mass and the generating quadratic. At m0=−1/2m_0=-1/2 that form is the pure quadratic, M¯known=−12+2.1191F2+0.5228(F2)2, M_ known=- 12+2.1191\,F^2+0.5228\,(F^2)^2, (19) against the 2.1252.125 and 0.5250.525 of Eq. (18). At the other two pinned masses the form with the fewest symbolic parameters that reproduces the reconstruction includes a Gaussian function, a line with one at m0=−3/4m_0=-3/4 and a quadratic with one at −1/4-1/4, so the count is two at the generating scalar mass against four and five elsewhere, with no tie. The rule prefers the reconstruction whose shape needs the fewest parameters, and on this dome that is the one held at the generating scalar mass, because a reconstruction held elsewhere has to bend across the blind region to reach its pinned mass, and representing that bend requires additional Gaussian functions. Figure 2: The 101101-parameter reconstruction of the phase boundary generated by Eq. (18), at each pinned mass m0m_0, against MgenM_ gen (dashed) and the closed form M¯known M_ known of Eq. (19) (orange dotted). Shading is the horizon range and the vertical dotted line Fmax2F^2_ . The end marked with diagonal lines, on which only the two highest dopings depend, is left out of the closed-form comparison. Left, the whole range. Right, the blind region alone. Figure 3: The phase boundaries regenerated by the reconstructions of Fig. 2 and by the closed form M¯known M_ known of Eq. (19), against the phase boundary of MgenM_ gen itself (dashed), on which the 100100 points of the dome lie. Left, on a linear scale. Right, below the smallest doping on a logarithmic scale, where the reconstructions part according to their pinned masses and every curve ends where τc _c falls below about 10−610^-6. Figure 2 shows which parts of M the data constrain. Over the horizon range, apart from the end marked with diagonal lines, the three reconstructions agree with MgenM_ gen and with each other. Over the blind region the curves are set by the pinned mass and the shape requirement. Only the curve held at −1/2-1/2 stays within 0.0230.023 of MgenM_ gen, and the other two deviate from it near Fmax2F^2_ and end at their own pinned masses. The phase boundaries they regenerate, in Fig. 3, lie on the phase boundary of MgenM_ gen over the range of the data and part only below the smallest doping, according to their pinned masses, while the closed form M¯known M_ known stays on the phase boundary of MgenM_ gen there too. A change of m0m_0 is absorbed by coefficients in the blind region, which the data constrain only weakly, while for the quadratic, whose values on the horizon range depend on its value at F2=0F^2=0, changing the pinned mass changes the curve where the data are, and the fit to the data becomes worse. In other words, identifiability, whether the data single out one M, is a property of the class of functions and not of the dome. One phase boundary and one set of pinned masses leave the answer undetermined at 101101 parameters and determine it at two, and Fig. 2 shows the region where the difference arises. On the two synthetic targets of Sec. 5.2 the scan is flat in the same way, while on the two measured ones the phase-boundary error grows toward the most negative pinned mass, by factors of 2.42.4 and 1.61.6. 5.2 Four reconstructions Four domes are inverted, and training uses the phase boundary and nothing else. Two are specified directly as domes in (x,τ)(x,τ), a Gaussian in x, which is the example of Ref. Kim et al. (2025), and a sum of two Gaussians, harder because the minimum between the peaks requires compensating structure at intermediate F2F^2. No M is known for either. Two are measured. The first is the critical temperature of YBCO against hole doping p Sato et al. (2017), the data used by the earlier method Kim et al. (2025), digitized from the published figure into 3434 points. The second is the carrier-density phase diagram of 2M-WS2 Zhao et al. (2021), which peaks at 8.68.6 K and whose abscissa is already a density. The axes of the bulk theory are dimensionless, so each measured phase diagram enters with two scale factors, x=axpx=a_xp and τ=aτTcτ=a_τT_c with p the measured abscissa, chosen once so that the data lie in the part of the (x,τ)(x,τ) plane that the background covers, the material-dependent ambiguity noted in Ref. Kim et al. (2025). They are ax=1.78a_x=1.78 and aτ=1.339×10−4a_τ=1.339× 10^-4 K-1 for YBCO, with p the hole doping, and ax=0.2a_x=0.2 and aτ=1.5×10−3a_τ=1.5× 10^-3 K-1 for 2M-WS2, with p the carrier density in units of 102110^21 cm-3. Both maps are linear through the origin, so any phase boundary below returns to measured units through p=x/axp=x/a_x and Tc=τ/aτT_c=τ/a_τ. Every target enters training as 100100 points, sampled from a smoothing spline on the measured ones. Figure 4: Top, the reconstructed M at each pinned mass m0m_0 and, in orange, the closed form the rule keeps, at the m0m_0 in the legend, with the horizon range shaded and the blind region right of the dotted line at Fmax2F^2_ . Bottom, the phase boundaries τc(x) _c(x) they regenerate over the target (gray points), with the range of x used in training shaded. Top legends give the mean phase-boundary error and ϵ2 _2 of Sec. 4.3, against its threshold of 5%5\%. Figure 4 is the scan, four targets by three pinned masses at the single pair of weights. Every trained reconstruction reproduces the phase boundary of its target. The peak is reproduced closely on three of the four, and on 2M-WS2 it comes out a few percent low, because the reconstruction smooths the sharp maximum at x=0.28x=0.28. The top row shows the three reconstructions of each target in full. The structure visible in such a plot, the positions and depths of the local minima of M and the plateau on 2M-WS2, is common to all three and is therefore a property of the dome. The value of a curve at F2=0F^2=0 is not, being its pinned mass, and neither is its shape across the blind region, which the shape requirement determines and which is therefore an inference rather than a reconstruction. The two criteria of Sec. 4.3 leave one pinned mass per target. Every one of the twelve reconstructions rises monotonically across the blind region. At the 5%5\% threshold on the deviation from a parabola, the two least negative pinned masses of 2M-WS2 fail the first criterion and every other reconstruction passes. The second criterion is applied now where the answer is not known, again without the end on which only the two highest dopings depend, between five and seven of the 250250 curve points on each target. The counts are not small. The Gaussian target needs eleven symbolic parameters at every pinned mass, the double Gaussian fourteen or fifteen, YBCO eleven at every pinned mass and 2M-WS2 fifteen to eighteen, with eighteen, the largest form in the library, on the one reconstruction that passes the first criterion. No pure polynomial passes on any of the twelve reconstructions. In three of the four forms kept a center offset is at its smallest allowed value, one in the double Gaussian and YBCO forms and two in the 2M-WS2 form, so the bound of the library, not the curve, fixes the closest pair of Gaussian functions there, and in the Gaussian form one amplitude is at its bound |dl|=20|d_l|=20. The Gaussian target ties at eleven on all three pinned masses, with regenerated errors of 2.2912.291, 2.2862.286 and 2.287%2.287\% from m0=−3/4m_0=-3/4 to −1/4-1/4, so the one kept, −1/2-1/2, is decided by differences below 0.01%0.01\% and is in effect undetermined. YBCO ties at eleven on all three pinned masses, where 0.662%0.662\% at −1/4-1/4 is lower than 0.7500.750 and 1.325%1.325\%, and the double Gaussian and 2M-WS2 have no tie. The kept closed forms regenerate the phase boundary less accurately than the reconstructions (Table 2), because matching a curve to 0.6%0.6\% of the range of M cannot give a phase boundary accurate to a tenth of a percent. Table 2 collects one value per target, the one the rule infers, with the dimension it implies. Appendix B writes the closed forms kept for the four targets out. phase boundary m0m_0 Δ (nc,nf)(n_c,n_f) k spline closed generated by Eq. (18) −1/2-1/2 2.8232.823 (2,0)(2,0) 22 0.247%0.247\% 0.538%0.538\% Gaussian −1/2-1/2 2.8232.823 (2,3)(2,3) 1111 0.172%0.172\% 2.286%2.286\% double Gaussian −3/4-3/4 2.7252.725 (2,4)(2,4) 1414 0.512%0.512\% 1.003%1.003\% YBCO −1/4-1/4 2.9142.914 (2,3)(2,3) 1111 0.306%0.306\% 0.662%0.662\% 2M-WS2 −3/4-3/4 2.7252.725 (3,5)(3,5) 1818 2.398%2.398\% 2.424%2.424\% Table 2: The pinned mass each phase boundary keeps under the two criteria, with the dimension it implies, under the conventions of Sec. 4.3. The form (nc,nf)(n_c,n_f) gives the polynomial degree and the number of Gaussian functions of Eq. (17), and k=nc+3nfk=n_c+3n_f its symbolic parameters. The last two columns are the mean error of the phase boundary regenerated from the 101101-parameter reconstruction and from the closed form. The first row is the test of Sec. 5.1. Eqs. (19) and (21) to (24) write the forms out. Further tests, made after the rule was fixed, show where it stops working. On domes generated from variants of Eq. (18), with M(0)M(0) between pinned masses or a Gaussian function added inside the blind region, it still finds the generating scalar mass, or the nearest one, together with the quadratic. But when the dome is generated from one of the closed forms of Appendix B at its own m0m_0, every pinned mass needs the same number of symbolic parameters for YBCO, the double Gaussian and the Gaussian, and the smallest count falls at the wrong one for 2M-WS2, so the generating scalar mass is found in two of the four cases, both through the tie-break. On each of the four targets, changing a threshold within a factor of two or another convention of the rule moves the kept pinned mass. 6 Discussion The value this paper reports for each target is the output of three choices, the window of Sec. 3 and the two criteria of the rule, each of which could be made differently. There is a window, the values of m0m_0 at which no critical temperature survives at zero doping, set by the stability bound of the near-horizon AdS2. There is a shape requirement on the blind region. And there is a preference for the reconstruction reproduced by the closed form with the fewest symbolic parameters. None of the three comes from the residual, which does not determine M(0)M(0) and constrains the rest of the blind region in only a few combinations. What the training of Sec. 4.2 contributes is that the rule can be applied at all, to fifteen reconstructions with a closed-form search for each. What would falsify the rule, as opposed to disagreeing with its output, is a case where the generating M is known and the rule does not return it. Section 5.1 is a case where the generating M is known, and the rule returns it on a dome made after the conventions of the rule had been fixed. Such cases also include the four domes at the end of Sec. 5.2, whose generating M are the closed forms of Appendix B, and there the rule returns the generating scalar mass only twice, each time through the tie-break. On the four targets of Sec. 5.2, however, the pinned masses are decided by ties, by a single parameter or by the threshold of the first criterion, so the values kept are not evidence for their scalar masses. A second observable could take the place of the rule. The dome is one function of doping and leaves M(0)M(0) open, but observables of the condensed phase, or of pair fluctuations above TcT_c, depend on the dimension Δ of the dual operator, and the pair susceptibility has been proposed as a measurement of that dimension in quantum critical metals She et al. (2011). Computing one of them from the reconstructions at the different pinned masses would show how far apart they are in measurable terms. There are three things this paper does not establish. It does not measure any material’s scalar mass, since both measured phase diagrams enter with two scale factors chosen by hand and the number reported moves when they change. It does not claim that the mass functions displayed are unique, since the reconstructions at different m0m_0 agree to a few times 10−210^-2 on the horizon range and differ by the full spacing of the pinned masses only in the blind region. And it does not claim to generalize beyond the RN black hole with axions. The construction relies on |Fh2|=2Q2|F^2_h|=2Q^2 growing along every dome, which is why localized features are the right ansatz in Sec. 4.1, and a background on which |Fh2||F^2_h| does not grow monotonically along a dome would require that ansatz to be reconsidered. Two assumptions about the condensed phase are not tested here. Every TcT_c above is the onset of the condensation mode, which is the transition temperature only if the transition is continuous, and couplings of the order parameter to the gauge field can make it first order Franco et al. (2010), with the transition temperature then above the onset. Below TcT_c the coupling feeds back into the Maxwell equation through the factor 1+4M′(F2)|ψ|21+4M (F^2)|ψ|^2 of Eq. (3), which has to stay positive for the gauge field to keep a positive kinetic term, and that bounds the condensate where M′<0M <0, at more negative F2F^2 than each minimum of every reconstruction. Both questions need the hairy black holes below TcT_c, which this paper does not construct. Switching the shape requirement off shows what it removes. With ρ=0ρ=0 a measured reconstruction develops a shallow local minimum between Fmax2F^2_ and F2=0F^2=0, and removing it is what raises the phase-boundary error at the common weights. Such a minimum requires a sufficiently wide blind region, and the geometry determines its width. Its share of the range of F2F^2 is (Qmin/Qmax)2(Q_ /Q_ )^2, so a narrower doping range gives a wider blind region. The measured doping ranges are the narrower ones, but a narrow range alone does not produce the minimum, since the two Gaussian domes restricted to the measured ranges develop it weakly or not at all, and which feature of a measured dome produces it we leave open. In the language of inverse problems the method is a least-squares fit constrained by an ordinary differential equation. The condition for a condensation mode is a Sturm–Liouville problem, as in analytic estimates of TcT_c Siopsis and Therrien (2010), its derivative Eq. (1) is first-order perturbation theory for a self-adjoint operator, known elsewhere as the adjoint-state method Plessix (2006), and which parts of M the dome determines is the question of inverse Sturm–Liouville theory Borg (1946) and of resolution analysis Backus and Gilbert (1968), to which the blind region is the answer here. Table 1 and the update counts after it separate the two factors of the speed-up, the analytic Jacobian, which makes an update about np/2n_p/2 times cheaper in integrations than a finite-difference one, and Levenberg–Marquardt, which needs about a hundred times fewer updates than Adam, so a finite-difference Jacobian with Levenberg–Marquardt would keep the second factor and lose the first. What the dome constrains, then, is the mass function where its horizons reach, with the same minima at every pinned mass, and not the scalar mass, which the window of Sec. 3 bounds and a second observable would have to fix. Acknowledgements.The author is supported by a KIAS Individual Grant (AP103401) via the Center for AI and Natural Sciences at the Korea Institute for Advanced Study. The code for the method and for the computations reported here, and the drafts of this manuscript, were written with the assistance of Claude (Anthropic). The code and the data behind every figure are available from the author on request. The formulation of the problem, the design of the tests and the interpretation of the results are the author’s own, and the author takes full responsibility for the content. Appendix A Numerical parameters Integration uses DOP853, the explicit Runge–Kutta method of order eight of Dormand and Prince, with relative and absolute accuracy 10−1110^-11 and 10−1310^-13 during training, and 10−910^-9 and 10−1210^-12 when a phase boundary is regenerated from a reconstructed M by root finding. The boundary branch starts at z0=5×10−4z_0=5× 10^-4 and the horizon branch at a distance 10−610^-6 from the horizon, which the leading-order near-horizon expansion used there needs to be small compared with 4πT/|M(Fh2)|4π T/|M(F^2_h)|, a condition met with a wide margin at τc≥10−4 _c≥ 10^-4. The matching point is zmid=0.5z_ mid=0.5, and the integral of Eq. (1) is evaluated on 12001200 points of a grid denser at both ends. The coefficients are bounded by 55 in absolute value, the box within which the trust-region reflective solver takes its steps Branch et al. (1999), a Levenberg–Marquardt variant that reflects steps off the bounds and solves its subproblem by singular value decomposition with column scaling taken from the Jacobian. Every reconstruction minimizes Eq. (13) with this solver. The Jacobian of Eq. (10) keeps the kernel and drops the derivative of the normalization, which is exact at the solution because that term multiplies W W itself. Against a central finite difference its columns agree to 3.4×10−33.4× 10^-3 of the largest entry on the phase boundary of Sec. 5.1 at ρ=0ρ=0 and λ=1.4×10−4λ=1.4× 10^-4, where the median |W^|| W| is 1.9×10−81.9× 10^-8, the remaining difference coming from the numerical integration of the kernel. At the common weights W W does not vanish and the dropped term adds to this, so the agreement is 2.4×10−22.4× 10^-2 on the same phase boundary and 6.9×10−26.9× 10^-2 on the Gaussian target at its selected pinned mass, at median |W^|| W| of 9.9×10−69.9× 10^-6 and 3.6×10−53.6× 10^-5. The solver therefore stops at a stationary point of an objective that differs from Eq. (13) by terms proportional to W W, which the same two solutions could remove, since the derivative of the normalization follows from them by variation of parameters. The misfit is W W and not the relative error in τc _c, which it becomes to first order after division by τi∂τW^i _i _τ W_i, so the points are weighted by the sensitivity of W W to τ, and the weights λ and ρ are set by shape rather than by a discrepancy principle in τc _c. Every free coefficient leaves M(0)M(0) fixed, ∂M(0)∂θj=0, ∂ M(0)∂ _j=0, (20) so the boundary term that a variation of Δ contributes appears only in the derivative with respect to the pinned mass, which is held fixed. Because a Jacobian evaluation integrates both branches across the whole bulk, the near-extremal backgrounds at the lowest dopings occasionally fail to integrate, five of a hundred on the Gaussian target, all at κ/κmax>0.998κ/ _ >0.998, where κmax(x) _ (x) is the largest axion scale at which a horizon still exists. Such a point is given a large residual and a zero Jacobian row, so that the trust region shrinks instead of the solver stopping with an error. Appendix B The closed forms kept Written out with the coefficients rounded to four significant figures, the closed forms kept for the four targets are M¯Gauss M_ Gauss =−12−1.720F2−0.1329(F2)2 =- 12-1.720\,F^2-0.1329\,(F^2)^2 −0.6817Φ(−2.058,0.5540)+3.856Φ(−0.1112,1.373) -0.6817\, (-2.058,0.5540)+3.856\, (-0.1112,1.373) +20.00Φ(1.174,0.7243), +20.00\, (1.174,0.7243), (21) M¯2Gauss M_ 2Gauss =−34+1.610F2+0.07314(F2)2 =- 34+1.610\,F^2+0.07314\,(F^2)^2 −0.7547Φ(−3.929,0.6025)−4.727Φ(−3.441,2.422) -0.7547\, (-3.929,0.6025)-4.727\, (-3.441,2.422) −8.193Φ(−0.4487,1.739)−0.7738Φ(−0.04944,0.2911), -8.193\, (-0.4487,1.739)-0.7738\, (-0.04944,0.2911), (22) M¯YBCO M_ YBCO =−14+0.9616F2+0.1558(F2)2 =- 14+0.9616\,F^2+0.1558\,(F^2)^2 −0.8962Φ(−2.077,1.525)−0.7821Φ(−0.9308,0.4321) -0.8962\, (-2.077,1.525)-0.7821\, (-0.9308,0.4321) −0.5790Φ(−0.6120,0.2209), -0.5790\, (-0.6120,0.2209), (23) M¯WS2 M_ WS_2 =−34+0.4336F2+0.4112(F2)2+0.07788(F2)3 =- 34+0.4336\,F^2+0.4112\,(F^2)^2+0.07788\,(F^2)^3 −1.748Φ(−3.092,0.3804)−1.438Φ(−2.411,0.7924) -1.748\, (-3.092,0.3804)-1.438\, (-2.411,0.7924) −0.9339Φ(−1.358,0.6616)−0.1731Φ(−1.143,0.1802) -0.9339\, (-1.358,0.6616)-0.1731\, (-1.143,0.1802) −0.3447Φ(−0.9286,0.09334), -0.3447\, (-0.9286,0.09334), (24) with Φ the Gaussian function of Eq. (17), while the test of Sec. 5.1 keeps Eq. (19). Every amplitude dld_l in them is negative except the last two of the Gaussian form, and the last of those is at the bound dl=20d_l=20 with its center at F2=1.174F^2=1.174, where no background reaches since F2≤0F^2≤ 0 on all of them. References Ahn et al. (2024) B. Ahn, H. Jeong, K. Kim, and K. Yun Deep learning bulk spacetime from boundary optical conductivity. JHEP 03, p. 141. External Links: 2401.00939, Document Cited by: §1. Akutagawa et al. (2020) T. Akutagawa, K. Hashimoto, and T. Sumimoto Deep Learning and AdS/QCD. Phys. Rev. D 102 (2), p. 026020. External Links: 2005.02636, Document Cited by: §1. Andrade and Gentle (2015) T. Andrade and S. A. Gentle Relaxed superconductors. JHEP 06, p. 140. External Links: 1412.6521, Document Cited by: §2. Andrade and Withers (2014) T. Andrade and B. Withers A simple holographic model of momentum relaxation. JHEP 05, p. 101. External Links: 1311.5157, Document Cited by: §2, §2, §2. Backus and Gilbert (1968) G. Backus and F. Gilbert The resolving power of gross Earth data. Geophys. J. R. Astron. Soc. 16, p. 169–205. External Links: Document Cited by: §6. Baggioli and Goykhman (2015) M. Baggioli and M. Goykhman Phases of holographic superconductors with broken translational symmetry. JHEP 07, p. 035. External Links: 1504.05561, Document Cited by: §2. Baggioli and Goykhman (2016) M. Baggioli and M. Goykhman Under The Dome: Doped holographic superconductors with broken translational symmetry. JHEP 01, p. 011. External Links: 1510.06363, Document Cited by: §1. Baydin et al. (2018) A. G. Baydin, B. A. Pearlmutter, A. A. Radul, and J. M. Siskind Automatic differentiation in machine learning: a survey. J. Mach. Learn. Res. 18 (153), p. 1–43. External Links: 1502.05767 Cited by: §4.2. Bea et al. (2024) Y. Bea, R. Jimenez, D. Mateos, S. Liu, P. Protopapas, P. Tarancón-Álvarez, and P. Tejerina-Pérez Gravitational duals from equations of state. JHEP 07, p. 087. External Links: 2403.14763, Document Cited by: §1. Bilson (2008) S. Bilson Extracting spacetimes using the AdS/CFT conjecture. JHEP 08, p. 073. External Links: 0807.3695, Document Cited by: §1. Borg (1946) G. Borg Eine Umkehrung der Sturm–Liouvilleschen Eigenwertaufgabe. Acta Math. 78, p. 1–96. External Links: Document Cited by: §6. Branch et al. (1999) M. A. Branch, T. F. Coleman, and Y. Li A Subspace, Interior, and Conjugate Gradient Method for Large-Scale Bound-Constrained Minimization Problems. SIAM J. Sci. Comput. 21 (1), p. 1–23. External Links: Document Cited by: Appendix A. Breitenlohner and Freedman (1982) P. Breitenlohner and D. Z. Freedman Stability in Gauged Extended Supergravity. Annals Phys. 144, p. 249. External Links: Document Cited by: §3. Cai et al. (2015) R. Cai, L. Li, L. Li, and R. Yang Introduction to Holographic Superconductor Models. Sci. China Phys. Mech. Astron. 58 (6), p. 060401. External Links: 1502.00437, Document Cited by: §1. Cai and Sin (2021) W. Cai and S. Sin The superconducting dome for holographic doped Mott insulator with hyperscaling violation. Eur. Phys. J. C 81 (6), p. 565. External Links: 2009.00381, Document Cited by: §1. Chen et al. (2016a) J. Chen, S. Dai, D. Maity, and Y. Zhang Engineering holographic phase diagrams. Phys. Rev. D 94 (8), p. 086004. External Links: 1603.08259, Document Cited by: §1. Chen et al. (2018) R. T. Q. Chen, Y. Rubanova, J. Bettencourt, and D. Duvenaud Neural Ordinary Differential Equations. Adv. Neural Inf. Process. Syst. 31. External Links: 1806.07366 Cited by: §4.2. Chen et al. (2016b) T. Chen, B. Xu, C. Zhang, and C. Guestrin Training Deep Nets with Sublinear Memory Cost. External Links: 1604.06174, Document Cited by: §4.2. Denef and Hartnoll (2009) F. Denef and S. A. Hartnoll Landscape of superconducting membranes. Phys. Rev. D 79, p. 126008. External Links: 0901.1160, Document Cited by: §3. Engelhardt and Horowitz (2017) N. Engelhardt and G. T. Horowitz Recovering the spacetime metric from a holographic dual. Adv. Theor. Math. Phys. 21, p. 1635–1653. External Links: 1612.00391, Document Cited by: §1. Faulkner et al. (2011) T. Faulkner, H. Liu, J. McGreevy, and D. Vegh Emergent quantum criticality, Fermi surfaces, and AdS2_2. Phys. Rev. D 83, p. 125002. External Links: 0907.2694, Document Cited by: §3. Filev (2026) V. G. Filev Holographic entanglement entropy, Wilson loops, and neural networks. JHEP 09, p. 073. External Links: 2604.05970, Document Cited by: §1. Franco et al. (2010) S. Franco, A. García-García, and D. Rodríguez-Gómez A General class of holographic superconductors. JHEP 04, p. 092. External Links: 0906.1214, Document Cited by: §6. Ganguli et al. (2013) S. Ganguli, J. A. Hutasoit, and G. Siopsis Superconducting Dome from Holography. Phys. Rev. D 87 (12), p. 126003. External Links: 1302.5426, Document Cited by: §1. Gubser et al. (1998) S. S. Gubser, I. R. Klebanov, and A. M. Polyakov Gauge theory correlators from noncritical string theory. Phys. Lett. B 428, p. 105–114. External Links: hep-th/9802109, Document Cited by: §1. Gubser (2008) S. S. Gubser Breaking an Abelian gauge symmetry near a black hole horizon. Phys. Rev. D 78, p. 065034. External Links: 0801.2977, Document Cited by: §1. Hammersley (2006) J. Hammersley Extracting the bulk metric from boundary information in asymptotically AdS spacetimes. JHEP 12, p. 047. External Links: hep-th/0609202, Document Cited by: §1. Hartnoll et al. (2008a) S. A. Hartnoll, C. P. Herzog, and G. T. Horowitz Building a Holographic Superconductor. Phys. Rev. Lett. 101, p. 031601. External Links: 0803.3295, Document Cited by: §1. Hartnoll et al. (2008b) S. A. Hartnoll, C. P. Herzog, and G. T. Horowitz Holographic Superconductors. JHEP 12, p. 015. External Links: 0810.1563, Document Cited by: §1. Hartnoll (2009) S. A. Hartnoll Lectures on holographic methods for condensed matter physics. Class. Quant. Grav. 26, p. 224002. External Links: 0903.3246, Document Cited by: §1. Hashimoto et al. (2021) K. Hashimoto, H. Hu, and Y. You Neural ordinary differential equation and holographic quantum chromodynamics. Mach. Learn. Sci. Tech. 2 (3), p. 035011. External Links: 2006.00712, Document Cited by: §1. Hashimoto et al. (2026) K. Hashimoto, H. Jeong, K. Kim, D. Takeda, and K. Yun Deep learning emergent spacetime from fermionic spectral functions in holography. External Links: 2609.18566 Cited by: §1. Hashimoto et al. (2018a) K. Hashimoto, S. Sugishita, A. Tanaka, and A. Tomiya Deep Learning and Holographic QCD. Phys. Rev. D 98 (10), p. 106014. External Links: 1809.10536, Document Cited by: §1. Hashimoto et al. (2018b) K. Hashimoto, S. Sugishita, A. Tanaka, and A. Tomiya Deep learning and the AdS/CFT correspondence. Phys. Rev. D 98 (4), p. 046019. External Links: 1802.08313, Document Cited by: §1. Hashimoto (2019) K. Hashimoto AdS/CFT correspondence as a deep Boltzmann machine. Phys. Rev. D 99 (10), p. 106017. External Links: 1903.04951, Document Cited by: §1. Herzog (2009) C. P. Herzog Lectures on Holographic Superfluidity and Superconductivity. J. Phys. A 42, p. 343001. External Links: 0904.1975, Document Cited by: §1. Horowitz (2011) G. T. Horowitz Introduction to Holographic Superconductors. Lect. Notes Phys. 828, p. 313–347. External Links: 1002.1722, Document Cited by: §1. Hubeny (2012) V. E. Hubeny Extremal surfaces as bulk probes in AdS/CFT. JHEP 07, p. 093. External Links: 1203.1044, Document Cited by: §1. Iqbal et al. (2015) N. Iqbal, H. Liu, and M. Mezei Quantum phase transitions in semilocal quantum liquids. Phys. Rev. D 91 (2), p. 025024. External Links: 1108.0425, Document Cited by: §3. Jensen et al. (2010) K. Jensen, A. Karch, D. T. Son, and E. G. Thompson Holographic Berezinskii-Kosterlitz-Thouless Transitions. Phys. Rev. Lett. 105, p. 041601. External Links: 1002.3159, Document Cited by: §3. Jeong et al. (2025) H. Jeong, H. Kim, K. Kim, G. Yun, H. Yu, and K. Yun AdS/Deep-Learning made easy I: neural network-based approaches to holography and inverse problems. External Links: 2511.22522 Cited by: §1. Jimenez et al. (2026) R. Jimenez, D. Mateos, P. Protopapas, P. Solé-Vilaró, P. Tarancón-Álvarez, and P. Tejerina-Pérez Gravitational Duals from Equations of State I: Large Hierarchies and False Vacua. External Links: 2606.30117 Cited by: §1. Karniadakis et al. (2021) G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, and L. Yang Physics-informed machine learning. Nature Rev. Phys. 3 (6), p. 422–440. External Links: Document Cited by: §1. Keimer et al. (2015) B. Keimer, S. A. Kivelson, M. R. Norman, S. Uchida, and J. Zaanen From quantum matter to high-temperature superconductivity in copper oxides. Nature 518 (7538), p. 179–186. External Links: Document Cited by: §1. Kim et al. (2015) K. Kim, K. K. Kim, and M. Park A Simple Holographic Superconductor with Momentum Relaxation. JHEP 04, p. 152. External Links: 1501.00446, Document Cited by: §2. Kim et al. (2025) S. Kim, K. K. Kim, and Y. Seo Phase diagram from nonlinear interaction between superconducting order and density: toward data-based holographic superconductor. JHEP 02, p. 077. External Links: 2410.06523, Document Cited by: §1, §1, §2, §3, §4.1, §4.2, Table 1, §5.2, §5.2. Kim (2026) S. Kim Learning the inverse Ryu-Takayanagi formula with transformers. JHEP 04, p. 128. External Links: 2511.06387, Document Cited by: §1. Kingma and Ba (2015) D. P. Kingma and J. Ba Adam: A Method for Stochastic Optimization. In International Conference on Learning Representations, External Links: 1412.6980 Cited by: §1. Kiritsis and Li (2016) E. Kiritsis and L. Li Holographic Competition of Phases and Superconductivity. JHEP 01, p. 147. External Links: 1510.00020, Document Cited by: §1. Levenberg (1944) K. Levenberg A method for the solution of certain non-linear problems in least squares. Quart. Appl. Math. 2, p. 164–168. External Links: Document Cited by: §1. Liu et al. (2025) Z. Liu, Y. Wang, S. Vaidya, F. Ruehle, J. Halverson, M. Soljačić, T. Y. Hou, and M. Tegmark KAN: Kolmogorov-Arnold Networks. In International Conference on Learning Representations, External Links: 2404.19756 Cited by: §4.1, §4.1. Maldacena (1998) J. M. Maldacena The Large N limit of superconformal field theories and supergravity. Adv. Theor. Math. Phys. 2, p. 231–252. External Links: hep-th/9711200, Document Cited by: §1. Marquardt (1963) D. W. Marquardt An algorithm for least-squares estimation of nonlinear parameters. J. Soc. Indust. Appl. Math. 11, p. 431–441. External Links: Document Cited by: §1. Park et al. (2022) C. Park, C. Hwang, K. Cho, and S. Kim Dual geometry of entanglement entropy via deep learning. Phys. Rev. D 106 (10), p. 106017. External Links: 2205.04445, Document Cited by: §1. Park et al. (2026) C. Park, S. Kim, and J. H. Lee Holography transformer. Mod. Phys. Lett. A 41 (10), p. 2650046. External Links: 2311.01724, Document Cited by: §1. Plessix (2006) R.-E. Plessix A review of the adjoint-state method for computing the gradient of a functional with geophysical applications. Geophys. J. Int. 167 (2), p. 495–503. External Links: Document Cited by: §6. Raissi et al. (2019) M. Raissi, P. Perdikaris, and G. E. Karniadakis Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. J. Comput. Phys. 378, p. 686–707. External Links: Document Cited by: §1. Rissanen (1978) J. Rissanen Modeling by shortest data description. Automatica 14 (5), p. 465–471. External Links: Document Cited by: §1. Rumelhart et al. (1986) D. E. Rumelhart, G. E. Hinton, and R. J. Williams Learning representations by back-propagating errors. Nature 323 (6088), p. 533–536. External Links: Document Cited by: §4.2. Sato et al. (2017) Y. Sato, S. Kasahara, H. Murayama, Y. Kasahara, E.-G. Moon, T. Nishizaki, T. Loew, J. Porras, B. Keimer, T. Shibauchi, and Y. Matsuda Thermodynamic evidence for a nematic phase transition at the onset of the pseudogap in YBa2_2Cu3_3Oy_y. Nature Phys. 13 (11), p. 1074–1078. External Links: Document Cited by: §5.2. Schmidt and Lipson (2009) M. Schmidt and H. Lipson Distilling Free-Form Natural Laws from Experimental Data. Science 324 (5923), p. 81–85. External Links: Document Cited by: §1. Seo et al. (2024) Y. Seo, S. Kim, and K. K. Kim Construction of superconducting dome and emergence of quantum critical region in holography. Phys. Rev. D 110 (8), p. 086005. External Links: 2312.06321, Document Cited by: §1, §2. She et al. (2011) J. She, B. J. Overbosch, Y. Sun, Y. Liu, K. Schalm, J. A. Mydosh, and J. Zaanen Observing the origin of superconductivity in quantum critical metals. Phys. Rev. B 84, p. 144527. External Links: 1105.5377, Document Cited by: §6. Shukla et al. (2024) K. Shukla, J. D. Toscano, Z. Wang, Z. Zou, and G. E. Karniadakis A comprehensive and FAIR comparison between MLP and KAN representations for differential equations and operator networks. Comput. Methods Appl. Mech. Eng. 431, p. 117290. External Links: 2406.02917, Document Cited by: §4.1. Siopsis and Therrien (2010) G. Siopsis and J. Therrien Analytic Calculation of Properties of Holographic Superconductors. JHEP 05, p. 013. External Links: 1003.4275, Document Cited by: §6. Spillane (2013) M. Spillane Constructing Space From Entanglement Entropy. External Links: 1311.4516 Cited by: §1. Udrescu and Tegmark (2020) S. Udrescu and M. Tegmark AI Feynman: a Physics-Inspired Method for Symbolic Regression. Sci. Adv. 6 (16), p. eaay2631. External Links: 1905.11481, Document Cited by: §1. Witten (1998) E. Witten Anti de Sitter space and holography. Adv. Theor. Math. Phys. 2, p. 253–291. External Links: hep-th/9802150, Document Cited by: §1. Xiao et al. (2026) H. Xiao, Z. He, Z. Xian, and S. Wu Holographic Learning from Fermionic Spectra: Application to Strange Metal Phenomenology. External Links: 2607.02861 Cited by: §1. Yan et al. (2020) Y. Yan, S. Wu, X. Ge, and Y. Tian Deep learning black hole metrics from shear viscosity. Phys. Rev. D 102 (10), p. 101902. External Links: 2004.12112, Document Cited by: §1. You et al. (2018) Y. You, Z. Yang, and X. Qi Machine Learning Spatial Geometry from Entanglement Features. Phys. Rev. B 97, p. 045153. External Links: 1709.01223, Document Cited by: §1. Zhao et al. (2021) C. Zhao, X. Che, Z. Zhang, and F. Huang P-type doping in 2M-WS2_2 for a complete phase diagram. Dalton Trans. 50 (11), p. 3862–3866. External Links: Document Cited by: §5.2.