{"id":"574dd333-a33c-4bb2-a02f-a8332c4d961f","arxiv_id":"1908.01470","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Training a physics-encoded neural network on boundary data recovers the Reissner-Nordström-AdS black hole metric, with mean squared errors ranging from 0.0015 to 0.28 across six charge and topology settings.","lead":"A neural network built from the equations of motion of a test field successfully recovered the geometry of a charged black hole from labeled boundary data in four of six tested configurations. The work extends the 'holography as deep learning' program, which tries to reconstruct curved spacetime from boundary information, from neutral to charged black holes.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (14) does not follow from Eq. (8) under the stated transformation (13): the learned weights are not the metric f(r(η)) itself, so the central claim is unsupported as written.","rationale":"The reader's weakest assumption—per-case regularizer tuning with the answer in hand—is legitimate and experimentally grounded in Table I, where both the coefficient and the η-exponent change from case to case with no selection rule. However, a more load-bearing defect appears earlier in the argument: the coordinate transformation from (8) to (14) is algebraically inconsistent. Under dη = dr/√f, the operator ∂η = √f ∂r, so the term f π' in (8) becomes √f ∂η π, not ∂η π. If π is redefined as ∂η φ, as (15) requires for the Euler step, then the coefficient multiplying π is f_η/(2f) + 2√f/r, not f. For pure AdS this coefficient is 3, while f(r(η)) is e^{2η}; these are plainly different functions. Because the data set is generated by propagating the same (incorrect) e.o.m. with R = f, the neural network can fit the injected R without ever solving the actual scalar e.o.m. in the RN-AdS geometry. In that case the successful MSE values in Appendix A only measure self-consistency of the pipeline, not recovery of the metric. This is a correctness risk that no amount of regularizer tuning can repair. I therefore recommend rejecting the paper in its present form, while noting that a corrigendum deriving the correct η-coordinate e.o.m. and defining the learned quantity precisely could make the benchmark meaningful.","tokens_in":10540,"tokens_out":15767,"duration_ms":157617,"concrete_test":"Derive the η-coordinate e.o.m. from Eq. (8) using Eq. (13), and for the pure AdS case f = r^2 (L = 1, n = 4) compare the coefficient of π with R(η) = e^{2η}: the standard derivation gives a constant coefficient 3, not e^{2η}. Then regenerate the RN1 dataset by propagating the correct e.o.m. and retrain the network with the same architecture and the Table I regularizer; if the learned weights do not match f(r(η)) to the reported MSE, the paper's identification of the learned weights with the RN-AdS metric fails.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's central identification is that the network weight R(η) equals the RN-AdS metric f(r(η)). This rests on Eq. (14), which is asserted to be the η-coordinate e.o.m. obtained from Eq. (8) using dη = dr/√f, Eq. (13). But substituting (13) into (8) does not give (14). Eq. (15) implies π = ∂ηφ, and the correct transformation of (8) is ∂ηπ + [f_η/(2f) + 2√f/r] π − m^2φ − V'(φ) = 0, where f_η = df/dη. For pure AdS with f = r^2, L = 1, n = 4, this coefficient is the constant 3, whereas Eq. (14) would use R(η) = e^{2η}. Therefore the weight W^(n)_22 = 1 − Δη R(η(n)) is not the metric f at the grid point. Moreover, the training labels are generated by propagating the same (incorrect) e.o.m. with R = f, so the network can recover that input R without ever encoding the actual scalar-field dynamics in RN-AdS. The comparison in Appendix A is thus between the network output and the same quantity put into the label-generation rule, not with the physical metric. This issue is prior to the regularizer-tuning problem: even with a fixed regularizer, the object being learned is not the metric. If R(η) were instead intended as the coefficient f'/√f + 2√f/r (or f_η/(2f)+2√f/r), then the figures and tables comparing R with the analytic metric conflate two different functions. Either way, the central claim is not supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript adapts the deep-learning approach of Hashimoto et al. (Ref. [16]) to the Reissner-Nordström-AdS black hole. The authors construct a ten-layer neural network in which the weights are proposed to be the discretized scalar-field equation of motion in a new radial coordinate η defined by dη = dr/√f. They generate a training set by propagating boundary data through the analytic RN-AdS metric, label points with a horizon regularity condition, and train the network to reproduce the labels. The trained weights are interpreted as the RN-AdS metric function R(η), and the results are compared with the analytic metric for two charges (Q=0.5, 0.25) and three horizon topologies (k=0, ±1). The paper reports MSE values between 0.00155 and 0.279 and discusses the influence of learning rate, batch size, and initialization on the loss.","tokens_in":10851,"tokens_out":11529,"duration_ms":115661,"significance":"If the central identification of the network weight with the metric were correct, the paper would extend the holographic deep-learning program to charged black holes with nontrivial topology. The authors are transparent in reporting per-case numerical errors and in showing hyperparameter sensitivity, and they provide a quantitative comparison table. However, the main mathematical step connecting the coordinate-transformed equation to the network weight is incorrect, and the reported agreement is between the network output and the same quantity inserted into the data-generation procedure. As a result, the central claim that the neural network learns the RN-AdS metric from boundary data is not supported by the current derivation and experiments.","major_comments":[{"comment":"Under the coordinate transformation dη = dr/√f in Eq. (13), the scalar equation (8) does not become Eq. (14). If π in (14) is taken to be the new momentum ∂_η φ, the correct equation reads ∂_η(∂_η φ) + [f'(r(η))/(2√f) + 2√f/r(η)] ∂_η φ - m²φ - V' = 0 in the four-dimensional case used in the numerical section; if π is still the old φ'(r), then ∂_ηπ acquires a 1/√f factor on the source term and the coefficient is f'/√f + 2√f/r. In neither case is the coefficient equal to f(r(η)). Therefore the weight W_{22}^{(n)} = 1 - Δη R(η^{(n)}) in Eq. (16) is not the metric at the grid point, and the comparison in Appendix A between R and the analytic f is a comparison between two different quantities. The central claim in Section V that 'the expected metric can be obtained by training this network' is consequently unsupported.","section":"Section II, Eqs. (13)-(16)"},{"comment":"The training set is produced by propagating Eq. (14) through the analytic RN-AdS metric and then labeling the data with the horizon condition (18), which is derived from the same metric. The network architecture in Eq. (16) is exactly the discretized version of the same equation, with R(η) as the trainable weights. Training therefore recovers the R(η) that was used to assign the labels; the agreement in Table II is a check that the optimization finds the generating function of the forward model, not an independent inference of the physical metric from boundary data. To support the holographic claim, the authors should break this circularity—for example, by generating data with the r-coordinate equation (8) and testing whether the η-coordinate network still recovers f—or explicitly restrict the claim to inverting the network's own forward model.","section":"Section III.A and Section IV"},{"comment":"The regularizer terms in Table I have a different coefficient and a different power of η for each of the six RN cases, and Section IV states only that these values were found 'according to experience.' No rule is given for selecting them, and the two cases with the largest errors (RN4: MSE 0.176; RN6: MSE 0.279) show visible mismatches in Fig. 7(d) and 7(f) that are not discussed. This is load-bearing because it means the claimed successes are not reproducible from the described procedure and the method is fragile with respect to the auxiliary terms.","section":"Table I and Fig. 7"}],"minor_comments":[{"comment":"Typos: 'INTRODUCTON' (Section title), 'T raining' (Section III.B heading), 'TABEL' (Appendix A), 'wether' (Section I), 'TyTorch' (Section I) versus 'PyTorch' (Section III.B), and 'epoches' (Section IV) should be corrected.","section":"Throughout"},{"comment":"The second activation function is written as x2 + Δr δV(x1)/(f δφ), which is not a well-formed expression; after the coordinate transformation, Eq. (17) uses δV/δx1. The notation should be made consistent and dimensionally correct.","section":"Eq. (12) and Eq. (17)"},{"comment":"The horizon condition uses 2/η π without derivation. Since this condition is used to assign every label, the paper should derive it from the near-horizon limit of Eq. (14) or from the regularity condition of the scalar field.","section":"Eq. (18)"},{"comment":"The statement that 'the optimal learning rate is around 0.001' should be supported by a quantitative criterion; Fig. 6 shows only that lr=0.1 fails to converge, which does not establish optimality.","section":"Section IV"},{"comment":"The r values differ from case to case; please state how the radial grids are chosen and whether all cases use the same η-range and layer count.","section":"Appendix A, Table II"},{"comment":"The paper does not provide the code, the initialization scheme, or the exact training schedule (number of epochs, optimizer parameters), so the numerical results are not independently reproducible from the text alone.","section":"Reproducibility"},{"comment":"Panels (e) and (f) are hard to read because the two curves have similar line styles; use distinct markers or colors for the reappeared and emergent metrics.","section":"Fig. 7"}],"recommendation":"reject","confidential_remarks":"I recommend rejection because the coordinate-transformation error in Section II invalidates the identification of the learned weights with the metric. Correcting this requires re-deriving the network equations, re-running all experiments, and reinterpreting the results, which is beyond a standard major revision. The per-case regularizer tuning and the unexplained failures for RN4 and RN6 further weaken the reproducibility of the claimed results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nI read Tan and Chen's paper on learning the RN-AdS metric with a deep network. It is a direct extension of Hashimoto et al. to include charge and three topologies, and parts of the implementation are sensible. But the central identification of the network weight with the metric is wrong.\n\nThe problem is in the coordinate transformation. The authors state that with dη = dr/√f, the e.o.m. becomes ∂ηπ + R(η)π − m²φ − V' = 0, with R(η) the reproduced metric f(r(η)). That equation is not what you get from the original (8). If you actually substitute and define π = ∂ηφ, the coefficient of π is f_η/(2f) + 2√f/r, not f(r). In pure AdS f = r², that combination is about 4 − 1/r, not r². So the weight W^(n)_22 = 1 − Δη R(η) does not represent the metric at the grid point, unless R is redefined to mean the coefficient combination—but then comparing R to the analytic f(r) in Table II conflates two different functions.\n\nBecause the same (incorrect) e.o.m. is used to generate the boundary data and labels, the network is recovering the function R that was inserted into the data-generation rule. The Appendix A comparison is a self-consistency check, not evidence that the physical RN-AdS metric was learned.\n\nOther issues are secondary but real. The regularizer coefficients and η-exponents in Table I are chosen 'according to experience,' with no rule for choosing them, and they differ for every case. Two of the six configurations (RN4 and RN6) fail with MSE 0.176 and 0.279; the paper does not explain why. No code or data are provided, which makes it impossible to verify the training details.\n\nCredit where it is due: the idea of moving to a coordinate where the metric drops out of the activation function is a real technical step, and the numerical integration over η is a reasonable workaround for non-integrable metric functions. The scan over k = 0, ±1 is useful. If the transformation error were fixed, this could become a valid proof-of-concept.\n\nAs it stands, the central claim is unsupported. I would not send this to a serious referee for the science; a desk rejection is appropriate. The only value in outside review would be to put the derivation error on record in a few lines. For a reading group, it is a useful cautionary example of how a change of variables can hide a missing term.\n\nRecommendation: desk reject.","headline":"The charged black hole extension rests on a wrong change-of-variables, so the network learns an input function rather than the metric.","tokens_in":11467,"tokens_out":7634,"would_cite":false,"duration_ms":73063,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A ten-layer neural network whose weights encode the scalar field equation recovers the charged anti-de Sitter black hole metric from boundary data.","keywords":["deep learning","holography","AdS/CFT correspondence","Reissner-Nordström black hole","metric reconstruction","neural network","emergent spacetime"],"falsifier":"Fix a single regularization coefficient and $\\eta$-exponent, say the RN1 values, and train all six cases; if the recovered metrics separate into good and bad fits with errors far outside the reported range, the per-case tuning is doing essential work and the metric is not being learned from data alone.","tokens_in":10213,"feed_emoji":"🕳️","tokens_out":9042,"duration_ms":80488,"temperature":0.7,"pith_summary":"The paper tries to show that a deep neural network can recover the Reissner-Nordström–AdS black hole metric from boundary data alone, extending an earlier Schwarzschild demonstration to charged black holes and to planar, spherical, and hyperbolic horizons. The network is not a generic black box: its layer weights are the discretized scalar-field equation of motion in a coordinate $\\eta$ defined by $d\\eta = dr/\\sqrt{f(r)}$, so training the weights is equivalent to learning the metric function $R(\\eta)$ at ten radial layers. If the claim is right, it is concrete evidence that bulk geometry can emerge from boundary information plus known bulk dynamics, which is what the holographic correspondence asserts. The paper also reports how learning rate, batch size, and initialization affect convergence, with an optimal learning rate near 0.001.","feed_headline":"Neural nets recover the charged black hole metric","feed_subtitle":"Ten-layer network trained on boundary data matches the analytic charged-black-hole metric across six cases.","key_machinery":"The load-bearing object is a 10-layer feedforward network in which the linear transformation between layers is the discretized scalar equation of motion in the coordinate $\\eta$; the weight matrix is $W^{(n)} = \\begin{pmatrix} 1 & \\Delta\\eta \\\\ \\Delta\\eta\\,m^2 & 1-\\Delta\\eta\\,R(\\eta^{(n)}) \\end{pmatrix}$, and the activation function encodes the potential $V(\\varphi)$. A coordinate transformation $d\\eta = dr/\\sqrt{f}$ removes the metric function $f$ from the weights and activation, so the only unknown to be learned is the metric function $R(\\eta)$ at each layer. The final activation is a smoothed step function of the horizon boundary condition, and the loss is $L^1$ distance to the labels plus a per-case regularization term.","core_discovery":"On the paper's own terms, the central discovery is that the expected metric can be obtained by training this network: for each of six cases (charge $Q=0.5$ or $0.25$; topology $k=0,1,-1$), the trained ten-point emergent metric matches the analytic metric with mean-square errors between 0.00155 and 0.27903. The architecture encodes the discretized equation of motion $\\partial_\\eta \\pi + R(\\eta)\\pi - m^2\\varphi - \\delta V/\\delta\\varphi = 0$ in its weights, and the data are boundary values of $\\varphi$ and $\\pi$ labeled by whether they satisfy the horizon condition at $\\eta_{\\rm fin}$. The authors take this as evidence that by learning boundary CFT data, the network reproduces the AdS metric even for spacetimes with charge and different topology.","pith_inferences":["A sharper version of the paper's test would hold out the analytic metric: fix one regularization rule, train on all six cases, and compare errors; the reported table does not show that this succeeds.","If the regularization hurdle can be removed, the same discretized-equation network could be aimed at unknown bulk metrics, with the horizon labeling condition as the only input beyond boundary data.","The ten-layer discretization means only ten points of $R(\\eta)$ are learned; scaling to more layers and studying interpolation error would show whether the method converges to the full metric."],"forward_implications":["For any of the tested charges $Q=0.5,0.25$ and topologies $k=0,1,-1$, a trained network yields a metric whose ten sampled values approximate the analytic Reissner-Nordström–AdS metric.","Training with a suitable learning rate and batch size makes the loss converge to the same value given enough epochs, while a learning rate near 0.1 prevents convergence; this gives practical guidance for similar reconstructions.","Because the coordinate transformation $d\\eta = dr/\\sqrt{f}$ moves the metric out of the activation function, the same architecture can in principle be applied to other static black holes once the transformation is computed numerically.","The recovered metric is discrete by construction, so the method directly produces the metric only at the ten $\\eta$ layers used in the network."],"supporting_citations":[{"why":"Supplies the original deep-learning metric reconstruction scheme—network built from discretized equations of motion, data generation, loss, and activation—that this paper adapts to the charged case.","marker":"[16]"},{"why":"Defines the Reissner-Nordström charged black hole solution family that is the target metric the network learns and is compared against.","marker":"[17–20]"},{"why":"Establishes the AdS/CFT correspondence that supplies the reason learning boundary data should reproduce a bulk metric.","marker":"[1–3]"}],"fun_headline_variants":["AI learns charged black hole metric from boundary data","Deep learning reconstructs RN metric across six cases","Neural net matches analytic charged black hole","Charged holographic metric learned by deep net","Deep learning recovers Reissner-Nordstrom black hole"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reconstruction stands on the assumption that the per-case regularization term can be chosen without already knowing the answer; the paper gives no rule for selecting the coefficient or the $\\eta$-exponent, and the two largest reported errors suggest the choice is fragile.","fun_headline_variants_meta":{"raw":{"variants":["AI learns charged black hole metric from boundary data","Deep learning reconstructs RN metric across six cases","Neural net matches analytic charged black hole","Charged holographic metric learned by deep net","Deep learning recovers Reissner-Nordstrom black hole"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000229,"raw_usage":{"total_tokens":1414,"prompt_tokens":818,"completion_tokens":596,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":434,"completion_tokens_details":{"reasoning_tokens":522}},"tokens_in":434,"tokens_out":596,"duration_ms":6498,"temperature":1.0,"reasoning_tokens":522,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:12:58.091720+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix a single regularization coefficient and $\\eta$-exponent, say the RN1 values, and train all six cases; if the recovered metrics separate into good and bad fits with errors far outside the reported range, the per-case tuning is doing essential work and the metric is not being learned from data alone.","supporting_citations":[],"review_version":1}