{"id":"185e6837-bb4f-47c6-a881-8b038cdbaa4f","arxiv_id":"2412.00527","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A mixed least-squares deep neural network recovers anisotropic conductivity tensors from internal gradient measurements, with error estimates and 2D/3D numerical demonstrations.","lead":"This paper introduces a deep neural network method, MLS-DNN, that reconstructs an anisotropic conductivity tensor from internal measurements of solution gradients. The authors derive error bounds and demonstrate the method in 2D and 3D tests with up to 10% noise, comparing it to finite elements and physics-informed networks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The error bounds apply only to exact global minimizers, while the reported reconstructions come from ADAM on a nonconvex loss; the O(δ^(1/2)) prediction therefore does not control the 10%-noise results.","rationale":"I agree with the reader that the weakest link is the gap between the global-minimizer analysis and the ADAM-trained solutions. The paper is otherwise valuable: the numerical evidence is extensive (12 examples, 2D/3D, partial data, up to 10% noise), and the mixed least-squares formulation with a matrix projection is a sensible extension of [30]. No formal verification exists, and the hyperparameters are tuned to minimize the reported error, which inflates confidence. The proof flaw in Theorem 3.1 is real but appears repairable by replacing the asserted equality with the projection inequality and bounding (A†, e) from above; it does not by itself invalidate the numerical results. A concrete multiseed robustness experiment would settle whether the local-minimum issue actually undermines the headline accuracy claim. If the method is robust across seeds and optimizers, the CONDITIONAL verdict can be upgraded; if not, the paper should be revised to add an optimization-error analysis or at least qualify the theoretical contribution. Since the reader already assigned CONDITIONAL, my recommendation is UNCHANGED.","tokens_in":27059,"tokens_out":11333,"duration_ms":178828,"concrete_test":"Rerun Example 4.1 (Neumann, exact data and δ=10%) with 20 independent random initializations of the DNN parameters, keeping the reported architecture, sampling sizes, and hyperparameters fixed; record the final relative L2 error e(Â) for each seed. Also apply L-BFGS starting from the ADAM solution and record whether the empirical loss and e(Â) improve. If the spread of e(Â) is small (e.g., all within a factor of 2 of the reported values) and L-BFGS does not substantially reduce the loss, the local-minimum concern is not decisive; if some seeds produce errors an order of magnitude larger, the reported accuracy is not robust, and the global-minimizer theory does not explain the observed performance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorems 3.1 and 3.2 bound the reconstruction error of any global minimizer (θ*,κ*) or (θ̂*,κ̂*) of the population or empirical loss. The implemented method minimizes the empirical loss (2.7)/(2.13) with ADAM over tanh networks, and the loss landscape is nonconvex. The paper explicitly concedes (Section 4) that 'the optimizer may get trapped in a local minimum, leading to pronounced optimization errors' and that for exact data 'the reconstruction cannot be made arbitrarily accurate.' Consequently, the numerical solutions—including the headline 2D/3D results at 10% noise—are local minimizers, not the global minimizers appearing in the theorems. Without a quantitative estimate of the optimization error, the theoretical claim that the method recovers the conductivity at the O(δ^(1/2)) rate is not connected to the reported reconstructions. A secondary internal issue is that the proof of Theorem 3.1 asserts an equality (A†, A†−PK(A*θ)) = ((A†−PK(A*θ))∇u†, ∇ψ) that follows from the projection property only as an inequality; this is repairable but further undermines the theorem as written. The central claim 'accurately recover ... up to 10% noise' rests on the numerical experiments; those experiments are not yet shown to be stable with respect to initialization and hyperparameter choice, and the hyperparameters were explicitly tuned to minimize the reported error.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a deep-learning approach (MLS-DNN) for recovering an anisotropic conductivity tensor A from internal measurements of solutions to an elliptic boundary value problem. The method is based on a mixed least-squares formulation in which the flux sigma and the conductivity A are represented by tanh neural networks, with the constraint A in K enforced by a matrix-valued projection. The authors prove error estimates for the L2 reconstruction error of the projected DNN conductivity for global minimizers of both the population loss (Theorem 3.1) and the empirical loss (Theorem 3.2), assuming a projected source condition, DNN approximation rates, and Rademacher-type statistical bounds. They also present numerical experiments for Neumann and Dirichlet problems in 2D and 3D, with full or partial data and up to 10% noise, comparing with FEM and PINN.","tokens_in":27253,"tokens_out":8616,"duration_ms":80324,"significance":"If the theoretical results were fully rigorous and connected to the implemented algorithm, the paper would offer an important extension of the Kohn-Lowe mixed formulation and of the isotropic MLS-DNN framework to anisotropic problems, with the first numerical demonstrations in 3D. The numerical study is genuinely useful: it shows stable recovery for several nontrivial anisotropic examples, including partial interior data, and the authors state that code will be released. The main caveats are that the error bounds apply only to exact global minimizers, which the ADAM-trained networks need not be, and that one step in the proof of Theorem 3.1 rests on an equality that is not implied by the source condition. These issues limit the theoretical support for the headline claim of accurate recovery up to 10% noise, although the empirical evidence stands independently.","major_comments":[{"comment":"The displayed equality (A†, A†−PK(A*θ)) = ((A†−PK(A*θ))∇u†, ∇ψ) is not an identity under Assumption 3.2; the assumption gives A† = PK(∇u†⊗∇ψ), not A† = ∇u†⊗∇ψ. From the projection inequality (2.4) one only obtains the inequality (A†, A†−PK(A*θ)) ≤ ((A†−PK(A*θ))∇u†, ∇ψ). Since the subsequent argument bounds the absolute value, replacing the equality by this inequality appears to repair the proof, but the current text contains a false equality and should be corrected.","section":"Section 3.1, proof of Theorem 3.1"},{"comment":"The theorems bound the reconstruction error of global minimizers of the population and empirical losses, but the reported reconstructions are obtained by ADAM on the nonconvex empirical loss. The paper itself states in Section 4 that 'the optimizer may get trapped in a local minimum, leading to pronounced optimization errors' and that for exact data 'the reconstruction cannot be made arbitrarily accurate.' Consequently, the O(δ^{1/2}) error bound is not established for the computed solutions, and the numerical results do not verify the theorem quantitatively. The authors should either add an optimization-error estimate or an explicit condition under which the ADAM solution is close to a global minimizer, or clearly separate the theoretical claim (for global minimizers) from the numerical demonstration (which is then purely empirical).","section":"Section 3 (Theorems 3.1 and 3.2) and Section 4"},{"comment":"The theoretical noise model in (2.3) and (2.11) measures δ = ∥u† − zδ∥ in H^1(Ω) or H^{3/2}(Ω), while the experiments add i.i.d. Gaussian noise pointwise to ∇u† (Section 4: 'i.i.d. Gaussian noise is added pointwise to ∇u†'). A pointwise-perturbed gradient is generally not in H^1(Ω), so the experimental data do not satisfy the regularity assumptions of Theorem 3.2. If the noisy gradient is meant to be a smoothed version of raw data, the smoothing procedure and its effect on δ should be described; otherwise the numerical experiments do not directly test the noise-dependence predicted by the theory.","section":"Section 4, noise generation"},{"comment":"Lemma 3.2 is central to the empirical-loss bound, but its proof is omitted and is referred to [30, Theorem 3.4]. The loss here involves the anisotropic projection PK(Aθ) and additional boundary terms, so it is not immediately obvious that the argument transfers verbatim. A proof or a precise statement of the required changes should be included, rather than delegating to the authors' previous paper.","section":"Lemma 3.2"}],"minor_comments":[{"comment":"The tuple γ is defined as (γσ, γb, γq) ∈ R^3_+, but γq never appears; the third component should be γA.","section":"Section 2.2, after (2.5)"},{"comment":"The sentence 'This strategy is often used for tuning hyper-parameters in deep learning' refers to choosing hyperparameters that attain the smallest reconstruction error on the test examples; this selection procedure should be stated in the abstract or conclusions if the claimed accuracy is based on it.","section":"Section 4, paragraph after Table 1"},{"comment":"There are several typos: 'b e tween' after (2.4), 'reonstruction' in Table 2 heading, 'off-shelf' in Section 4, and 'Combing' in the proof of Theorem 3.1. These should be corrected.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The reader's report raised the false-equality step and the global-minimizer gap; both are genuine and are reflected in the major comments. The reliance on [30] for Lemma 3.2 is a self-citation concern; I would ask the authors to provide the proof or a careful transfer argument rather than simply citing. The numerical experiments are generally convincing and the 3D results are a useful addition; the paper is close to publishable after the theory is repaired and reframed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a useful numerical paper that extends the isotropic MLS-DNN method to anisotropic conductivity tensors. The numerics are the strongest part: 2D and 3D reconstructions, partial internal data, noise up to 10%, comparisons against FEM and PINN, and a code release. The matrix-valued projection onto the admissible set and the source-condition-based analysis are a genuine extension of the prior isotropic work [30], and the 3D experiments are, as far as I know, the first numerical demonstrations for anisotropic conductivity in 3D in this line. If you work on DNN-based inverse problems or conductivity imaging, this is worth knowing about.\n\nThe soft spots are in the theory. Theorem 3.1's proof uses the identity (A†, A†−P) = ((A†−P)∇u†, ∇ψ), where P = P_K(A*θ). That identity would require A† = ∇u†⊗∇ψ, but Assumption 3.2 only gives A† = P_K(∇u†⊗∇ψ). The projection property (2.4) is a variational inequality, not an equality, so the identity doesn't follow. This is a genuine gap in the written proof, and it carries into Theorem 3.2 through the same argument. I think it's repairable, but as stated the error bound isn't proven.\n\nThe second issue is the usual one for DNN methods: Theorems 3.1 and 3.2 bound the reconstruction error of a global minimizer of the population or empirical loss. The reported reconstructions come from ADAM, and the paper itself admits the optimizer can get stuck in local minima and that there is an accuracy floor even for exact data. So the O(δ^(1/2)) rate shouldn't be read as covering the experiments. The authors are transparent about this, which is to their credit, but the abstract's claim 'the method can accurately recover...' leans entirely on the numerics.\n\nLemma 3.2 is quoted from the prior isotropic paper [30] without proof. That's acceptable, but it is a reliance on a self-citation that a referee will want checked. Also, hyperparameters were tuned to minimize the reported error, so the error tables are optimistic—standard practice in deep learning, but worth remembering.\n\nOverall: the numerical contribution is real and likely to be useful; the theory is not reliable as written. I'd send it to peer review, but I'd ask for a corrected proof of Theorem 3.1 and a clearer separation of the ideal global-minimizer result from the practical optimization result.","headline":"Solid anisotropic extension of the MLS-DNN method with convincing numerics, but the proof of the main error bound has a real gap and the theory only covers global minimizers, so the theorem is unproven as written.","tokens_in":27924,"tokens_out":5488,"would_cite":false,"duration_ms":49677,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35R30","65N21","65N30","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper develops a mixed least-squares deep neural network that approximates the internal flux and the anisotropic conductivity tensor simultaneously from internal measurements of the solution, and proves error bounds of order…","keywords":["anisotropic conductivity","inverse problem","deep neural network","mixed least-squares","internal measurements","error estimate","elliptic PDE","conductivity imaging"],"falsifier":"Choose a smooth full anisotropic tensor $A^\\dagger$ that does not satisfy the projected source condition $A^\\dagger=P_K(\\nabla u^\\dagger\\otimes\\nabla\\psi)$ and run the MLS-DNN with exact data, taking $\\gamma_A\\to 0$ and large DNN capacity. If the empirical $L^2$ error $e(\\hat A)$ stops decreasing at a positive plateau rather than following the predicted $O(\\delta^{1/2})$ rate, the bound's premise --- a global minimizer plus the source condition --- is not what governs the reported accuracy.","tokens_in":26699,"feed_emoji":"🧮","tokens_out":10166,"duration_ms":87020,"temperature":0.7,"pith_summary":"This paper claims that an anisotropic conductivity tensor in an elliptic boundary value problem can be recovered from internal measurements of the solution by training two deep neural networks --- one for the conductivity and one for the internal flux --- against a mixed least-squares loss that enforces the first-order system. If correct, this gives a practical, mesh-free route to anisotropic conductivity imaging in two and three dimensions without paired training data, and the analysis supplies error bounds for this DNN formulation. The authors prove an $O(\\delta^{1/2})$ bound on the $L^2$ reconstruction error under a projected source condition, for both the population loss and its Monte Carlo empirical version, and support the claim with experiments in 2D and 3D, for Neumann and Dirichlet problems, full and partial internal data, and up to 10 percent noise.","feed_headline":"Neural nets recover full conductivity tensor to 10% noise","feed_subtitle":"Mixed least-squares DNN images anisotropic conductivity in 2D and 3D, with proved error bounds and FEM/PINN tests.","key_machinery":"The load-bearing object is the mixed least-squares loss built from the first-order system $\\sigma_i=A\\nabla u_i$, $-\\nabla\\cdot\\sigma_i=f_i$, with the relevant boundary condition, so that the conductivity and the flux are both unknowns fitted to the internal measurements. The loss uses a matrix-valued projection $P_K$, applied pointwise by eigenvalue decomposition and clipping eigenvalues to $[c_0,c_1]$, to keep $A_\\theta$ inside the admissible set of uniformly elliptic symmetric matrices; the projection also enters the error analysis through the variational inequality that characterizes it. The convergence proof combines three tools: DNN approximation rates for $\\tanh$ networks (giving the representation errors $\\epsilon_A,\\epsilon_\\sigma$), the projected source condition $A^\\dagger=P_K(\\nabla u^\\dagger\\otimes\\nabla\\psi)$ that turns the $L^2$ error into computable residuals, and Rademacher-complexity bounds for the Monte Carlo empirical loss.","core_discovery":"The central claim is that the mixed least-squares deep neural network recovers the anisotropic conductivity tensor $A^\\dagger$ by minimizing the loss $J_\\gamma(\\theta,\\kappa) = \\sum_i(\\|\\sigma_{i,\\kappa}-P_K(A_\\theta)\\nabla z_i^\\delta\\|^2_{L^2(\\Omega)^d} + \\gamma_\\sigma\\|\\nabla\\cdot\\sigma_{i,\\kappa}+f_i\\|^2_{L^2(\\Omega)} + \\gamma_b\\|n\\cdot\\sigma_{i,\\kappa}-g_i\\|^2_{L^2(\\partial\\Omega)} + \\gamma_A\\|P_K(A_\\theta)\\|^2_{L^2(\\Omega)^{d,d}})$ for the Neumann problem, with a Dirichlet analogue, where $P_K$ is the orthogonal projection onto the admissible set of uniformly elliptic symmetric matrices. Theorem 3.1 gives $\\|A^\\dagger - P_K(A^*_\\theta)\\|_{L^2(\\Omega)^{d,d}} \\le c(\\gamma_A^{-1}\\eta^2+\\eta+(1+\\gamma_\\sigma^{-1}+\\gamma_b^{-1})\\gamma_A)^{1/2}$ with $\\eta$ combining DNN approximation error, noise level $\\delta$, and the penalties, and Theorem 3.2 extends the bound to the empirical loss with Monte Carlo quadrature error terms. With the parameter choice $\\gamma_A\\sim\\delta$, $\\epsilon_A=\\epsilon_\\sigma\\sim\\delta$, the rate is $O(\\delta^{1/2})$, comparable to output least-squares Galerkin methods. The experiments demonstrate reconstruction of full matrix-valued conductivities in 2D and 3D with data noise up to 10 percent, including from partial internal data.","pith_inferences":["A testable extension would run the same scheme on conductivities that violate the projected source condition; the theory makes no prediction for them, so a numerical rate study would show whether the method's practical accuracy outruns the proved bound.","The reported insensitivity to the penalty parameter $\\gamma_A$ suggests the DNN's implicit spectral bias is doing regularization work that the analysis assigns to the explicit $L^2$ penalty; separating these effects would require measuring the loss gap between the found local minimizer and the global one.","The same mixed least-squares structure could be adapted to power-density data $A\\nabla u_i\\cdot\\nabla u_j$ or to time-harmonic problems, where internal data are phase-sensitive; the projected source condition would need re-derivation for those settings."],"forward_implications":["An anisotropic conductivity tensor in 2D and 3D can be reconstructed from internal solution measurements without labeled training pairs, using an unsupervised DNN that fits flux and conductivity simultaneously.","With $\\gamma_A\\sim\\delta$ and representation errors of order $\\delta$, the reconstruction error in $L^2(\\Omega)^{d,d}$ is of order $\\delta^{1/2}$, the same rate as classical output least-squares Galerkin methods.","The method handles partial internal data: reconstructions remain accurate over the whole domain even when measurements are confined to a boundary band, where the paper's FEM comparison shows poor stability.","The mesh-free DNN formulation makes 3D implementation nearly identical to 2D, and the paper demonstrates 3D anisotropic reconstructions at 5 percent noise.","For both Neumann and Dirichlet problems, the reconstruction degrades only mildly at 10 percent noise, while the PINN variant requires boundary knowledge $A^\\dagger|_{\\partial\\Omega}$ and second-order derivatives."],"supporting_citations":[{"why":"Supplies the mixed least-squares variational formulation the loss is based on.","marker":"[33]"},{"why":"Provides the tanh-DNN approximation rates used to bound the representation errors.","marker":"[19]"},{"why":"Gives the Galerkin FEM error analysis and source-condition framework the rate comparison is made against.","marker":"[14]"},{"why":"Is the isotropic MLS-DNN predecessor whose proof pattern is extended to the anisotropic case.","marker":"[30]"},{"why":"Supplies the Rademacher-complexity learning-theoretic tools behind the empirical-loss bound.","marker":"[4]"},{"why":"Supplies the uniform-convergence and risk-bound results used for the Monte Carlo quadrature errors.","marker":"[8]"},{"why":"Establishes convergence and regularity results for anisotropic conductivity imaging that motivate the assumptions.","marker":"[35]"},{"why":"Proves identifiability of anisotropic conductivity from current densities, the uniqueness backdrop for the inverse problem.","marker":"[6]"}],"fun_headline_variants":["DNN recovers anisotropic conductivity to 10% noise","Mixed least-squares DNN images full conductivity tensor","Neural net maps conductivity tensor from internal data","Imaging conductivity tensor with MLS-DNN to 10% noise","Proved bounds for conductivity tensor recovery by DNN"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The error bounds apply only to an exact global minimizer of the nonconvex training loss, but the actual computations use a gradient optimizer that may stop at a local minimum; the paper states that optimization errors can be pronounced and that for exact data the reconstruction cannot be made arbitrarily accurate.","fun_headline_variants_meta":{"raw":{"variants":["DNN recovers anisotropic conductivity to 10% noise","Mixed least-squares DNN images full conductivity tensor","Neural net maps conductivity tensor from internal data","Imaging conductivity tensor with MLS-DNN to 10% noise","Proved bounds for conductivity tensor recovery by DNN"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000192,"raw_usage":{"total_tokens":1401,"prompt_tokens":1056,"completion_tokens":345,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":672,"completion_tokens_details":{"reasoning_tokens":266}},"tokens_in":672,"tokens_out":345,"duration_ms":3832,"temperature":1.0,"reasoning_tokens":266,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:17:31.401719+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Choose a smooth full anisotropic tensor $A^\\dagger$ that does not satisfy the projected source condition $A^\\dagger=P_K(\\nabla u^\\dagger\\otimes\\nabla\\psi)$ and run the MLS-DNN with exact data, taking $\\gamma_A\\to 0$ and large DNN capacity. If the empirical $L^2$ error $e(\\hat A)$ stops decreasing at a positive plateau rather than following the predicted $O(\\delta^{1/2})$ rate, the bound's premise --- a global minimizer plus the source condition --- is not what governs the reported accuracy.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the mixed least-squares variational formulation the loss is based on."},{"cited_title":"G¨ uhring and M","cited_arxiv_id":null,"evidence_quote":"Provides the tanh-DNN approximation rates used to bound the representation errors."},{"cited_title":"Deckelnick and M","cited_arxiv_id":null,"evidence_quote":"Gives the Galerkin FEM error analysis and source-condition framework the rate comparison is made against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the isotropic MLS-DNN predecessor whose proof pattern is extended to the anisotropic case."},{"cited_title":"Anthony and P","cited_arxiv_id":null,"evidence_quote":"Supplies the Rademacher-complexity learning-theoretic tools behind the empirical-loss bound."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the uniform-convergence and risk-bound results used for the Monte Carlo quadrature errors."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes convergence and regularity results for anisotropic conductivity imaging that motivate the assumptions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Proves identifiability of anisotropic conductivity from current densities, the uniqueness backdrop for the inverse problem."}],"review_version":1}