{"id":"1583307f-a267-4258-8058-513959f9a4a3","arxiv_id":"2411.15693","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Proposes decoupling and direct iterative residual losses that extend neural-network eigenvalue solvers to multi-group neutron diffusion, validated on 1D, 2D and 3D reactor benchmarks.","lead":"Two new loss functions let neural networks solve multi-group neutron diffusion eigenvalue problems, including 2D and 3D reactor benchmarks, without using prior data. The method reports more accurate eigenvalues and fluxes than an earlier single-group neural network approach on the TWIGL and IAEA problems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The fixed-point iteration in Algorithm 1 is unanalyzed, and the 3D IAEA results suggest it does not converge to the principal eigenpair, so the claim of solving multi-group problems is not yet established.","rationale":"The paper's core claim is empirical: the proposed decoupling and direct-iterative losses allow neural networks to solve multi-group critical problems without prior data. For this to hold, the fixed-point iteration defined by Algorithm 1 must converge to the principal eigenpair. The reader correctly flags this as the weakest assumption. My stress test sharpens the concern: the loss functions (23) and (25) are residuals against frozen previous-iterate fluxes, so a small loss at the end of training only certifies that the current iterate is consistent with the previous iterate, not with the original coupled equation (1) unless the fixed point is actually reached. The 3D IAEA numbers (Table 9) and the paper's own limitation statement in Section 5 indicate that the fixed point is not fully reached in that benchmark. This is not a disagreement with standard practice; it is a correctness risk in the central claim. The proposed test (re-evaluating the original residual at the final model) would settle whether the reported solutions genuinely satisfy the continuous problem. I find no reason to change the reader's CONDITIONAL verdict: the method is plausible and works on simpler benchmarks, but the convergence of the iteration and the 3D accuracy need demonstration.","tokens_in":18749,"tokens_out":4105,"duration_ms":39496,"concrete_test":"Re-evaluate the saved 3D IAEA model from Table 9: compute the normalized residual of the original coupled two-group equations (1) at the final neural-network fluxes with k = k^{NN}_eff, i.e. (||R_1||_2 + ||R_2||_2) / ||fission source||_2, where R_1 and R_2 are the left-hand minus right-hand sides of (1). If this relative residual is not below 1e-2, the decoupling iteration has not reached a fixed point and the reported flux/eigenvalue errors do not establish that the original eigenvalue problem was solved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Algorithm 1 (Section 3.4) defines a fixed-point iteration: at each epoch the decoupling loss (23) and direct iterative loss (25) freeze the previous network outputs phi^{NN(i-1)} as targets while updating phi^{NN(i)}, and the eigenvalue is updated via the Rayleigh quotient (30) evaluated at the current approximate eigenfunction. No convergence analysis, contraction estimate, or safeguard against locking onto a non-principal mode is given. The numerical evidence in Table 9 suggests the fixed point is not reached in harder problems: the 3D IAEA thermal flux L2 error is 2.6125e-01 and the fast flux error is 1.2472e-01, and Section 5 concedes that 'the accuracy achieved by the neural network is not yet satisfactory.' If the iteration stalls before the fixed point, a small decoupling residual does not imply that the original coupled equation (1) is satisfied, and the Rayleigh-quotient eigenvalue estimate can be biased. Since the paper's central claim is that the new loss functions solve multi-group eigenvalue problems 'without prior data,' the absence of a convergence guarantee combined with large 3D errors is load-bearing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two new physics-informed residual loss functions—a decoupling loss (23) and a direct iterative loss (25)—for solving multi-group neutron diffusion eigenvalue problems with neural networks, following the source-iteration idea of classical reactor physics. The neural network is a residual network with per-region outputs, trained on residual, boundary, and interface point sets, with the eigenvalue updated via a Rayleigh quotient (30) and a power normalization step (Algorithm 1). Numerical experiments are presented for 1D (Ringhals-4), 2D (TWIGL, 3-region TWIGL, IAEA), and 3D (TWIGL, IAEA) benchmarks and compared against FreeFEM++ references. The reported eigenvalue relative errors are on the order of 1e-4 to 6e-3 and most 2D flux relative L2 errors are below about 5%, but 3D flux errors are considerably larger (e.g., IAEA-3D thermal flux L2 error of 2.6e-1). The central claim is the first neural-network solution of multi-group eigenvalue problems without prior data.","tokens_in":19050,"tokens_out":5691,"duration_ms":48386,"significance":"If correct, the paper's contribution is a practical way to extend physics-informed neural networks from single-group to multi-group eigenvalue problems, including discontinuous interface problems. A strength is that the reference solutions are used only for evaluation, not training, so the reported accuracy is not circular. The sampling-rate and interface-condition ablation studies are informative for practitioners. However, the contribution is weakened by the lack of any convergence analysis for the proposed fixed-point iteration, by the unexamined use of the Rayleigh quotient for a non-self-adjoint problem, and by the large 3D errors, which conflict with the abstract's claim to 'solve' 3D problems. These issues make the central claim currently incompletely supported.","major_comments":[{"comment":"The fixed-point iteration in Algorithm 1 is unanalyzed. The residual losses freeze the previous iterate phi^{NN(i-1)} as a known source while minimizing a squared residual for a single unknown, and then the eigenvalue is updated using the current eigenfunction in the Rayleigh quotient (30). No convergence proof, contraction estimate, or safety check is given to show that a fixed point of this iteration solves the original coupled eigenvalue problem (1) and selects the smallest eigenvalue (keff). The large 3D errors in Table 9 (IAEA-3D ER,2(phi2) = 2.6125e-01) are consistent with the iteration failing to reach the fixed point. Because the abstract claims the method 'solves' multi-group problems, this missing guarantee is load-bearing. I ask for either a convergence argument, a practical stopping rule based on the original coupled residual, or a scoped claim that does not assert 3D solving.","section":"Section 3.4, Algorithm 1 and Eqs. (22)-(25), (30)"},{"comment":"The Rayleigh quotient in (30) is used without discussing its validity for the non-self-adjoint generalized eigenvalue problem. The two-group neutron diffusion operator is not symmetric when scattering and material discontinuities are present, so the quotient <L phi, phi>/<F phi, phi> does not generally provide the correct eigenvalue for the intended right eigenpair unless the operator is self-adjoint or a left-eigenfunction weighting is used. The paper should justify that the iteration's fixed point has the quotient equal to 1/keff + sigma, or replace the update by a more robust estimator (e.g., a fission-source normalization), and demonstrate numerically that the estimated eigenvalue converges to the FreeFEM value. Without this, the eigenvalue results are not fully supported.","section":"Section 3.3, Eq. (30)"},{"comment":"The abstract claims that the new loss functions solve one-, two-, and three-dimensional multi-group problems without prior data. The reported 3D results do not support this: Table 9 lists ER,2(phi2) = 2.6125e-01 for IAEA-3D, and Section 5 explicitly states that 'the accuracy achieved by the neural network is not yet satisfactory.' Please either present the 3D experiments as preliminary, relax the abstract to 1D/2D (or to 'with moderate 3D accuracy'), or add acceptance criteria that the 3D results meet.","section":"Abstract and Section 5, Table 9"}],"minor_comments":[{"comment":"Several entries in Tables 3, 5, and 7 appear as concatenated repeated values (e.g., '1.0351e-041.0351e-041.0351e-04' in Table 3). Please fix the formatting so that each entry appears once.","section":"Tables 3, 5, 7"},{"comment":"The phrase 'significant infinite error in phi2 in the IAEA problem' is confusing; Table 9 reports L2 errors, so please clarify whether L-infinity error is meant and add that metric if so.","section":"Section 4.4"},{"comment":"The denominator notation 'MP' is not defined; it should be a sum over the M residual points.","section":"Eq. (30)"},{"comment":"The symbol n is used both for the number of unknown solution components and for the outward normal direction; please use distinct symbols to avoid ambiguity.","section":"Section 3.1 and Eq. (2)"},{"comment":"No random seeds, repeated runs, or code/data are reported; since neural-network training is stochastic, please provide statistics over several runs or at least seeds and code for reproducibility.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"In my view the paper is appropriate for a journal in computational mathematics if the 3D claim is toned down and the convergence/non-self-adjointness concerns are addressed. The authors may benefit from consulting the adjoint or left-eigenvector formulation of the k-eigenvalue problem. I would also encourage the editors to ask for code/data release given the stochastic nature of the method."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper delivers a real but narrow advance — two new residual losses that adapt source iteration to neural-network training for multi-group neutron diffusion eigenvalue problems — and the 1D and 2D numbers are credible. The 3D claim is overreaching and the paper lacks reproducibility details, but it deserves referee time, not a desk rejection.\n\nWhat's new: Eqs. (23) and (25), the decoupling and direct-iterative losses, freeze the previous epoch's fluxes as fixed sources and update k_eff via the Rayleigh quotient, letting one network track two energy groups with no training data. Prior PC-GIPMNN did single-group; earlier two-group PINN work (ref. [33], acknowledged in their own Table 1) existed, so the honest novelty is 'multi-group without prior data, plus first 3D demo' — narrower than the abstract's blanket claim, but genuine. The ablation work is the paper's best feature: the sampling-rate and interface-condition studies are systematic and show clearly that interface constraints are load-bearing. That is a useful engineering finding.\n\nSoft spots, in proportion:\n- The biggest practical gap is reproducibility: no code, no data, single runs, no error bars. The FreeFEM reference solutions are standard benchmarks, so I don't doubt them, but nothing here can be independently checked.\n- The Rayleigh quotient in (30) is applied to a non-symmetric two-group operator without comment. The variational/minimization property of the RQ does not carry over to non-normal pencils, so as written the eigenvalue update is heuristic. The 1D/2D results suggest it behaves, but this needs a justification.\n- Algorithm 1's fixed-point iteration gets no convergence analysis. The stress-test inference that the 26% thermal-flux error in 3D IAEA shows the iteration failed to reach the fixed point is plausible but not established — that error could equally come from network capacity or the reflector region where nu*Sigma_f,2 = 0 makes the eigenfunction stiff. The paper itself concedes 3D accuracy is 'not yet satisfactory,' so the abstract's 'solves ... three-dimensional' claim outruns the evidence.\n- Minor: the duplicated digits in several tables and the typographies suggest light proofreading.\n\nBottom line: the central method is plausible and the 1D/2D evidence supports it; the 3D claims should be softened and the fixed-point/RQ gaps addressed, ideally with code. Reader and stress-test are roughly right, though the stress-test overreads the 3D error as proof of non-convergence. I'd send this to a serious referee — it's a legitimate contribution that needs revision, not rejection.","headline":"Incremental but real: clean loss-function adaptation of source iteration for multi-group eigenvalue PINNs, credible in 1D/2D, overreaching in 3D, and short on reproducibility details.","tokens_in":19545,"tokens_out":5943,"would_cite":true,"duration_ms":52450,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N99","82D75","68T07","65N25"],"pacs":[],"model":"deepseek-v4-flash","headline":"Neural networks with a decoupling residual loss solve two-group neutron diffusion eigenvalue problems in 1D, 2D, and 3D without prior data.","keywords":["multi-group neutron diffusion","eigenvalue problem","physics-informed neural networks","decoupling residual loss","direct iterative loss","Rayleigh quotient","nuclear reactor benchmarks","effective multiplication factor"],"falsifier":"Run Algorithm 1 on a two-group benchmark with a known reference solution in which the second eigenvalue is close to the fundamental; if the iteration reports a $k_{\\mathrm{eff}}$ that matches the second mode or a flux with sign changes, the central claim fails. A direct test is the 3D IAEA problem: if the thermal-flux $L^2$ relative error does not fall substantially below 26 percent when the number of residual and interface points is increased, then the claim that the method solves three-dimensional multi-group problems would be contradicted.","tokens_in":18541,"feed_emoji":"⚛️","tokens_out":5541,"duration_ms":52501,"temperature":0.7,"pith_summary":"The paper claims that multi-group neutron diffusion eigenvalue problems, which determine the effective multiplication factor $k_{\\mathrm{eff}}$ of a reactor core, can be solved by neural networks using only the equations, boundary conditions, and interface conditions, with no prior flux data. Two new residual loss functions are proposed: a decoupling loss that turns each energy-group equation into a single-unknown PDE residual and a direct iterative loss that keeps the two-group coupling while freezing one group from the previous iteration. On the Ringhals, TWIGL, and IAEA benchmarks, the reported $k_{\\mathrm{eff}}$ relative errors range from about $10^{-4}$ to $6\\times10^{-3}$, and two-dimensional flux errors are mostly below ten percent. The three-dimensional IAEA thermal flux error reaches 26 percent, which the paper acknowledges as not yet satisfactory. The central practical claim is that this is the first neural-network solution of real three-dimensional multi-group reactor problems without observational data.","feed_headline":"Neural nets solve two-group reactor benchmarks without prior data","feed_subtitle":"Decoupling residual losses plus Rayleigh-quotient updates give keff errors near 1e-4 to 6e-3 on TWIGL and IAEA problems.","key_machinery":"The load-bearing object is the decoupling residual loss function (Eq. 23): one residual per energy group, in which the current-iterate flux of that group appears on the left while all cross-group terms come from the previous iterate, making each residual a single-unknown PDE. The direct iterative loss (Eq. 25) is a second variant that keeps the two-group coupling inside the current step but freezes $\\phi_1$ from the previous step in the thermal-group equation, which is needed when $\\nu\\Sigma_{f,1}=0$. Around this, the algorithm alternates three operations: gradient descent on the total loss, power normalization of the flux, and a Rayleigh-quotient update of $\\tilde\\lambda = 1/k_{\\mathrm{eff}} + \\sigma$. The network outputs one flux block per material region, and interface losses enforce flux continuity and $D$-weighted normal-current continuity.","core_discovery":"The paper's central claim is that the two-group neutron diffusion eigenvalue problem—the coupled system whose largest eigenvalue is the reactor's effective multiplication factor—can be solved by a neural network without any prior data. The discovery is a way of organizing the residual loss so that each energy group is solved as if it were a single-unknown PDE: in the decoupling residual loss, the current iteration's flux appears only in its own group's equation, while all cross-group fluxes are taken from the previous iteration; in the direct iterative loss, the coupling is kept on the left-hand side but the thermal-group source uses the previous fast-group flux, which matters when the fast fission cross-section vanishes. Around either loss, the algorithm alternates gradient updates, power normalization of the flux, and a Rayleigh-quotient update of the eigenvalue. The paper reports that this scheme reaches $k_{\\mathrm{eff}}$ errors around $10^{-4}$ to $6\\times10^{-3}$ on one-, two-, and three-dimensional benchmarks, with interface conditions essential: without them the problem becomes unsolvable in the tested cases.","pith_inferences":["Beyond the paper: the decoupling loss is a neural translation of the traditional source-iteration sweep, so if the underlying fixed-point iteration were shown to converge, the method would inherit a classical convergence theory rather than resting on heuristics.","Beyond the paper: the shift parameter $\\sigma$ plays a role analogous to a Wielandt shift in classical eigensolvers; choosing it near the inverse of the target eigenvalue could accelerate convergence in harder geometries, a testable extension.","Beyond the paper: the per-material output block architecture will struggle as the number of regions grows; combining the residual loss with a partition-of-unity or level-set network could remove the need for explicit interface training points.","Beyond the paper: the same decoupling recipe extends naturally to $G>2$ groups by ordering the groups so that only lower-to-higher scattering appears on the right-hand side, yielding a general neural source iteration."],"forward_implications":["Multi-group neutron diffusion eigenvalue problems can be solved by neural networks using only equation, boundary, and interface information, with no prior flux data.","The decoupling and direct iterative losses outperform the previous inverse-power-method residual, especially when the fast-group fission cross-section vanishes as in the IAEA benchmark.","Enforcing interface conditions is necessary: removing them makes the tested problems unsolvable by the proposed network.","Increasing the number of residual and interface training points improves accuracy and convergence speed, though a fixed network size limits the attainable precision.","In three dimensions the method captures the main features of both TWIGL and IAEA, but the 26 percent thermal-flux error on IAEA shows that three-dimensional accuracy is not yet satisfactory."],"supporting_citations":[{"why":"Supplies the previous single-group inverse-power-method neural network and its residual loss, which the new decoupling loss extends to multi-group problems.","marker":"[29]"},{"why":"Provides the source-iteration and power-iteration framework that motivates the decoupling of the multi-group equations.","marker":"[35]"},{"why":"Supplies the one-dimensional Ringhals-4 benchmark problem and its material data.","marker":"[38]"},{"why":"Supplies the TWIGL benchmark geometry and two-group cross-section data used in the two- and three-dimensional tests.","marker":"[39]"},{"why":"Supplies the IAEA PWR benchmark, including the geometric layout and material cross-sections for the two- and three-dimensional cases.","marker":"[34]"},{"why":"Provides the FreeFEM++ reference solutions used to define the relative error metrics for fluxes and eigenvalues.","marker":"[40]"},{"why":"Provides automatic differentiation, which is used to compute all derivative terms in the residual losses.","marker":"[37]"},{"why":"Provides the power-method and inverse-power-method neural-network formulation that underlies the baseline LossIPM residual.","marker":"[28]"}],"fun_headline_variants":["Decoupling loss lets NNs solve multi-group criticality","Neural nets nail keff in multi-group benchmarks","No prior data: NN solves multi-group eigenvalue","Coupled-group diffusion solved by single NN loss","Rayleigh-quotient NNs hit 1e-4 keff errors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire method rests on the alternating training loop—gradient updates on the residual loss, a Rayleigh-quotient refresh of the eigenvalue, and power normalization—actually converging to the fundamental mode; the paper supplies no convergence analysis and no safeguard against locking onto a spurious mode.","fun_headline_variants_meta":{"raw":{"variants":["Decoupling loss lets NNs solve multi-group criticality","Neural nets nail keff in multi-group benchmarks","No prior data: NN solves multi-group eigenvalue","Coupled-group diffusion solved by single NN loss","Rayleigh-quotient NNs hit 1e-4 keff errors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000876,"raw_usage":{"total_tokens":3789,"prompt_tokens":944,"completion_tokens":2845,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":2778}},"tokens_in":560,"tokens_out":2845,"duration_ms":19693,"temperature":1.0,"reasoning_tokens":2778,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:00:38.610893+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 1 on a two-group benchmark with a known reference solution in which the second eigenvalue is close to the fundamental; if the iteration reports a $k_{\\mathrm{eff}}$ that matches the second mode or a flux with sign changes, the central claim fails. A direct test is the 3D IAEA problem: if the thermal-flux $L^2$ relative error does not fall substantially below 26 percent when the number of residual and interface points is increased, then the claim that the method solves three-dimensional multi-group problems would be contradicted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the previous single-group inverse-power-method neural network and its residual loss, which the new decoupling loss extends to multi-group problems."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the source-iteration and power-iteration framework that motivates the decoupling of the multi-group equations."},{"cited_title":"Demazi` ere, C","cited_arxiv_id":null,"evidence_quote":"Supplies the one-dimensional Ringhals-4 benchmark problem and its material data."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the TWIGL benchmark geometry and two-group cross-section data used in the two- and three-dimensional tests."},{"cited_title":"Theler, F","cited_arxiv_id":null,"evidence_quote":"Supplies the IAEA PWR benchmark, including the geometric layout and material cross-sections for the two- and three-dimensional cases."},{"cited_title":"Hecht, New development in FreeFem++, J","cited_arxiv_id":null,"evidence_quote":"Provides the FreeFEM++ reference solutions used to define the relative error metrics for fluxes and eigenvalues."},{"cited_title":"Paszke, S","cited_arxiv_id":null,"evidence_quote":"Provides automatic differentiation, which is used to compute all derivative terms in the residual losses."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the power-method and inverse-power-method neural-network formulation that underlies the baseline LossIPM residual."}],"review_version":1}