{"id":"39123383-c4b2-4016-9a16-42715b4aa286","arxiv_id":"2509.05799","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A hybrid model that replaces the EM solver in an HPM breakdown simulation with a Fourier Neural Operator reproduces streamer evolution and runs about 60x faster than the standard FDTD solver.","lead":"This paper combines a Fourier Neural Operator with a plasma fluid solver to speed up simulations of high-power microwave breakdown, replacing the electromagnetic field solver while keeping the plasma equations intact. It reports about 60x speedup and close agreement with the standard FDTD method on unseen input amplitudes.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Closed-loop error accumulation is unaddressed: per-cycle APE≈12% in a nonlinear ionization feedback loop can amplify over hundreds of cycles, so the claimed 'almost perfect overlap' is not yet established.","rationale":"The reader's conditional verdict is appropriate. The paper's central claim is that the FNO surrogate can replace the EM solver inside the coupled plasma loop and still reproduce streamer dynamics. The most load-bearing requirement is that per-cycle surrogate errors are not amplified by the plasma update. The paper provides no evidence on this: it reports one-step accuracy on FDTD-generated test inputs and shows final trajectory overlays, but no closed-loop error metric or stability analysis. Because ionization depends exponentially on the local effective field, a 12% error near the streamer tip can change the local ionization rate substantially, and because the updated density changes the next field prediction, errors can compound. The absence of such an analysis is therefore a genuine gap. I do not think this invalidates the contribution; conditional acceptance with a request for an error-growth experiment is appropriate. Secondary concerns—normalization statistics computed over train+test, ambiguity about whether scattered or total E_rms is fed to the plasma solver, and the fairness of the CPU-vs-GPU speedup comparison—are noteworthy but less decisive; a clean closed-loop error-growth check would also require reporting these implementation details.","tokens_in":17621,"tokens_out":8828,"duration_ms":111384,"concrete_test":"For the test case E0 = 2.75 MV/m, instrument a full hybrid run. Every 10 plasma cycles, freeze the hybrid density n_e^H and compute the reference E_rms^ref by running the FDTD EM update for one wave period with that density (using the same code as the training-data generator). Compute the per-cycle input-conditional APE: ||FNO(n_e^H) - E_rms^ref|| / ||E_rms^ref||, and also record the 2-norm density error ||n_e^H - n_e^FDTD|| / ||n_e^FDTD|| over time, where n_e^FDTD is the reference trajectory. If the input-conditional APE stays within the 12% seen on FDTD test inputs and the density error does not grow monotonically or exponentially over the full run, closed-loop accumulation is benign. If input-conditional APE grows with cycle number, the FNO is being driven out of its training distribution and the 'almost perfect overlap' claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.2 validates the hybrid model by overlaying streamer length, growth rate, and tip-field curves against the FDTD reference for five E0 values. This does not establish closed-loop fidelity because the FNO is not evaluated on the density distribution it actually encounters inside the loop. Each plasma cycle uses the previous FNO output to update n_e via the nonlinear, ionization-dominated continuity equation (Eq. 6); ionization frequency depends exponentially on the local effective field, so a per-cycle field error of APE≈12% can be rectified and compounded. The paper itself warns in Section 5 that 'even a small discontinuity or mismatch in prediction of E_rms could lead to uncontrolled plasma streamer growth over long durations,' yet provides no quantitative error-growth analysis: no per-cycle error measured along the hybrid trajectory, no perturbed-initial-condition experiments, no amplification/Lyapunov estimate, and no uncertainty bars on the Figure 11/12 curves. Consequently, the central claim that the hybrid 'reproduces streamer shape, velocity, and temporal evolution' with near-perfect overlap is not yet supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a hybrid simulation framework for high-power microwave (HPM) breakdown in which a Fourier Neural Operator (FNO) replaces the FDTD-based electromagnetic solver inside an established EM-plasma fluid model. The FNO maps a two-channel input (plasma density and incident electric field) to the scattered RMS electric field; the plasma continuity solver and the Python/C coupling are retained. The FNO is trained on data generated by the same FDTD-fluid solver, and the hybrid model is tested on incident amplitudes (2.55, 2.65, 2.75, 2.85, 2.95 MV/m) not used in training. The authors report FNO stand-alone metrics (MSE 0.0394, APE 0.12, SSIM 0.9999), near-perfect overlap of streamer length, growth rate, and tip field between the hybrid and full FDTD-fluid models, and a speedup of roughly 57-63x. The manuscript also documents a modular Python-C implementation that may facilitate adoption of ML surrogates in legacy plasma codes.","tokens_in":17888,"tokens_out":3877,"duration_ms":45660,"significance":"If the central claims survive scrutiny, the paper demonstrates a practical and transferable strategy for accelerating multiscale EM-plasma simulations: the FNO surrogate is coupled in a closed loop with a conventional plasma-fluid solver, achieving an order-of-magnitude speedup while closely matching the reference solution. The work's strengths include the use of an established FDTD plasma-fluid model for data generation, a non-trivial closed-loop coupling, and a concrete runtime comparison (Table 2). The reported speedup is attractive for parametric studies and for extending simulations to larger domains. However, the validation is limited in important ways: the FNO is trained on and compared against the same numerical solver, the test amplitudes are interpolated within the training range, the normalization uses test-set statistics, and the closed-loop error accumulation is not quantified. These issues do not invalidate the framework but materially weaken the current evidence for 'near-perfect' fidelity and broad generalization.","major_comments":[{"comment":"The min-max normalization is performed using global minimum and maximum values computed over the entire training and testing sets, for both the input channels and the output Erms. This is a data-leakage issue: the test-set statistics are used during training preprocessing, which can artificially improve the reported MSE/APE/SSIM. Since the incident-field channel encodes E0 directly and the test E0 values lie between training values, the benefit may be substantial. The authors should re-normalize using training-set statistics only and report the resulting metrics.","section":"Section 4.1"},{"comment":"The five 'unseen' incident amplitudes, 2.55, 2.65, 2.75, 2.85, and 2.95 MV/m, are exactly the midpoints between the training amplitudes (2.5, 2.6, ..., 3.0 MV/m). This is an interpolation test, not a test of generalization to 'any electric field within the specified range' or to new physical regimes. Because the FNO input includes the incident field amplitude, the model can exploit the local linearity of the training sweep. The claims of generalization should be restricted to interpolation within the tested range, and the authors should include an extrapolation case (e.g., E0=3.1 MV/m) or a test with a different frequency/domain to support broader claims.","section":"Section 4.1 / Table 1"},{"comment":"The closed-loop stability of the FNO-fluid coupling is not quantitatively established. The manuscript itself warns that 'even a small discontinuity or mismatch in prediction of E_rms could lead to uncontrolled plasma streamer growth over long durations' (Section 5). The standalone APE is about 12%, and the plasma update (Eq. 6) is nonlinear and ionization-dominated; errors in the predicted Erms can be amplified or rectified over hundreds of update cycles. The current validation overlays streamer length, growth rate, and tip-field curves but provides no per-cycle error measured along the actual hybrid trajectory, no perturbed-initial-condition experiments, and no error-growth bound. Without such an analysis, the statement that hybrid predictions show 'almost perfect overlap' is not yet supported. At minimum, the authors should report error bars or quantiles for the curves in Figures 11-12","section":"Section 5, Section 5.2"},{"comment":"It is ambiguous whether the FNO predicts the scattered Erms or the total Erms. The text in Section 3 says the FNO 'predict the scattered electric field (E_rms)', while Section 5.1 compares 'FNO predicted and actual FDTD generated E_rms data'. In the coupled model, the plasma solver needs the total (or effective) field to compute ionization and transport coefficients, but if the FNO outputs only the scattered field, the incident contribution must be added analytically. The paper does not clearly state which quantity is used in Eq. (6). This is essential for interpreting the reported MSE/APE and for reproducing the hybrid implementation. Please clarify the output definition and how the total Erms is assembled in the loop.","section":"Sections 3 and 5.1"}],"minor_comments":[{"comment":"Typos: 'sceince' in the abstract, 'suare' in Section 2.2, 'therfore' in Section 2.1. Please proofread.","section":"Abstract / throughout"},{"comment":"The speedup numbers are inconsistent: Section 5.2 reports 57.10-62.77x, the abstract says 'of the order of 60X', and the conclusion says 'between 55x to 60x'. The Table 2 values should be cited consistently.","section":"Section 5.2 / Table 2 / Conclusion"},{"comment":"The text refers to 'Figure 2 (a)' when discussing plasma density evolution and streamer dynamics, but the relevant panels appear to be in Figure 3. Please correct the cross-references.","section":"Section 2.2"},{"comment":"The description of the Fourier layer is inconsistent: it says only the lowest m=16 modes are retained, but Eq. (8) says 'leaving higher modes unchanged'. Please clarify whether higher modes are set to zero, retained, or processed differently.","section":"Section 4.2.1"},{"comment":"All metrics (MSE, APE, SSIM) are computed on the min-max normalized, 0-255-rescaled images. Reporting values on this scale, rather than in physical units (V/m and m^-3), makes the errors harder to interpret. Please report at least one error metric in physical units.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"I see no signs of deliberate misrepresentation; the paper appears to be a genuine engineering integration effort. The main concerns are methodological (train/test normalization leakage, interpolated test points) and the absence of closed-loop error-growth analysis. These are fixable within the scope of the manuscript, but they need to be addressed before the central fidelity claim can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere is my take on 2509.05799. The paper is best read as a proof-of-concept: it replaces the EM solve inside a coupled EM-plasma fluid model for HPM breakdown with an FNO, keeps the C-based plasma density solver, and reports a measured 60x speedup on 2D test cases. Several results are genuinely useful. The closed-loop integration of Python-driven FNO with ctypes-linked C routines is nontrivial, and the paper is honest that the EM solver dominates (>99%) the runtime and that speedup will vary with problem size. The qualitative streamer agreement in Figures 11 and 12 is encouraging.\n\nNow the soft spots. The min-max normalization uses global min/max computed over training and test sets together (Section 4.1). That is a data leak; it does not invalidate the approach, but the reported APE/SSIM are over-optimistic and should be recomputed with test-set-independent scaling. The test amplitudes (2.55, 2.65, ..., 2.95) are all interpolated between training values, so the generalization claim is narrower than stated. More importantly, the paper never analyzes closed-loop error accumulation. The authors themselves warn in Section 5 that even small E_rms mismatches can drive uncontrolled streamer growth, yet they provide no per-cycle error along the hybrid trajectory, no perturbed-initial-condition runs, and no error bars on the streamer length/velocity curves. Because ionization depends exponentially on the effective field, a per-cycle APE of 12% could in principle amplify; the current validation does not rule that out. Also, the text is ambiguous about whether the FNO output is the scattered field or the total RMS field, and the hybrid loop description says the FNO 'predicts E_rms' without specifying how the incident field is added. That needs clarification.\n\nThese issues are real but not fatal. The central argument that a neural-operator surrogate can be embedded in an existing discharge solver and accelerate parameter sweeps is plausible and supported by the evidence. This is a computational acceleration paper, not a new physics discovery, and the authors mostly frame it that way. I would send it to peer review, but request revisions: fix the normalization, clarify the field convention, and add at least one closed-loop error diagnostic (e.g., per-cycle field/density error at a test amplitude, or an ensemble with different FNO training seeds). Code/data release would substantially raise confidence.\n\nFor a reading group on surrogate modeling for plasma, it is worth a look; for a general plasma physics audience, it is niche. I would not cite it in my current work, but I would engage with it in review.","headline":"A useful proof-of-concept for embedding an FNO surrogate in a coupled EM-plasma fluid solver, with a credible 60x speedup and good qualitative streamer agreement, but the paper overstates its closed-loop fidelity and has a normalization leakage that should be fixed.","tokens_in":18359,"tokens_out":3885,"would_cite":false,"duration_ms":43401,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["52.65.-y","52.80.Pi"],"model":"deepseek-v4-flash","headline":"A neural network that replaces the electromagnetic field solver reproduces microwave streamer growth and runs roughly 60 times faster than the full FDTD-plasma fluid model.","keywords":["high-power microwave breakdown","microwave streamer","plasma fluid model","Fourier neural operator","FDTD","hybrid simulation","EM–plasma interaction","machine learning surrogate"],"falsifier":"Run the closed-loop hybrid and the full FDTD-fluid model on the same unseen case, say E0 = 2.75 MV/m, but continue past the paper's 80%-of-domain termination condition; if streamer length or tip field diverges from the reference after the training horizon, the no-error-accumulation assumption fails.","tokens_in":17530,"feed_emoji":"⚡","tokens_out":8239,"duration_ms":85675,"temperature":0.7,"pith_summary":"High-power microwave breakdown simulations couple Maxwell's equations to a plasma continuity equation, and the electromagnetic update consumes more than 99 percent of the runtime, making large or long simulations impractical. This paper claims that a Fourier Neural Operator can learn the mapping from plasma density plus incident wave to the time-averaged scattered electric field, and that plugging this surrogate into the coupled solver—while keeping the plasma density update physics-based—reproduces microwave streamer formation for incident field amplitudes not seen in training. The reported result is close overlap with the full FDTD-fluid model in streamer shape, growth rate, tip field, and temporal evolution, at 57–63x speedup that turns multi-day runs into sub-hour runs. The paper also shows how an existing compiled C solver can be wrapped and driven from Python to host such a surrogate. If the claim holds, multiscale HPM breakdown studies that currently require months of serial computation become feasible on ordinary hardware.","feed_headline":"Neural EM surrogate matches streamer growth at 60x speed","feed_subtitle":"Fourier neural operator replaces the costly Maxwell solve while the plasma solver stays physics-based, cutting multi-day runs to minutes.","key_machinery":"The load-bearing component is a Fourier Neural Operator used as an image-to-image surrogate: the input is a two-channel image of normalized plasma density and incident electric field, and the output is the scattered RMS electric field. Four Fourier layers, each retaining 16 low-frequency modes, learn the global EM-plasma coupling in the spectral domain, while a pointwise convolution bias captures localized streamer-tip features. In the closed loop, the surrogate replaces the per-wave-cycle FDTD EM update; the C-based plasma solver advances electron density once per EM wave cycle using transport coefficients computed from the predicted RMS field.","core_discovery":"The paper's central claim is that the expensive finite-difference time-domain electromagnetic solver inside a coupled EM-plasma fluid model for high-power microwave breakdown can be replaced by a Fourier Neural Operator surrogate without losing the physics of streamer formation. Trained on scattered-field snapshots from an in-house FDTD-plasma fluid solver, the surrogate maps a two-channel snapshot of plasma density and incident field to the RMS scattered electric field. When cycled against the unchanged plasma continuity solver, the hybrid reproduces streamer length, streamer growth rate, and tip electric field enhancement for unseen incident fields of 2.55, 2.65, 2.75, 2.85, and 2.95 MV/m,","pith_inferences":["The 60x figure is tied to the EM update consuming more than 99 percent of runtime; a problem where other physics takes a larger share would show a smaller speedup.","Training and testing cover only 2.5–3.0 MV/m at 110 GHz and 760 Torr, so behavior outside that interval of field amplitude, pressure, or frequency remains an open question.","At a reported per-cycle average percent error near 12 percent, the long-horizon reliability of the closed loop depends on error cancellation; measuring the growth rate of streamer-length error over many cycles would make that assumption quantitative.","Because the surrogate predicts time-averaged RMS fields, sub-cycle electromagnetic dynamics that matter for gas heating or higher-harmonic effects are outside what this architecture captures."],"forward_implications":["A 10λ × 10λ domain that the paper estimates at roughly 500 days with the reference FDTD-fluid model would become a few days of computation at the measured speedup.","Parameter sweeps over incident field amplitude inside the trained 2.5–3.0 MV/m range could be run quickly, making breakdown-threshold and streamer-dynamics studies practical.","The hybrid retains the plasma continuity equation solver, so the acceleration comes without replacing the fluid description of plasma evolution.","The demonstrated Python–shared-library–C interface offers a template for retrofitting other legacy compiled EM-plasma codes with machine-learning surrogates."],"supporting_citations":[{"why":"supplies the EM-plasma fluid physical model, effective-field approximation, and transport coefficients used to generate training data","marker":"[10]"},{"why":"provides the direct-integration discretization used for electron velocity and field updates in the reference solver","marker":"[11]"},{"why":"establishes the FDTD-fluid solver as reproducing microwave streamer and filament observations for validation against","marker":"[9]"},{"why":"defines the Fourier Neural Operator architecture that the surrogate is built on","marker":"[36]"},{"why":"supplies the scattered-field FDTD formulation that the FNO surrogate replaces","marker":"[34]"},{"why":"quantifies the domain-scaling cost for large HPM breakdown runs that motivates the speedup","marker":"[22]"},{"why":"supplies the absorbing boundary conditions used by the reference electromagnetic solver","marker":"[21]"},{"why":"describes the original monolithic C plasma solver whose restructuring into a Python-C hybrid is demonstrated","marker":"[8]"}],"fun_headline_variants":["FNO-based EM surrogate speeds HPM breakdown sims 60x","Hybrid FNO-plasma model matches streamers, runs 60x faster","Neural EM operator cuts HPM streamer sim time by 60x","Fourier neural operator replaces Maxwell solver in plasma model: 60x faster","Microwave breakdown streamer sims hit 60x speedup with FNO hybrid"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"That the roughly 12 percent average error in each predicted field snapshot does not compound through the nonlinear field–plasma feedback into uncontrolled streamer growth over the hundreds of update cycles of a full simulation.","fun_headline_variants_meta":{"raw":{"variants":["FNO-based EM surrogate speeds HPM breakdown sims 60x","Hybrid FNO-plasma model matches streamers, runs 60x faster","Neural EM operator cuts HPM streamer sim time by 60x","Fourier neural operator replaces Maxwell solver in plasma model: 60x faster","Microwave breakdown streamer sims hit 60x speedup with FNO hybrid"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000737,"raw_usage":{"total_tokens":3149,"prompt_tokens":784,"completion_tokens":2365,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":2261}},"tokens_in":528,"tokens_out":2365,"duration_ms":19168,"temperature":1.0,"reasoning_tokens":2261,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T04:57:12.927937+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the closed-loop hybrid and the full FDTD-fluid model on the same unseen case, say E0 = 2.75 MV/m, but continue past the paper's 80%-of-domain termination condition; if streamer length or tip field diverges from the reference after the training horizon, the no-error-accumulation assumption fails.","supporting_citations":[{"cited_title":"Chaudhury, J","cited_arxiv_id":null,"evidence_quote":"supplies the EM-plasma fluid physical model, effective-field approximation, and transport coefficients used to generate training data"},{"cited_title":"Chaudhury and J-P Boeuf","cited_arxiv_id":null,"evidence_quote":"provides the direct-integration discretization used for electron velocity and field updates in the reference solver"},{"cited_title":"Chaudhury, J","cited_arxiv_id":null,"evidence_quote":"establishes the FDTD-fluid solver as reproducing microwave streamer and filament observations for validation against"},{"cited_title":"Kunz and R.J","cited_arxiv_id":null,"evidence_quote":"supplies the scattered-field FDTD formulation that the FNO surrogate replaces"},{"cited_title":"Ghosh and B","cited_arxiv_id":null,"evidence_quote":"quantifies the domain-scaling cost for large HPM breakdown runs that motivates the speedup"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the absorbing boundary conditions used by the reference electromagnetic solver"},{"cited_title":"Chaudhury, A","cited_arxiv_id":null,"evidence_quote":"describes the original monolithic C plasma solver whose restructuring into a Python-C hybrid is demonstrated"}],"review_version":1}