{"id":"2fe0eefe-8fef-4823-9249-2c5ce126b31e","arxiv_id":"2507.09591","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A data-free physics-informed neural network matches finite-element thermal histories of wire-arc directed energy deposition to about 7% relative L2 error, reporting up to 98.6% compute-time reduction versus a fine-mesh FEM baseline.","lead":"A neural network trained only on the heat equation, with no simulation data, reproduces the temperature history of a wire-arc 3D-printing process computed by finite elements, with a relative error of about 7 percent. The authors report up to 98.6 percent lower computation time versus a very fine finite-element mesh, which could make large-scale process simulation cheaper.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline 98.6% speed-up is only against the finest FEM baseline, while the coarsest FEM-CC runs in 1m33s and is 29x faster than PINN; without a mesh-convergence study showing FEM-CC is inaccurate, the efficiency claim is unsubstantiated.","rationale":"The paper's strongest quantitative claim is the 98.6% computational reduction. That claim depends entirely on FEM-FF being the required fidelity baseline, yet the paper offers no mesh-convergence study, no comparison of FEM-CC with FEM-FC/FF, and no engineering accuracy tolerance. Since FEM-CC is 29x faster than PINN, the efficiency superiority of PINN disappears if the coarse FEM is acceptable for the intended use. The reader identified this same weakest assumption, and my reading confirms it; the additional hardware mismatch further inflates the speed-up. A secondary internal inconsistency is that Table 1 lists v=10.0 mm/s while Figure 6 shows v=3 mm/s, which should be resolved, and no artifacts or code are released, so reproducibility is limited. However, the decisive technical gap remains the missing mesh-convergence/baseline justification, and the concrete test above would settle it. Because the reader already returned a conditional verdict requiring exactly this evidence, my stress-test does not change the verdict.","tokens_in":14711,"tokens_out":7490,"duration_ms":77591,"concrete_test":"Run an intermediate FEM mesh (e.g., 0.25mm spatial / 10ms time) plus the existing FEM-CC and FEM-FF on the same CPU cores, and compute full-domain relative L2 differences (FEM-CC vs FEM-FF, intermediate vs FEM-FF, and PINN vs FEM-FF), along with P1-P5 thermal histories. If L2(FEM-CC, FEM-FF) is at or below the reported 7.267e-2, the coarse FEM is already at least as close to the converged solution as the PINN, and the 98.6% speed-up is an artifact of an unnecessary baseline; if the coarse FEM error is much larger than the PINN error, the fine baseline may be justified. Report all timings with matched hardware.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.3/Table 4 report PINN training time 45m40s against FEM-CC (1m33s), FEM-FC (2h01m), and FEM-FF (2d06h). The 98.6% reduction is solely relative to FEM-FF. Against FEM-CC, the PINN is slower by a factor of ~29, so the central claim that PINNs outperform FEM reverses unless FEM-CC is shown to be inadequate. The paper provides no mesh-convergence study and no L2 comparison between FEM-CC and FEM-FC/FF; the only stated reason for preferring FEM-FF is that it 'most closely resembles the discretization on which the PINN was trained,' which is a resolution-matching argument, not an accuracy argument. The comparison is further skewed by hardware: FEM ran on 4 CPU cores of a laptop while PINN trained on an RTX 3070Ti GPU. Since the paper's headline efficiency conclusion depends on using the longest FEM run as the required baseline, the missing justification of that baseline is the load-bearing weakness.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a physics-informed neural network (PINN) surrogate for predicting thermal histories in wire-arc directed energy deposition (DED). The PINN is trained using only the heat equation, boundary/initial conditions, and a Goldak heat source, without external FEM or experimental data. The authors introduce Sobol' sequence-based collocation point sampling, a power-law z-warp, hard initial conditions, and self-adaptive loss weighting. They compare the trained PINN against three FEM simulations (FEM-CC, FEM-FC, FEM-FF) with increasing spatial and temporal resolution, reporting a relative L2 error of 7.267e-2 and computational time reductions of up to 98.6% relative to the finest FEM model. The paper also discusses scalability and future directions for large-scale DED simulation.","tokens_in":14966,"tokens_out":5095,"duration_ms":53453,"significance":"If the reported efficiency gains hold, the work would be a useful step toward making PINN surrogates practical for repeated thermal simulation of large-scale DED components, where FEM runs are expensive. The claimed novelty lies in data-free training combined with efficient quasi-Monte Carlo sampling. The manuscript gives a clear description of the network architecture, loss weighting, sampling strategy, and training schedule, which facilitates reproducibility. The main limitations are that the central efficiency claim is not yet adequately supported because the comparison baseline and hardware settings are not justified, and the single-run error estimate does not demonstrate robustness.","major_comments":[{"comment":"The headline efficiency claim of up to 98.6% reduction is computed against FEM-FF (2d 6h), while the coarsest FEM-CC baseline runs in 1m33s and is approximately 29 times faster than the PINN training time (45m40s). The paper's only justification for preferring FEM-FF as the comparison baseline is that it 'most closely resembles the discretization on which the PINN was trained' (Section 3.3). This is a resolution-matching argument, not an accuracy argument. The authors need to provide a mesh-convergence study, reporting L2 errors and thermal-history differences between FEM-CC, FEM-FC, and FEM-FF, and then justify which FEM resolution is sufficient for engineering accuracy. Without this, the efficiency conclusion reverses if FEM-CC is adequate.","section":"Section 3.3, Table 4"},{"comment":"The timing comparison mixes hardware and workflows: PINN training was performed on an NVIDIA RTX 3070Ti GPU, while FEM runs used four CPU cores of a laptop. The PINN time is a one-time training cost, and the paper does not quantify inference time for new queries, storage cost, or the break-even number of repeated simulations. For a single thermal-history prediction, the FEM-CC run is already faster than PINN training. To make the efficiency claim meaningful, the authors should specify the total cost model (training plus inference) and state for how many parameter variations or repeated simulations the PINN becomes cheaper.","section":"Section 3.3, Table 4"},{"comment":"The reported relative L2 error of 7.267e-2 is a single value with no information about variation across random seeds, network initializations, or Sobol' sequence draws, and no spatial or temporal decomposition of the error. Given the stochastic Adam optimizer and the sensitivity of PINN training, one training run does not establish the robustness of the error. The manuscript should report the mean and standard deviation over at least several independent training runs, together with the precise definition of the error norm over the spatiotemporal domain.","section":"Section 3.3, L2 error"},{"comment":"The 'super-resolution' property is invoked as an advantage, but it is not demonstrated. The PINN was trained with time discretized at 5 ms and the FEM-FF baseline also uses 5 ms temporal and 0.1 mm spatial resolution, so the comparison does not show prediction beyond the discretization of the reference. To substantiate the super-resolution claim, the authors should evaluate the trained PINN at spatial and temporal resolutions finer than any FEM run (for example, 2.5 ms or 0.05 mm) and compare against an appropriate refined reference or interpolation, or clearly define what 'super-resolution' means in this context.","section":"Section 3.3, super-resolution"}],"minor_comments":[{"comment":"The unit for power P is given as 'g mm/s2', which is not dimensionally consistent with the Goldak source term in Eq. (2). Please verify the unit and use standard SI or consistent CGS units throughout.","section":"Table 1"},{"comment":"The summation index i and the sample count N are not defined for the gradient-based weight update; please specify that N corresponds to the number of collocation points in the respective loss term.","section":"Section 2.2, Eq. (10)"},{"comment":"The phrase 'collapsed time dimension' is unclear; please add a sentence explaining that the points are projected onto the spatial domain.","section":"Figure 7 caption"},{"comment":"The statement that Sobol' sampling gives 'approximately half of the necessary computation time per epoch' lacks quantitative support; report the measured epoch times for uniform and Sobol' sampling.","section":"Section 3.2.1"},{"comment":"In the Conclusion, the reduction for FEM-FC is stated as 62.5%, whereas Table 4 implies 62.4%; please make the numbers consistent.","section":"Section 4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the scope of cs.CE and the central idea is potentially useful. The efficiency claim, however, is the key selling point and is currently not properly justified. I recommend that the authors perform a mesh-convergence study and re-benchmark on comparable hardware, or at least clearly state the limitations of the comparison. Making the training code available would also improve reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real contribution here is a PINN trained without any FEM data in the loss, using Sobol-sequence collocation points for a wire-arc DED thermal problem, with a quantitative comparison against FEM. That is worth knowing. The paper is honest about being a groundwork study and it describes the network and training choices in enough detail to reproduce.\n\nThe efficiency claim, though, does not survive contact with Table 4. The 98.6% reduction is only against FEM-FF, the finest mesh (2d6h). Against FEM-CC, the coarsest FEM (1m33s), the PINN training time of 45m40s is about 29x slower. The paper justifies FEM-FF only as “most closely resembling the discretization on which the PINN was trained,” which is a resolution-matching argument, not an accuracy argument. Without a mesh-convergence study showing FEM-CC is inadequate, the headline speed-up is an artifact of the chosen baseline. The hardware mismatch (GPU for PINN, laptop CPU for FEM) makes the comparison even less clean.\n\nOther soft spots: the L2 error of 7.267e-2 is a single number with no seed variation or error bars; no code or data are released; and there is an internal inconsistency in the velocity (Section 3.1 text says v=3 mm/s, while Table 1 lists v=10.0 mm/s). These are fixable, but they need to be addressed.\n\nWhat the paper does well: it clearly isolates the PDE-only setting, applies hard initial conditions, uses the Goldak source, and compares against three FEM resolutions. The FEM model was validated experimentally in the authors' prior work [10], which gives some credibility to using it as a benchmark. Also, because both PINN and FEM solve the same heat equation, agreement is expected if both converge; the informative claim is speed, not independent physical validation, and the paper mostly frames it that way.\n\nWho is this for? PINN practitioners working on additive manufacturing and structural-engineering folks who want a concrete benchmark. It is not a breakthrough, but it is a solid application study with a clear methodology. I would send it to peer review, but with major revision: add a mesh-convergence study, a matched-hardware or properly caveated timing comparison, error bars across seeds, resolve the velocity inconsistency, and ideally release code and data. Then the efficiency claim can be evaluated on its merits.","headline":"A useful data-free PINN benchmark for wire-arc DED, but the 98.6% speed-up is an artifact of comparing against the finest FEM mesh; the paper needs a mesh-convergence study and matched-baseline timing before the efficiency claim can be taken at face value.","tokens_in":15530,"tokens_out":2043,"would_cite":false,"duration_ms":25578,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A physics-informed neural network trained solely on the heat equation and boundary conditions reproduces FEM thermal histories for wire-arc DED with a relative L2 error of 7.267e-2 and up to 98.6% reduction in reported compute time.","keywords":["physics-informed neural networks","wire-arc directed energy deposition","thermal history simulation","Sobol' sequences","collocation point sampling","data-free surrogate","Goldak heat source","finite element method benchmark"],"falsifier":"Run a mesh- and time-step-convergence study on the same 40 mm × 6 mm × 4 mm block, comparing peak temperatures at the five probe points across FEM-CC, FEM-FC, and FEM-FF; if FEM-CC agrees with FEM-FF within engineering tolerance, then the 98.6% speed-up over FEM-FF is not the relevant comparison and the PINN is slower than FEM-CC. Alternatively, compare PINN predictions to thermocouple measurements on a deposited wall: a relative error much larger than 7.3% would show the PDE-only training misses physics that the FEM benchmark already approximates.","tokens_in":14515,"feed_emoji":"🔥","tokens_out":9266,"duration_ms":95781,"temperature":0.7,"pith_summary":"The paper tries to show that a physics-informed neural network (PINN), trained with no experimental or FEM data, can serve as a fast surrogate for thermal-history simulation in wire-arc directed energy deposition. On a single-layer 40 mm × 6 mm × 4 mm block, the network reproduces the temperature field of a fine finite-element benchmark with relative L2 error $7.267\\times10^{-2}$ while cutting the reported compute time by up to 98.6%. The key enabler is a sampling strategy based on Sobol' sequences, which covers the space-time domain with far fewer collocation points than uniform grids. Because the trained network can be evaluated at any point in space and time, it offers \\u201csuper-resolution\\u201d that FEM would need a finer mesh to match. If the result holds at larger scales, repeated thermal simulations for process planning could move from days to under an hour.","feed_headline":"Data-free PINN cuts wire-arc DED thermal sim time by 98.6 percent","feed_subtitle":"A physics-informed network matches fine FEM temperature fields at 7.3 percent relative error and adds super-resolution.","key_machinery":"The carrying mechanism is a feed-forward MLP that outputs temperature $\\hat{u}(x,y,z,t)$ and is trained by minimizing normalized residuals of the heat-conduction PDE (with the Goldak double-ellipsoid moving heat source $Q_{\\text{goldak}}$), the Robin convection/radiation boundary condition, and the Dirichlet bottom condition, all computed by automatic differentiation. The Goldak source is a Gaussian-shaped, double-ellipsoidal volumetric heat flux that is standard for welding and DED. Three design choices carry the scalability argument: hard enforcement of the initial condition, which removes one loss term; a self-adaptive weighting scheme that balances boundary and PDE residuals during training; and Sobol' low-discrepancy quasi-random collocation points, which cover the four-dimensional space-time domain with far fewer samples than uniform grids and change from time step to time step. The Sobol' sequence, a low-discrepancy point set designed to cover a unit hypercube uniformly, is the central scalability component: it keeps the collocation-point count and per-epoch cost low as domain size and, later, parameter dimensionality grow.","core_discovery":"The central claim is that a fully connected feed-forward neural network, constrained only by the transient heat-conduction equation with a Goldak double-ellipsoid volumetric heat source and by convection/radiation (Robin) plus Dirichlet boundary conditions, can learn the thermal history of a wire-arc DED deposition without any external training data. With hard enforcement of the uniform ambient initial condition, self-adaptive loss weighting, Sobol'-sampled collocation points, and an Adam-then-L-BFGS schedule, the trained network matches the finest FEM benchmark (0.1 mm mesh, 5 ms steps) to a relative L2 error of $7.267\\times10^{-2}$ and reproduces peak temperatures along the heat-source path to an average difference of about 85 \\u00b0C (6.5%). The authors conclude that a strictly PDE-driven PINN is a justifiable substitute for FEM when fine-resolution thermal histories are needed, and that folding even one FEM run into training would erase the computational advantage for a single use case.","pith_inferences":["If a mesh-convergence study shows the coarse FEM already captures the thermal history, the honest comparison is PINN versus FEM-CC; then PINN wins only when the same network is reused for many different queries or process parameters, not for a single thermal history.","The 'super-resolution' property could be exploited the other way: instead of matching a fine FEM mesh, use the PINN to interpolate sparse thermocouple data onto any spatial-temporal grid, effectively turning the surrogate into a measurement-enhancement tool.","The single-layer, constant-geometry test leaves open how the method behaves with layer-by-layer deposition, where the heat source moves over newly added material and boundary conditions change; a natural next check is a two- or three-layer wall with the same Sobol' sampling and a comparison to FEM or experiment.","The 45-minute training cost is paid once; amortized over many simulations, even a slower-than-coarse-FEM training time can be economically justified, an argument the paper makes only implicitly through its parametric-PINN outlook."],"forward_implications":["For applications that genuinely require FEM-FF-level resolution (0.1 mm mesh, 5 ms steps), a single PINN training run replaces a multi-day FEM run, and additional queries at new points or times cost only a forward pass.","Because the PINN is trained from the governing equations alone, it sidesteps the shortage of large experimental datasets for wire-arc DED; the same workflow could be applied to a new geometry or material by redefining the residual and initial/boundary conditions.","Repeated what-if simulations for heat management, process-parameter choice, or printing strategy become feasible: the network can be evaluated thousands of times at negligible marginal cost.","Sobol' collocation reduces per-epoch cost and sample count relative to uniform multi-grid sampling, and the advantage grows with domain size and parameterized inputs, pointing toward parametric PINNs trained over ranges of process parameters."],"supporting_citations":[{"why":"Validates the FEM model against experiments in the authors' prior work, making it the benchmark the PINN is compared to.","marker":"[10]"},{"why":"Establishes the hybrid PINN thermal-modeling approach for additive manufacturing and the claim that auxiliary labeled data can accelerate training, the data-free alternative this paper explores.","marker":"[23]"},{"why":"Supports the choice of the Goldak double-ellipsoid heat source as most representative for wire-arc DED.","marker":"[31]"},{"why":"Defines the Goldak double-ellipsoidal volumetric heat flux used in the PDE residual.","marker":"[32]"},{"why":"Supplies the self-adaptive loss-weighting scheme that balances boundary and PDE residuals during training.","marker":"[33]"},{"why":"Provides the hard initial-condition enforcement method that removes the initial-condition loss term.","marker":"[39]"},{"why":"Supplies the Goldak heat-source parameters (power, efficiency, ellipsoid dimensions, distribution factors) used in both FEM and PINN.","marker":"[42]"},{"why":"Supplies the convection and radiation boundary-condition parameters used in the models.","marker":"[43]"},{"why":"Defines Sobol' sequences, the low-discrepancy sampling used for collocation points.","marker":"[46]"},{"why":"Gives the error-scaling argument showing why low-discrepancy sampling needs fewer points than uniform sampling.","marker":"[49]"}],"fun_headline_variants":["No-data PINN matches fine FEM with 7.3% error for wire-arc DED","PINN surrogate cuts wire-arc DED thermal sim cost by 98.6%","Physics-informed net simulates wire-arc DED heat in 1.4% of FEM time","Data-free neural net predicts DED thermal histories 70x faster","Physics-informed net adds super-resolution to wire-arc DED thermal sim"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the finest FEM model (0.1 mm mesh, 5 ms time steps, about two days on a laptop CPU) is the required fidelity baseline, so the 98.6% speed-up is the right comparison; the paper also concedes in Section 3.4 that a strictly PDE-driven approach has inherent accuracy limits, and if the coarser 0.5 mm / 20 ms FEM (about 1.5 minutes) is already accurate enough, the PINN's 45-minute training is slower than FEM, a possibility no mesh-convergence study rules out.","fun_headline_variants_meta":{"raw":{"variants":["No-data PINN matches fine FEM with 7.3% error for wire-arc DED","PINN surrogate cuts wire-arc DED thermal sim cost by 98.6%","Physics-informed net simulates wire-arc DED heat in 1.4% of FEM time","Data-free neural net predicts DED thermal histories 70x faster","Physics-informed net adds super-resolution to wire-arc DED thermal sim"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001026,"raw_usage":{"total_tokens":4378,"prompt_tokens":1051,"completion_tokens":3327,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":667,"completion_tokens_details":{"reasoning_tokens":3216}},"tokens_in":667,"tokens_out":3327,"duration_ms":24341,"temperature":1.0,"reasoning_tokens":3216,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:52:14.388552+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a mesh- and time-step-convergence study on the same 40 mm × 6 mm × 4 mm block, comparing peak temperatures at the five probe points across FEM-CC, FEM-FC, and FEM-FF; if FEM-CC agrees with FEM-FF within engineering tolerance, then the 98.6% speed-up over FEM-FF is not the relevant comparison and the PINN is slower than FEM-CC. Alternatively, compare PINN predictions to thermocouple measurements on a deposited wall: a relative error much larger than 7.3% would show the PDE-only training misses physics that the FEM benchmark already approximates.","supporting_citations":[{"cited_title":"Dahaghin, M","cited_arxiv_id":null,"evidence_quote":"Validates the FEM model against experiments in the authors' prior work, making it the benchmark the PINN is compared to."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the hybrid PINN thermal-modeling approach for additive manufacturing and the claim that auxiliary labeled data can accelerate training, the data-free alternative this paper explores."},{"cited_title":"Exact Enforcement of Temporal Continuity in Sequential Physics-Informed Neural Networks","cited_arxiv_id":"2403.03223","evidence_quote":"Provides the hard initial-condition enforcement method that removes the initial-condition loss term."},{"cited_title":"Liang, Q","cited_arxiv_id":null,"evidence_quote":"Supplies the Goldak heat-source parameters (power, efficiency, ellipsoid dimensions, distribution factors) used in both FEM and PINN."}],"review_version":1}