{"id":"22ae68a9-86d6-4a5f-b96c-885123602fa4","arxiv_id":"2502.01820","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A sequential PINN predicts multi-track LPBF temperatures with about 2.5 percent mean error and trains 8.5 times faster than a PI-DeepONet alternative.","lead":"This paper compares two neural network strategies for predicting 3D temperature fields in laser powder bed fusion, including multiple laser tracks. A sequential physics-informed network beat a single operator network in accuracy and training time for multi-track cases, but the authors validate only against their own finite-difference solver.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The unvalidated finite-difference reference in Section 2.1 undermines the reported accuracy metrics; a mesh-convergence against an independent solution should precede acceptance.","rationale":"The reader's weakest_assumption correctly identifies the unvalidated finite-difference reference as the most load-bearing assumption. I agree: if the FD solver is not converged or is physically inaccurate, then the central claim—that sequential PINN maintains MAPE below 2.5% and melt-pool errors below 5% for all six paths—is unsupported, because those numbers are computed relative to that same solver. I considered an alternative concern: the conclusion calls the method 'scalable,' but Eq. (20) gives N_train = sum_{i=0}^{n-1} product_{j=0}^{i} (n-j), which grows factorially with the number of tracks (15 for n=3, 64 for n=4, 325 for n=5). That is a real overclaim for extrapolation beyond the demonstrated three-track case, and it is worth noting in the revision. However, it does not invalidate the reported n=3 measurements, whereas an inaccurate FD reference would invalidate every quantitative result. Thus the FD-validation concern is more load-bearing. The paper's own conclusion acknowledges possible error accumulation over multiple tracks but does not address the baseline validation gap. A mesh-convergence and independent-solution check is a concrete, inexpensive way to settle the concern. If it passes, the empirical comparison becomes credible; if it fails, the central claim is not established. The verdict should therefore remain CONDITIONAL, with this check as a condition for full acceptance.","tokens_in":19102,"tokens_out":6036,"duration_ms":62824,"concrete_test":"Run a convergence study for the FD solver on one single-track case and one three-track case: reduce the spatial step by factors of 2 (e.g., 0.05, 0.025, 0.0125 mm) while adjusting the time step to keep the CFL number fixed, and compare peak temperature and melt-pool length/width/depth between successive grids. If the coarsest and finest predictions differ by more than about 5%, the reference is not converged and the reported MAPE values cannot be interpreted as accuracy statements. In addition, compare the finest FD single-track solution against the Rosenthal analytical solution or an independent FEM solver; if that check fails, the reference itself is suspect and the surrogate comparison must be redone.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Every quantitative claim in the paper—MAPE 2.49%, melt-pool errors below 5%, and the 8.5x training-time advantage—is measured against the finite-difference solver of Section 2.1, Eq. (5). That solver is never validated. No spatial or temporal discretization sizes are reported, no grid-refinement study is shown, and no comparison is made to an analytic solution, a high-fidelity FEM result, or experimental thermocouple/IR data. This matters specifically because Eq. (5) is an explicit Euler scheme with a strongly temperature-dependent conductivity and an apparent-heat-capacity treatment of latent heat (Eq. 21). Such schemes have strict stability limits and can introduce artificial diffusion or phase lag near the moving Gaussian heat source, exactly where the melt-pool dimensions are measured. If the reference under-resolves the peak temperature or the melt-pool boundary, the reported 'errors' of the surrogates are not errors with respect to the true physics; they are errors with respect to a possibly biased numerical baseline. The paper itself only states that FD solutions are used 'as a baseline' (Section 2.1), and no code is provided to reproduce the reference. The central claim therefore rests on an unverified ground truth.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript develops physics-informed surrogates for the three-dimensional temperature field in laser powder bed fusion, with emphasis on multi-track scanning paths. The authors solve the transient heat equation in an unsupervised manner using a PINN and a PI-EnDeepONet, and they propose a sequential PINN that trains one small network per track transition and assembles full multi-track solutions by recalling the trained networks. Single-track results show MAPE below 2% for both architectures; for six three-track paths, the sequential PINN achieves a MAPE of 2.49% and melt-pool errors below 5%, while reducing training time by a factor of 8.5 relative to PI-EnDeepONet. All accuracy claims are measured against a finite-difference solution of the heat equation.","tokens_in":19343,"tokens_out":6412,"duration_ms":62895,"significance":"If the reported accuracy holds, the sequential-PINN idea is a pragmatic contribution: it replaces one hard multi-track operator-learning problem with many small, Markovian track-transition problems, uses no labeled data, and cuts training time substantially for the n=3 case. The training is unsupervised, with physics-based losses, and the FD reference is independent, so there is no circular fitting of the benchmark. The paper also names the error-accumulation limitation of the sequential approach explicitly. However, the single-run nature of the experiments and the absence of any validation of the FD reference mean that the quantitative conclusions should be read as promising demonstrations rather than established accuracy guarantees.","major_comments":[{"comment":"The finite-difference solution is the sole reference for all accuracy metrics in Tables 2 and 3, yet the scheme is never validated and its discretization parameters are not reported. Eq. (5) is an explicit Euler update with temperature-dependent conductivity and an apparent-heat-capacity treatment of latent heat; such schemes have restrictive stability limits, and under-resolution near the moving Gaussian source would directly bias the peak temperatures and melt-pool dimensions that the surrogates are compared against. Please report the grid spacings and time step actually used, add a spatial and temporal mesh-convergence study, and validate the converged FD solution against an independent benchmark (an analytical solution for a Gaussian heat source, a well-established FEM code, or experimental thermocouple/IR data). Without this, the MAPE values are errors with respect to a possibly biased numerical baseline rather than to the true thermal field.","section":"Section 2.1, Eq. (5); Tables 2 and 3"},{"comment":"All accuracy and training-time comparisons are based on a single training run per model. Physics-informed networks are sensitive to initialization and collocation-point sampling; the differences between PINN (MAPE 1.44%) and PI-EnDeepONet (1.77%) in single-track cases, and the melt-pool errors of 2.5-5%, may be within run-to-run variance. Report results over multiple random seeds, including the mean and standard deviation of MAPE and melt-pool errors, so that the claimed superiority of the sequential PINN over PI-EnDeepONet is statistically supported.","section":"Section 3.3, Tables 2 and 3"},{"comment":"The paper attributes the PI-EnDeepONet degradation in multi-track scenarios to the 'curse of dimensionality' but provides no diagnostic evidence. The branch-net input does grow with the number of tracks, but the error increase could equally be caused by insufficient collocation points, branch/trunk capacity, the fixed weighting scheme of Eq. (14), or optimization difficulty. Please add a controlled experiment (for example, varying the input discretization for a fixed path, or keeping the number of tracks fixed while changing the input size) to isolate the cause, and soften the conclusion accordingly if the evidence is inconclusive.","section":"Section 3.3 and Section 4"}],"minor_comments":[{"comment":"The abstract and conclusion state that the surrogate analyzes the effect of 'scanning paths and laser parameters,' but only the tool path is varied in the experiments; laser power, scan speed, and spot size are fixed in Table 1. Please restrict the parametric claims to tool-path variation or add experiments that vary laser parameters.","section":"Abstract and Section 4"},{"comment":"The text refers to 'Fig. 8 illustrates the temperature contours' when the contour plot is actually Figure 7; the temperature-history plot is Figure 8. Correct the cross-references.","section":"Section 3.3, Figures 7 and 8"},{"comment":"Table 1 lists a convection heat transfer coefficient of 10 W/(m2K), but the governing equation in Eq. (1) and the FD update in Eq. (5) neglect convective losses. Please clarify how the convective boundary condition is imposed in the finite-difference solver and in the PINN loss terms, or remove the unused parameter.","section":"Table 1 and Eq. (5)"},{"comment":"The number of required PINNs in Eq. (20) grows like the sum of falling factorials, exceeding the number of scenarios for n>2. The word 'scalable' in the conclusion should therefore be qualified as 'scalable relative to PI-EnDeepONet for the n=3 case' rather than suggesting that the approach extends to large multi-track problems without further changes.","section":"Section 2.3, Eq. (20)"},{"comment":"No code or data are provided for the finite-difference reference or the trained networks. Releasing the FD solver and the training configuration would materially aid reproducibility, given that all benchmark metrics depend on the reference implementation.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable contribution to physics-informed surrogates for AM, and the sequential track-decomposition idea is worth publishing once the reference solver is validated. The main technical risk is the unvalidated FD baseline: because every quantitative claim is measured against it, a referee cannot currently certify the reported MAPE and melt-pool errors as errors with respect to the true physics. The random-seed issue and the unsupported 'curse of dimensionality' attribution are secondary but should also be addressed. The overclaim about laser parameters in the abstract should be corrected regardless."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take on arXiv:2502.01820. The genuinely new bit is the multi-track LPBF setting and the sequential-PINN trick: break a path into track stages, train one PINN per stage with the predecessor's end state as initial condition, then chain them. It's simple, and it works for the six three-track paths they test: MAPE under 2.5% and melt-pool dimensions under 5%, while PI-EnDeepONet degrades to 6.8% MAPE and 8-10% melt-pool errors. That comparison is the paper's real contribution—a clear demonstration that the enriched DeepONet does not handle moving-source, multi-track inputs well, and that a staged composition of standard PINNs is a practical alternative.\n\nThe methodology is mostly sound. The PINN training is unsupervised with physics losses, so the reported errors are not fits to the benchmark. The material model with temperature-dependent conductivity and apparent heat capacity is standard, and the FD scheme is independent. I think the paper is honest: they state the PI-EnDeepONet failure and acknowledge error accumulation in multi-track sequences.\n\nThe soft spots, in order of severity. The finite-difference reference is never validated. No grid sizes or time steps are given, no mesh-convergence study, no comparison to an analytic solution or experimental data. Explicit Euler with a moving Gaussian source and strong temperature-dependent coefficients can be diffusive or under-resolved. So the absolute accuracy numbers (2.49%, etc.) are only as trustworthy as that baseline; if the FD is biased near the melt pool, both surrogates inherit the bias. The relative comparison between the two methods is less affected, but the absolute claims should be softened. Second, there are no multiple-seed runs, so we don't know how stable the PINN results are. Third, the \"curse of dimensionality\" explanation for PI-EnDeepONet's failure is asserted, not demonstrated with scaling tests. Fourth, the sequential approach grows combinatorially: 15 networks for 3 tracks, and roughly factorial for n tracks. Calling it \"scalable\" is generous; it works for small n.\n\nMinor: the abstract says \"melting and consolidation\" but the model is conduction-only with apparent heat capacity—fine as a limitation, but the wording overreaches.\n\nNet: this is a solid application paper for the AM surrogate community. The main comparison is useful and likely reproducible, but the unvalidated benchmark and missing seed variance should be addressed before publication. I would send it to review.","headline":"A useful application paper showing sequential per-track PINNs beat an operator network on a small multi-track LPBF benchmark, but the unvalidated FD ground truth means the absolute error numbers are only as good as that baseline.","tokens_in":19861,"tokens_out":3559,"would_cite":true,"duration_ms":34072,"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 track-by-track PINN surrogate predicts multi-track laser powder bed fusion temperatures within 2.5% error.","keywords":["physics-informed neural networks","operator learning","laser powder bed fusion","multi-track scan paths","temperature prediction","surrogate modeling","DeepONet","sequential PINN"],"falsifier":"Run the same six three-track paths with a grid-converged finite-difference or finite-element thermal model, or with instrumented single-bead and multi-bead experiments, and recompute the surrogate's MAPE and melt pool errors: if the errors exceed the reported 2.5% and 5% thresholds, the central claim fails.","tokens_in":18909,"feed_emoji":"🔥","tokens_out":6429,"duration_ms":57149,"temperature":0.7,"pith_summary":"This paper tries to establish that multi-track laser powder bed fusion thermal histories can be predicted by a chain of physics-informed neural networks, one trained per track, instead of by a single neural operator trained on entire scan paths. The chained surrogate keeps mean absolute percentage error below 2.5% and melt pool dimension errors below 5% for all six orderings of three parallel tracks, while cutting training time by a factor of 8.5 relative to the enriched DeepONet it is compared against. That matters because scan-path planning is one of the most expensive parts of laser powder bed fusion simulation, and a cheap parametric surrogate opens the door to path optimization and real-time monitoring. The authors verify the surrogates against finite-difference solutions of the three-dimensional heat equation, not against experiments.","feed_headline":"Multi-track laser temperatures predicted to 2.5% error, 8.5x faster","feed_subtitle":"Track-by-track PINN surrogates beat one big DeepONet on all six three-track scan paths.","key_machinery":"The load-bearing mechanism is the sequential PINN decomposition: a scenario with n tracks is split into n time intervals, a dedicated PINN is trained for every track in every predecessor context (15 networks for n = 3), and the trained networks are stored and recalled in order to assemble any complete path. The initial condition for each track is the end-state of its predecessor, which is what keeps the stitched solution physically consistent. Accuracy near the moving laser comes from adaptive clustering of collocation points around the heat source, so the networks resolve the sharp thermal gradients that dominate melt pool dimensions.","core_discovery":"On the paper's own terms, the central discovery is that the sequential PINN is a scalable surrogate where the one-shot physics-informed operator fails: when the laser path is treated as an input function, the enriched DeepONet's error grows to a MAPE of 6.81% and melt pool errors of 8-10% on six three-track paths, whereas decomposing each path into tracks and chaining track-level PINNs brings MAPE to 2.49% and melt pool errors below 5%. The chain works because each track network takes the final temperature field of its predecessor as its initial condition, preserving continuity of the thermal history across track boundaries. The same chained model also tracks the cooling and consolidation phase with MAPE below 5% over time, which the authors identify as relevant for microstructure predictions.","pith_inferences":["A boundary the authors leave implicit: the number of stored PINNs in Eq. (20) grows factorially with the number of tracks, so for many-track parts the benefit of reuse shrinks unless tracks share common predecessors or the approach is paired with transfer learning.","The same chaining idea applies naturally to multi-layer builds: each new layer could start from the previous layer's final temperature field, which is the direct next test of whether error accumulation stays bounded.","A stronger validation the paper does not attempt would compare the chained surrogate's melt pool widths and depths against high-speed imaging or cross-section metallography, since the reference itself is an unvalidated finite-difference solve."],"forward_implications":["Scan-path optimization for laser powder bed fusion becomes feasible with a surrogate that holds melt pool errors below 5%, avoiding thousands of full finite-difference or finite-element solves.","Physics-informed surrogates are no longer restricted to single tracks: the sequential formulation extends the parametric idea to path-dependent processes without the accuracy collapse seen in the one-shot DeepONet.","Reusing stored track networks cuts training time from 17 hours to 2 hours for the six-path test case, so adding a new scan order costs only the recall of already-trained track models.","Because the chain resolves the cooling phase, the surrogate can feed microstructure models or support process control decisions that depend on thermal history, not just peak temperature."],"supporting_citations":[{"why":"Supplies the parametric PINN formulation, the Hastelloy X temperature-dependent properties, and the adaptive collocation clustering used by the sequential PINN.","marker":"[11]"},{"why":"Defines the DeepONet operator-learning architecture that the competing PI-EnDeepONet baseline builds on.","marker":"[19]"},{"why":"Introduces the enriched DeepONet combination operations used in the PI-EnDeepONet baseline.","marker":"[10]"},{"why":"Establishes the physics-informed loss formulation that both the PINN and PI-EnDeepONet use.","marker":"[23]"},{"why":"Motivates the L-BFGS optimizer choice for training the physics-informed networks.","marker":"[24]"}],"fun_headline_variants":["Sequential PINNs beat one-shot DeepONet on multi-track laser paths","Track-by-track PINN chains cut melt pool error below 5% in LPBF","Physics-informed surrogate predicts 3D heat flow in LPBF 8.5x faster","Multi-track LPBF temperature surrogate achieves 2.49% MAPE"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The finite-difference solver that generates the reference temperatures is an accurate and converged representation of the true physical process, yet no mesh-convergence, analytical-benchmark, or experimental validation of that solver is reported.","fun_headline_variants_meta":{"raw":{"variants":["Sequential PINNs beat one-shot DeepONet on multi-track laser paths","Track-by-track PINN chains cut melt pool error below 5% in LPBF","Physics-informed surrogate predicts 3D heat flow in LPBF 8.5x faster","Multi-track LPBF temperature surrogate achieves 2.49% MAPE"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000521,"raw_usage":{"total_tokens":2507,"prompt_tokens":917,"completion_tokens":1590,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":1503}},"tokens_in":533,"tokens_out":1590,"duration_ms":13775,"temperature":1.0,"reasoning_tokens":1503,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T14:20:09.419821+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same six three-track paths with a grid-converged finite-difference or finite-element thermal model, or with instrumented single-bead and multi-bead experiments, and recompute the surrogate's MAPE and melt pool errors: if the errors exceed the reported 2.5% and 5% thresholds, the central claim fails.","supporting_citations":[{"cited_title":"Single-track thermal analysis of laser powder bed fusion process: Parametric solution through physics-informed neural networks","cited_arxiv_id":null,"evidence_quote":"Supplies the parametric PINN formulation, the Hastelloy X temperature-dependent properties, and the adaptive collocation clustering used by the sequential PINN."},{"cited_title":"En-deeponet: An enrichment ap- proach for enhancing the expressivity of neural operators with applications to seismology","cited_arxiv_id":null,"evidence_quote":"Introduces the enriched DeepONet combination operations used in the PI-EnDeepONet baseline."}],"review_version":1}