{"id":"42a4ae0f-da80-4f0d-b5c2-952cd1ecdf3e","arxiv_id":"2411.18048","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"MeltpoolINR is a differentiable implicit neural network that maps LPBF process parameters to the temperature field, melt pool geometry, and their parametric sensitivities.","lead":"A new neural network, MeltpoolINR, learns the temperature field of a laser melting process as a smooth, differentiable function of laser power, scan speed, and preheating temperature. It reconstructs melt pool shape, cooling rates, and how fast the melt pool changes when a process setting is changed, all from simulation data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The v-sensitivity may omit the chain rule through laser position: dT/dv needs the extra term t(∂T_net/∂x_l − ∂T_net/∂ξ), which the paper neither states nor validates.","rationale":"The reader identified the unvalidated parametric derivatives as the weakest point. I agree in part, but the v-derivative has a concrete internal problem: the network input x_l encodes the scan position and depends on v, so the autodiff partial derivative is not the physical sensitivity. The paper's own Appendix A shows they know the moving-frame transformation for time derivatives, but no analogous formula appears for v. This is not merely a missing validation; it is a likely computational error that would invalidate a headline result. The test is straightforward: recompute with the chain rule and compare to finite-difference data already available in the dataset. If the curves match, the paper is fine but needs to state the implementation; if not, the central claim fails. The temperature-field surrogate and solidification-rate results are otherwise well supported, so a conditional verdict with this mandatory check is appropriate.","tokens_in":17017,"tokens_out":10882,"duration_ms":93746,"concrete_test":"Recompute the v-sensitivities in Eqs. (10)–(11) using the total derivative dT/dv = ∂T_net/∂v + t(∂T_net/∂x_l − ∂T_net/∂ξ) at t = 202 µs, and compare the resulting ∂l/∂v and ∂AR/∂v curves to the published 'MeltpoolINR (rate)' curves in Figure 10 and to central finite differences of the ground truth melt-pool length/aspect-ratio vs. speed data (v = 0.75, 1.0, 1.25, 1.5, 1.75, 2.25, 2.5 m/s). If the recomputed curves differ, the paper omitted the chain rule and the v-sensitivity claim is invalid.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The headline contribution — predicting ∂l/∂v, ∂w/∂v, and ∂AR/∂v at a fixed time — requires the physical total derivative of the temperature field with respect to scanning speed v at fixed lab coordinates (x,y,t). In the network, the temperature is represented as T_net(ξ, y, x_l; P, v, T_pre) with ξ = x − x_l and x_l = x0 + v t. Autodifferentiation w.r.t. v holding the other inputs (especially x_l) fixed yields only the partial derivative ∂T_net/∂v. The correct total derivative is ∂T/∂v = ∂T_net/∂v + t(∂T_net/∂x_l − ∂T_net/∂ξ), because varying v at fixed t changes the laser position x_l and hence ξ. The paper derives the analogous time-derivative identity in Appendix A but never states this v-chain-rule. If it is not applied when computing Eqs. (9)–(12), the 'MeltpoolINR (rate)' curves in Figure 10 are not the physical ∂l/∂v and ∂AR/∂v, even if the temperature fit is perfect. No finite-difference ground truth is provided in Figures 9–11 to catch this.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces MeltpoolINR, a fully-connected neural network surrogate for the top-view temperature field in single-track laser powder bed fusion. The model is parameterized by laser power, scanning speed, and preheat temperature, and is trained with Sobolev losses on the temperature and its spatial and temporal derivatives against SPH simulation data. The melt pool boundary is defined through the level-set function phi = T/T0 - 1, and the manuscript claims to infer the solidification rate and the rates of change of melt pool length and aspect ratio with respect to process parameters by differentiating the model. The temperature field, gradients, and cooling-rate predictions are compared with a CNN baseline on held-out SPH cases, with MeltpoolINR showing higher R2 values and lower Chamfer distances.","tokens_in":17317,"tokens_out":9875,"duration_ms":81821,"significance":"If the parametric derivative predictions are correct, MeltpoolINR would be a useful fast surrogate for LPBF process analysis and gradient-based optimization. The paper's strengths are its careful data generation and evaluation of the temperature field and its spatial/temporal derivatives: the Sobolev training is sensible, the comparison against a CNN baseline is informative, and the generalization to held-out parameter combinations is demonstrated. However, the headline contributions - the rates of change of melt pool geometry with respect to process parameters - are not validated against any ground truth, and there is a plausible missing chain-rule term in the scanning-speed sensitivity. The manuscript is therefore not yet ready for publication in its current form.","major_comments":[{"comment":"The rates of change of melt pool length and aspect ratio with respect to laser power, scanning speed, and preheat temperature are plotted as 'MeltpoolINR (rate)' curves, but they are never compared to finite-difference ground truth from the SPH data or to finite differences of the model's own geometry predictions. Without such a comparison, the central claim of the paper - prediction of the rate of change of melt pool geometry - is unsubstantiated. The statement in Section 5 that the model predicts these changes 'in close agreement with the simulation results' is only supported for the length and aspect ratio values, not for their derivatives.","section":"Section 4.2, Figs. 9-11"},{"comment":"The derivative of temperature with respect to scanning speed is not defined. The network represents T_net(xi, y, x_l; P, v, T_pre) with xi = x - x_l and x_l = x0 + v t. At fixed lab coordinates (x, y, t), the physical total derivative is dT/dv = partial T_net/partial v + t (partial T_net/partial x_l - partial T_net/partial xi). The paper derives the analogous identity for the time derivative in Eq. (A4) but never states or applies the v-derivative chain rule. If autodifferentiation is performed only with respect to the input v while holding x_l fixed, the rates shown in Figure 10 are not the physical rates, even if the temperature field is perfectly learned. The authors must state the total-derivative relation and confirm that it is used in the computation.","section":"Section 3.1 / Appendix A"},{"comment":"The Sobolev loss supervises the temperature and its x, y, and t derivatives, but not the derivatives with respect to the process parameters P, v, and T_pre. The parametric sensitivities used in Eqs. (9)-(12) are therefore not trained against any derivative targets. A neural network can fit the temperature field well while having inaccurate or noisy parameter derivatives, particularly in regions of sharp variation. The manuscript should either add derivative supervision (e.g., finite-difference targets from the SPH data) or provide a validation of the parametric derivatives against finite differences.","section":"Section 3.1, Eq. (4)"},{"comment":"The 'implicit neural representation of the melt pool boundary as a level set' is not a learned representation; phi is defined analytically as T/T0 - 1, and Eqs. (6)-(9) are used only as post-processing identities for the temperature isotherm. This framing should be clarified, and the claims in the Introduction and Conclusions about implicit neural representation should be tempered accordingly. The actual contribution is the differentiable temperature model that enables these isotherm computations, not a learned level-set function.","section":"Section 3.2, Eq. (8)"}],"minor_comments":[{"comment":"The axis label 'Rate of length change [s]' appears to be incorrect: the quantity dL/dv has units of time (e.g., microseconds), not seconds; the label should read '[us]' or '[um/(m/s)]'.","section":"Figure 10"},{"comment":"The functional form of the temperature network, including the exponent MLP and the parameter epsilon, is not given as an equation. The value of epsilon and the Fourier feature scales are not reported, which hampers reproducibility.","section":"Section 3.1 / Figure 3"},{"comment":"There are typographical issues such as 'T able 1' in the text before Table 1 and the author name 'Nathanä el' in the header line.","section":"Table 1"},{"comment":"The loss weights differ between MeltpoolINR (wx=5, wy=1, wt=5) and the CNN (wx=wy=wt=0.1); the choice of different weights for the two models is not justified and may affect the comparison of their derivative accuracy.","section":"Section 3.1"},{"comment":"The method for computing the temporal derivative (partial_t T)_data from the SPH data is not described; please clarify whether finite differences or another differentiation scheme was used.","section":"Section 2"}],"recommendation":"major_revision","confidential_remarks":"The paper's main selling point is the prediction of parametric sensitivities of melt pool geometry, but this is precisely the part that is least supported. The missing validation in Figures 9-11 and the potentially omitted chain-rule term for the v-derivative are load-bearing concerns that cannot be resolved by minor edits. I would encourage the authors to (i) derive and state the total derivative with respect to v, (ii) implement it consistently, and (iii) validate all three parametric sensitivities against finite-difference ground truth from the SPH data. If the validation reveals inaccuracies, the scope of the claims should be reduced accordingly. The temperature-field modeling itself appears solid and well-evaluated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the core surrogate works: a parametric implicit neural representation that learns the LPBF temperature field, its spatial gradients, and cooling rate from SPH data, with held-out R2 in the high 0.9s and a consistent win over a CNN baseline on gradients and melt pool boundary geometry (Chamfer distance). Second, the paper's headline claim — rates of change of melt pool geometry with respect to process parameters — is never validated against anything, and the v-sensitivity has an unstated chain-rule issue that could make those curves physically wrong even if the temperature fit is perfect.\n\nWhat is actually new here is the assembly: Fourier-feature MLPs with a parameter encoder, Sobolev training on temperature plus derivatives, an exponential Rosenthal-inspired term to handle the sharp front, and a level-set defined by phi = T/T0 - 1 that turns the learned field into melt pool boundary, solidification rate, and parametric geometry sensitivities. Each ingredient exists separately, but combining them for a parametric LPBF surrogate with autodiff sensitivities is a real step beyond the CNN surrogate work. The authors also define the geometry explicitly from the temperature field, which makes the geometry claims transparently reducible to temperature accuracy rather than an independent output. That is a design choice, not a flaw.\n\nThe soft spots, in proportion. Temperature, gradient, and cooling-rate claims are solid, and the level-set speed function F is checked against the SPH ground truth in Figure 8 with close agreement. What is not checked is the parametric sensitivity. Figures 9-11 show the \"MeltpoolINR (rate)\" curves with no finite-difference ground truth drawn from the 210 SPH cases, no error bars, no experimental check. That alone makes the headline claim conditional. On top of that, the v-sensitivity requires the total derivative of the network temperature at fixed lab coordinates: dT/dv = dT_net/dv + t(dT_net/dx_l - dT_net/dxi), because varying v at fixed time moves the laser. The paper derives the analogous time identity in Appendix A but never states the v version and never says which derivative feeds Eqs. (9)-(12). If they used only the partial derivative, the rate curves are not the physical dl/dv, dw/dv, or dAR/dv. I worked through the geometry terms: the -t dT/dxi part cancels between the two x-extremes for the length, but survives for the width and aspect ratio, and the t dT/dx_l part survives everywhere unless the moving-frame field is strictly quasi-steady. This is fixable — state the chain rule and validate the rates against finite differences from the SPH data — but it needs to be done. No code or data is released, which makes the unvalidated rates harder to audit.\n\nWho this is for: anyone building differentiable surrogates for process optimization in AM, and anyone combining INRs with Sobolev training for sharp-front fields. It deserves a serious referee, with a major revision expected.","headline":"Solid, well-validated temperature-field surrogate with a clean level-set geometry trick; the headline parametric sensitivities are unvalidated and the v-sensitivity chain rule is never stated, so serious review but expect a major revision.","tokens_in":17819,"tokens_out":20222,"would_cite":true,"duration_ms":157602,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A differentiable neural network learns the melt pool temperature field and its parametric derivatives in laser powder bed fusion.","keywords":["laser powder bed fusion","melt pool","implicit neural representation","Fourier features","Sobolev training","level set","solidification rate","surrogate modeling"],"falsifier":"Take a trained MeltpoolINR, compute the melt pool length and aspect ratio at a reference parameter set, then at small finite-difference increments of power, speed, and preheat, and compare the resulting slopes with the model's reported rate curves; a mismatch would show the parametric derivatives are not physical. Alternatively, measure melt pool length from high-speed imaging at varied parameters and compare the rates of change.","tokens_in":16857,"feed_emoji":"🔥","tokens_out":4012,"duration_ms":34742,"temperature":0.7,"pith_summary":"This paper proposes MeltpoolINR, a differentiable neural network that learns the top-view temperature field of a single-track laser powder bed fusion process as a continuous function of position, time, laser power, scan speed, and preheat temperature. The aim is not only to predict temperatures but also their spatial and temporal gradients and the shape of the melt pool, and—through automatic differentiation—the rate at which the melt pool changes when process parameters change. If it works, it would give fast surrogate predictions of quantities that currently require expensive thermo-fluid simulations, and would supply the parametric sensitivities needed for gradient-based process optimization and microstructure modeling. The paper validates the field predictions against a CNN baseline and uses the level-set representation to infer solidification rate and geometric rates of change.","feed_headline":"Neural net predicts melt pool shape and how it responds to laser settings","feed_subtitle":"Differentiable neural net with a level-set melt pool boundary yields temperature fields and parametric rates from SPH data.","key_machinery":"The central object is the differentiable Fourier-feature MLP combined with the level-set representation of the melt pool. Temperature is learned in the laser-attached frame using an exponential term patterned on Rosenthal's solution, which lets a small network represent the sharp front; the zero level set of $\\phi = T/T_0 - 1$ defines the boundary; and the speed function $F$ and the perturbation equation $\\partial\\phi/\\partial P_i + F_{P_i}|\\nabla\\phi| = 0$ convert learned parametric derivatives into rates of change of melt pool length, width, and aspect ratio via eqs. (10)-(12).","core_discovery":"MeltpoolINR treats the temperature field as an implicit neural representation: four multilayer perceptrons with Fourier feature encoding map the laser-attached coordinates ($\\xi = x - x_l$), the $y$-coordinate, laser position, and process parameters to the temperature, with an exponential output factor inspired by Rosenthal's moving point-source solution that helps capture the sharp front of the melt pool. The network is trained with a Sobolev loss that fits temperature and its $x$-, $y$-, and $t$-derivatives to SPH simulation data. The melt pool boundary is the zero level set of $\\phi = T/T_0 - 1$, so the speed function $F = -(\\partial\\phi/\\partial t)/|\\nabla\\phi|$ gives the solidification rate, and a level-set perturbation equation gives $\\partial\\phi/\\partial P_i$, from which rates of change of melt pool length, width, and aspect ratio with respect to laser power, scan speed, and preheat temperature are computed. The paper reports that MeltpoolINR matches high-fidelity simulation data closely, outperforms the CNN on gradients, cooling rate, and boundary geometry, and generalizes to held-out parameter combinations.","pith_inferences":["The level-set perturbation rates could be validated by finite-difference slopes from the same trained model; if they match, MeltpoolINR could replace adjoint solvers in LPBF optimization loops.","The Sobolev-trained derivatives may serve as weak supervision for physics-informed thermal solvers in regions where simulation data are sparse.","Applying the model to alloys with very different thermal diffusivity would probably require retuning the Rosenthal-inspired exponent, a check that could reveal how much of the accuracy is physics prior versus learned fit."],"forward_implications":["Fast inference of temperature field, gradient, and cooling rate for single-track LPBF without rerunning SPH simulations; held-out parameter errors are comparable to training-case errors.","Accurate melt pool boundary prediction via the implicit level set, with Chamfer distances better than the CNN baseline.","Solidification rate inferred directly from the level-set speed function, avoiding grid-interpolation noise that appears in the simulation-based reference.","Parametric rate predictions for melt pool length and aspect ratio as functions of power, speed, and preheat, enabling gradient-based process optimization.","The same architecture can be extended to 3D fields and multi-track printing, as the authors state."],"supporting_citations":[{"why":"Rosenthal's moving heat source solution supplies the exponential structure of the temperature output, used to capture the sharp front of the melt pool.","marker":"[13]"},{"why":"Fourier feature mapping, shown to let MLPs learn high-frequency functions, enables the network to represent sharp melt-pool features.","marker":"[47]"},{"why":"Sobolev training adds derivative matching to the loss, which is how the gradients and cooling rate are learned.","marker":"[48]"},{"why":"Ray-tracing SPH laser model generates the high-fidelity temperature data that the network is trained against.","marker":"[51]"},{"why":"Level set methods provide the evolution equation and the perturbation equation used for the melt pool boundary and its rates of change.","marker":"[56]"},{"why":"Neural radiance fields establish the implicit neural representation approach that the architecture builds on.","marker":"[46]"},{"why":"Adaptive SPH scheme provides the efficient single-track melt pool simulations used to create the dataset.","marker":"[23]"},{"why":"Convolutional network architecture used as the state-of-the-art baseline for comparative evaluation.","marker":"[55]"}],"fun_headline_variants":["Differentiable neural net predicts melt pool geometry and rates","Implicit neural representation of melt pool with level-set boundary","Sobolev loss and Fourier features drive melt pool predictions","MeltpoolINR: Differentiable net computes geometry and rates","Neural net predicts melt pool shape and rates from SPH data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the network's derivative of the temperature field with respect to laser power, scan speed, and preheat temperature is accurate, since the level-set rate predictions (eqs. (9)-(12)) inherit that accuracy; the paper does not validate these rates against finite-difference or experimental ground truth.","fun_headline_variants_meta":{"raw":{"variants":["Differentiable neural net predicts melt pool geometry and rates","Implicit neural representation of melt pool with level-set boundary","Sobolev loss and Fourier features drive melt pool predictions","MeltpoolINR: Differentiable net computes geometry and rates","Neural net predicts melt pool shape and rates from SPH data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000817,"raw_usage":{"total_tokens":3597,"prompt_tokens":980,"completion_tokens":2617,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":596,"completion_tokens_details":{"reasoning_tokens":2532}},"tokens_in":596,"tokens_out":2617,"duration_ms":16059,"temperature":1.0,"reasoning_tokens":2532,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:33:32.629112+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a trained MeltpoolINR, compute the melt pool length and aspect ratio at a reference parameter set, then at small finite-difference increments of power, speed, and preheat, and compare the resulting slopes with the model's reported rate curves; a mismatch would show the parametric derivatives are not physical. Alternatively, measure melt pool length from high-speed imaging at varied parameters and compare the rates of change.","supporting_citations":[{"cited_title":"Transactions of the American Society of Mechanical Engineers 68(8), 849–865 (1946)","cited_arxiv_id":null,"evidence_quote":"Rosenthal's moving heat source solution supplies the exponential structure of the temperature output, used to capture the sharp front of the melt pool."},{"cited_title":"Advances in neural information processing systems 33, 7537–7547 (2020)","cited_arxiv_id":null,"evidence_quote":"Fourier feature mapping, shown to let MLPs learn high-frequency functions, enables the network to represent sharp melt-pool features."},{"cited_title":"International Journal of Heat and Mass Transfer 214, 124378 (2023)","cited_arxiv_id":null,"evidence_quote":"Ray-tracing SPH laser model generates the high-fidelity temperature data that the network is trained against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Level set methods provide the evolution equation and the perturbation equation used for the melt pool boundary and its rates of change."},{"cited_title":"Computers & Mathematics with Applications 139, 7–27 (2023)","cited_arxiv_id":null,"evidence_quote":"Adaptive SPH scheme provides the efficient single-track melt pool simulations used to create the dataset."},{"cited_title":"The International Journal of Advanced Manufacturing Technology, 1–16 (2023)","cited_arxiv_id":null,"evidence_quote":"Convolutional network architecture used as the state-of-the-art baseline for comparative evaluation."}],"review_version":1}