{"id":"02d94a4f-54db-48b8-bb8d-996ae15a605c","arxiv_id":"2501.08922","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"Simulation-based machine learning and polynomial regression predict melt pool dimensions and spatter volume in LPBF from power and velocity, but the spatter labels come from a classifier transfer and the headline R2 values are partly computed on a log-scaled target that the text does not disclose.","lead":"The authors mapped 281 laser powder bed fusion simulations with machine learning and polynomial fits, claiming high accuracy in predicting melt pool dimensions and spatter volume from laser power and scan speed. If the fitted equations hold up, they would give additive manufacturing engineers a cheap, interpretable way to anticipate melt pool and spatter behavior without running expensive simulations.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Undisclosed log transformation of the spatter target makes the Table 2 R2/MAE comparisons invalid, so the central high-accuracy claim is unsupported.","rationale":"The reader's weakest_assumption targets the classifier transfer from OpenFOAM to FLOW-3D, which is a serious threat to the physical validity of the spatter labels. I consider that concern important, but the log-transformation issue is the single most load-bearing because it can be settled directly from the paper's own Table 2 and does not depend on interpreting simulation fidelity. The paper's central quantitative claim—high R2 for spatter from process conditions or melt pool dimensions—rests entirely on those metrics. If the target was transformed without disclosure, the R2 values in the log-input rows are not comparable to the non-log rows, the 'improvement' narrative in Section 3.2 collapses, and the fitted polynomial equations in Tables 5 and A.6 cannot be tied to the reported accuracy. This reinforces the reader's REJECT verdict; I see no need to move the verdict. A secondary concern remains that the target itself may be an artifact of unvalidated classifier transfer, but that is less directly checkable from the manuscript alone.","tokens_in":18478,"tokens_out":5721,"duration_ms":60228,"concrete_test":"Obtain the dataset and code, or rerun the pipeline with the exact train/test split and hyperparameters from Table 1. Fit the ExtraTree and RF models using the log-input feature sets ([Power, Velocity, logVelocity] and [logLength, Width, Depth, logWidth, logDepth]) with the target left in its original 'Volume Indicated as Spatter' units, and reproduce the Table 2 MAE values. If the original-scale target yields MAE around 20,000 rather than 0.2, the log-input rows must have used a log-transformed target, confirming the inconsistency. As a second check, fit the same models with the target log-transformed, exponentiate the predictions, and recompute R2 and MAE on the original scale; if the original-scale test R2 falls below 0.85, the claim that input log-transformation alone improves spatter prediction is false.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central accuracy claim for spatter is invalidated by an internal inconsistency in Table 2. For the same target ('Volume Indicated as Spatter'), the non-log rows have Train/Test MAE of roughly 20,000-35,000 (e.g., RF: 18,709.6 / 28,853.8), while the rows with log-transformed inputs have Train/Test MAE of roughly 0.1-0.3 (e.g., ExtraTree with log Length, Width, Depth, logWidth, logDepth: 0.1112 / 0.2173). Section 3.2 and the abstract state that only the model inputs were logarithmically transformed, leaving the target unchanged. But a five-orders-of-magnitude drop in MAE for the same target is mathematically impossible unless the target itself was also log-transformed or otherwise rescaled. This undisclosed target transformation means the reported 'improvement' in R2 from 0.80 to 0.875 is not a comparison on the original spatter-volume scale, and the headline ExtraTree R2 of 96.7% is not the R2 for the quantity named in the paper. The polynomial equations in Tables 5 and A.6 appear to be expressed for original-scale volume, but the reported high R2 values for the log-input models cannot be reconciled with those equations from the information given.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a machine-learning and polynomial-regression framework to predict melt pool dimensions (length, width, depth), melt pool geometry (cross-sectional area, volume), and a quantity called 'volume indicated as spatter' for laser powder bed fusion (LPBF). The dataset consists of 281 FLOW-3D simulations, with spatter labels generated by applying a classifier trained on OpenFOAM simulations to FLOW-3D output. The authors report R2 values above 95% for melt pool features and, after logarithmic transformation of model inputs, an ExtraTree model achieving R2 of 96.7% (train) and 87.5% (test) for spatter volume. They also provide polynomial equations intended as interpretable 'constitutive models' for these quantities.","tokens_in":18682,"tokens_out":3986,"duration_ms":42745,"significance":"If the reported accuracy were valid on the original physical scale and the spatter target were physically meaningful, the framework could provide inexpensive surrogate models for LPBF process mapping and a step toward interpretable quality-control models. The paper usefully compares five ML algorithms and reports explicit fitted equations, which is a strength. However, the central spatter accuracy claim is compromised by an internal inconsistency in the reported MAE values, and the spatter target itself is a classifier-generated quantity without experimental validation. The paper does not provide code or data, so the equations and metrics cannot be independently checked. The idea is potentially interesting, but the current manuscript does not establish the central claim.","major_comments":[{"comment":"The reported MAE values for the spatter target are internally inconsistent. Rows for inputs [Length, Width, Depth] report Train/Test MAE on the order of 23,000–35,000 (e.g., KNN: 27,269.7 / 27,395.2), while rows for inputs [log Length, Width, Depth, logWidth, logDepth] report Train/Test MAE of roughly 0.1–0.3 (e.g., ExtraTree: 0.1112 / 0.2173). The text in Section 3.2 and the abstract state that only the model inputs were logarithmically transformed, leaving the target unchanged. For the same target in the same units, a five-orders-of-magnitude drop in MAE is not possible unless the target was also log-transformed or otherwise rescaled. This means the claimed improvement in R2 from 0.80 to 0.875, and the headline ExtraTree R2 of 96.7%, are not demonstrated for the quantity named in the paper. The authors must disclose whether the target was transformed, and report R2 and MAE on the original spatter-volume scale, or explicitly reframe all claims as applying to a transformed target.","section":"Section 3.2, Table 2"},{"comment":"Table A.6 is captioned as the polynomial equation derived from logarithmic transformation of melt pool dimensions, but the equation is written in terms of P, V, and log(V) (process conditions), not length, width, and depth. This contradicts the caption, the text, and Table 5, and it makes the appendix unusable for reproducing the claimed 'volume indicated as spatter' model based on melt pool dimensions. The correct equation for the log-transformed melt-pool-dimension model must be provided, and the erroneous duplicate/placeholder equation removed.","section":"Section 3.3, Table A.6"},{"comment":"The spatter volume target is not a direct simulation output; it is produced by a classifier trained on OpenFOAM features and then applied to FLOW-3D data after 'aligning the range of each feature' with the OpenFOAM dataset. The manuscript does not quantify the classifier's transfer accuracy, the domain shift between OpenFOAM and FLOW-3D, or the sensitivity of the derived spatter values to the alignment procedure. Since FLOW-3D is acknowledged to lack the physics needed to produce realistic spatter (Section 2.1), the high R2 values for spatter may describe self-consistency of the simulation+classifier pipeline rather than physical spatter behavior. The authors should either provide a validation benchmark (e.g., against experimental spatter measurements or high-fidelity simulations) or clearly restrict all claims to the surrogate 'volume indicated as spatter' rather than physical spatter.","section":"Section 2.2, Figure 1"},{"comment":"The polynomial equations are presented as 'constitutive models,' but only training R2 is reported for the equations in Tables 4, 5, and A.6, while Table 3 gives both training and test R2. Without test-set performance for the displayed equations, and without any uncertainty quantification for the coefficients, the reader cannot assess whether these equations generalize or merely overfit the training data. At minimum, the test R2 should be listed next to each equation, and the term 'constitutive models' should be reconsidered, since these are empirical response surfaces rather than constitutive laws derived from physical principles.","section":"Section 3.3, Tables 3 and 5"}],"minor_comments":[{"comment":"The acronym 'LBPF' appears in the introduction; it should be 'LPBF' (laser powder bed fusion).","section":"Introduction"},{"comment":"The text refers to 'volume indicated as splinter' and 'avalanche meltpool volume'; these appear to be typographical errors or inconsistent feature names and should be corrected to match the terminology used elsewhere in the paper.","section":"Section 3.1, Figure 3 text"},{"comment":"The equation numbering is inconsistent: the text refers to equations (3), (4), and (5), but the displayed equations are numbered (1), (2), and (3), and equation (4) is used for velocity magnitude in Section 2.2. All in-text equation references should be renumbered or re-checked.","section":"Section 2.1"},{"comment":"The first spatter equation begins with '-47591.2675 - 0.1273 - 0.2616P', where the '-0.1273' term has no corresponding variable; it is likely a typographical artifact or a missing term and should be corrected.","section":"Table 5"},{"comment":"The captions for panels (F) and (G) are nearly identical and do not clearly distinguish the two input sets; the captions should be revised to state exactly which inputs (e.g., [Power, Velocity] versus [Power, Velocity, log(Velocity)]) are used in each panel.","section":"Figures 4 and 5"},{"comment":"The paper does not include a data availability or code availability statement. Given that the dataset is synthetic and the pipeline is complex, releasing the code and data (or a reproducible benchmark) would be important for the claims to be verifiable.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's core problem is the undisclosed or misreported target transformation in Table 2, which invalidates the headline spatter R2 as currently stated. This is fixable by re-running the analysis with a consistent target and reporting metrics on the original scale, or by clearly reframing the claims. The second major issue is the lack of any external validation for the classifier-generated spatter volume; even after fixing the metrics, the physical relevance of 'volume indicated as spatter' remains unestablished. I recommend major revision rather than rejection because the methodological core—comparing ML regressors and polynomial fits on simulation-derived melt pool features—is sound and could be salvaged with a corrected and more cautious presentation. However, if the authors cannot reconcile the MAE inconsistency or provide any validation of the spatter target, the paper should be rejected, since the primary claimed contribution would then be unsupported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read on arXiv:2501.08922. The paper has two distinct halves. The melt pool part—predicting length, width, depth, area, volume from power and velocity—appears sound and standard. R2 values above 95% on train/test are plausible for a simulation dataset, and the fitted polynomial equations in Table 4 are what you'd expect from response-surface fitting. That half is useful engineering.\n\nThe spatter-volume half, which is the advertised novelty, has a load-bearing problem. The stress-test note is right: Table 2 shows MAE around 20,000–35,000 for models using raw inputs, but MAE around 0.1–0.3 for models whose inputs include logs. The text and abstract say only the inputs were log-transformed. A five-orders-of-magnitude drop in MAE for the same target is not physically possible unless the target itself was also log-transformed or rescaled. So the reported R2 improvements (from 0.80 to 0.875 for ExtraTree, and the headline 96.7% on training) are not comparisons on the original spatter-volume scale. This is undisclosed.\n\nThat makes the central quantitative claim—ExtraTree achieves 96.7% R2 predicting spatter volume—unsupported as stated. The polynomial equations in Table A.6 may be for log-transformed targets, but the text doesn't say so, and the coefficients are presented as if predicting volume directly. This is a statistical reporting error that cannot be waved off.\n\nBeyond that, the \"constitutive models\" terminology overclaims. These are least-squares fits to a simulation-generated target, with no experimental validation and no shipped code or data to reproduce the pipeline. The spatter volume itself comes from a classifier trained on OpenFOAM and applied to FLOW-3D, which does not model spatter physics directly. The paper acknowledges this limitation to its credit, but it means the fitted equations describe the simulation pipeline, not necessarily physical spatter.\n\nWho should read this? Process engineers building surrogate models for LPBF might find the melt pool equations useful; the 281-case dataset could be a contribution if released. But the paper as submitted would mislead anyone relying on the spatter R2 numbers.\n\nMy recommendation: this deserves a serious referee, not a desk reject, because the pipeline is intelligible and the flaw looks fixable. The authors should retrain or re-report with the target on the original scale, disclose any log transformation of the target, and provide code/data. If they do, the melt pool half stands and the spatter half might become usable. As is, I would not cite the spatter numbers.","headline":"The melt pool predictions are fine, but the spatter R2 claims are invalidated by an undisclosed log transformation of the target.","tokens_in":19274,"tokens_out":3033,"would_cite":false,"duration_ms":29676,"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":"This paper reports that melt pool dimensions, area, volume, and spatter volume in laser powder bed fusion are predictable from laser power and scan velocity alone, with $R^2$ above 95% for melt pool geometry and up to 96.7% for spatter…","keywords":["laser powder bed fusion","spatter","melt pool dynamics","machine learning","symbolic regression","polynomial regression","process map","computational fluid dynamics"],"falsifier":"Run the fitted polynomial equations against experimentally measured spatter volumes from high-speed or synchrotron X-ray imaging of equivalent SS316L single tracks over the same power-velocity grid; if the equations systematically diverge, especially in regimes with deep keyholes or spatter modes absent from the training simulations, the FLOW-3D labels cannot be representing physical spatter.","tokens_in":18201,"feed_emoji":"⚙️","tokens_out":11828,"duration_ms":107437,"temperature":0.7,"pith_summary":"This paper aims to show that in laser powder bed fusion, the dimensions, area, and volume of the melt pool, together with the volume of material ejected as spatter, are set by just two process variables: laser power and scan velocity. The evidence is built by fusing two simulation tools: OpenFOAM, which resolves the gas dynamics behind spatter ejection but is expensive, and FLOW-3D, which runs quickly but approximates recoil pressure and mass transfer and omits tangential surface tension. A classifier trained on OpenFOAM output labels the fast FLOW-3D simulations, yielding a dataset of 281 single-track conditions from which melt pool features and 'volume indicated as spatter' are extracted. Machine learning regressors and polynomial symbolic regression fit this dataset, with $R^2$ above 95% for melt pool dimensions and geometry and best spatter values of 96.7% (training) and 87.5% (testing). If the classifier transfer is faithful, the fitted polynomial equations are interpretable constitutive models that could guide process maps, monitoring, and defect control.","feed_headline":"Power and scan speed predict melt pool shape and spatter volume","feed_subtitle":"If true, power and speed suffice to forecast melt pool defects, giving engineers an interpretable control law.","key_machinery":"The load-bearing mechanism is a two-stage pipeline: a spatter/melt pool classifier trained on OpenFOAM simulations is transferred to FLOW-3D output, after aligning feature ranges, to label 'volume indicated as spatter' in simulations that do not resolve spatter physics; then polynomial regression of degrees 2 to 6 transforms power and velocity, or melt pool dimensions, into polynomial and interaction terms and fits a linear model, producing interpretable equations. The logarithmic transformations of inputs, such as $\\log$ velocity, $\\log$ length, $\\log$ width, and $\\log$ depth, are a deliberate part of the machinery: they improve spatter $R^2$ substantially, indicating that spatter volume responds approximately linearly to log-transformed process and geometry variables. Feature importance is read off from the absolute sizes of the polynomial coefficients, which is what turns the regressions into claims about which physical variables drive melt pool shape and spatter.","core_discovery":"The central claim is that a low-dimensional, closed-form map connects process conditions to melt pool state and spatter: laser power and scan velocity alone fix the stable melt pool length, width, depth, cross-sectional area, and volume, and those same process conditions or the melt pool dimensions fix the volume indicated as spatter. In the paper's own terms, this is a discovery of constitutive models for spatter, namely explicit polynomial equations with $R^2$ above 95% for melt pool geometry features and 85% (training) and 82% (testing) for spatter in the polynomial fits. The best machine learning result, the ExtraTree model, reaches $R^2$ of 96.7% on training and 87.5% on testing for spatter when log-transformed melt pool dimensions are the inputs. The derived equations also rank the importance of physical inputs: power dominates the equations for length and depth, velocity dominates width, and width and depth dominate the spatter equations built from melt pool dimensions.","pith_inferences":["If the central claim holds, an untested consequence is that closed-loop control could regulate spatter by adjusting power and velocity from measured melt pool dimensions, without needing a direct spatter sensor.","Because the FLOW-3D simulations omit tangential surface tension, the fitted equations describe a reduced-physics world; comparing them with full-physics OpenFOAM output or experiments would reveal which polynomial terms are artifacts of that simplification.","A testable extension follows from the strong power-spatter correlation: the classifier's label is probably dominated by recoil-pressure ejection, so the equations should miss spatter generated by other mechanisms such as powder-bed entrainment or laser-plume interactions."],"forward_implications":["With power and velocity fixed before a build, the fitted equations yield forecasts of melt pool length, width, depth, area, volume, and spatter volume without additional simulation cost.","Because melt pool dimensions alone predict spatter volume, in-situ sensing of melt pool shape during printing can act as a surrogate spatter monitor for defect detection.","The feature-importance rankings in the equations identify which process knob to turn: power for length and depth, velocity for width, and width plus depth for spatter volume.","The consistent improvement from logarithmic inputs implies spatter volume scales in an approximately power-law or log-linear way with process conditions, not linearly."],"supporting_citations":[{"why":"Supplies the spatter/melt pool classifier and feature-alignment methodology trained on OpenFOAM output; it is the source of the spatter labels transferred to FLOW-3D.","marker":"[5]"},{"why":"Cited for the process-instability physics behind spatter; supports the paper's statement that FLOW-3D's simplifications omit spatter-relevant phenomena.","marker":"[47]"},{"why":"Provides the FLOW-3D melt pool simulation basis used to generate the 281 cases and is cited for its assumptions about recoil pressure and mass transfer.","marker":"[48]"},{"why":"Underpins the OpenFOAM solver configuration and surface-tension modeling choices used for the high-fidelity spatter simulations.","marker":"[2]"},{"why":"Supplies the icoReactingMultiphaseInterFoam CFD solver used for the OpenFOAM simulations that train the classifier.","marker":"[49]"},{"why":"Documents the OpenFOAM toolchain that produces the spatter-resolving simulation data.","marker":"[50]"},{"why":"Defines the Extremely Randomized Trees algorithm that achieves the best spatter $R^2$ values.","marker":"[56]"}],"fun_headline_variants":["Laser power and speed alone forecast melt pool and spatter","Melt pool and spatter equations from just power and speed","Two inputs, one equation: power and speed give melt pool and spatter","Machine learning finds simple laws for melt pool and spatter"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire 'volume indicated as spatter' result rests on a single transfer: a classifier trained on OpenFOAM output is assumed to label spatter correctly in FLOW-3D simulations, even though FLOW-3D approximates recoil pressure and mass transfer and omits tangential surface tension; if that transfer is unfaithful, every reported spatter equation describes an artifact, not physical spatter.","fun_headline_variants_meta":{"raw":{"variants":["Laser power and speed alone forecast melt pool and spatter","Melt pool and spatter equations from just power and speed","Two inputs, one equation: power and speed give melt pool and spatter","Machine learning finds simple laws for melt pool and spatter"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001091,"raw_usage":{"total_tokens":4604,"prompt_tokens":1039,"completion_tokens":3565,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":655,"completion_tokens_details":{"reasoning_tokens":3504}},"tokens_in":655,"tokens_out":3565,"duration_ms":21789,"temperature":1.0,"reasoning_tokens":3504,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:14:30.068844+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the fitted polynomial equations against experimentally measured spatter volumes from high-speed or synchrotron X-ray imaging of equivalent SS316L single tracks over the same power-velocity grid; if the equations systematically diverge, especially in regimes with deep keyholes or spatter modes absent from the training simulations, the FLOW-3D labels cannot be representing physical spatter.","supporting_citations":[{"cited_title":"Integrating Multi-Physics Simulations and Machine Learning to Define the Spatter Mechanism and Process Window in Laser Powder Bed Fusion","cited_arxiv_id":"2405.07823","evidence_quote":"Supplies the spatter/melt pool classifier and feature-alignment methodology trained on OpenFOAM output; it is the source of the spatter labels transferred to FLOW-3D."},{"cited_title":"Controlling process instability for defect lean metal additive manufacturing","cited_arxiv_id":null,"evidence_quote":"Cited for the process-instability physics behind spatter; supports the paper's statement that FLOW-3D's simplifications omit spatter-relevant phenomena."},{"cited_title":"Computational investigation of melt pool process dynamics and pore formation in laser powder bed fusion","cited_arxiv_id":null,"evidence_quote":"Provides the FLOW-3D melt pool simulation basis used to generate the 281 cases and is cited for its assumptions about recoil pressure and mass transfer."},{"cited_title":"Implementation of the marangoni effect in an open-source software environment and the influence of surface tension modeling in the mushy region in laser powder bed fusion (lpbf)","cited_arxiv_id":null,"evidence_quote":"Underpins the OpenFOAM solver configuration and surface-tension modeling choices used for the high-fidelity spatter simulations."},{"cited_title":"Cfd simulation of fluid flow during laser metal wire deposition using openfoam: 3d printing, 2019","cited_arxiv_id":null,"evidence_quote":"Supplies the icoReactingMultiphaseInterFoam CFD solver used for the OpenFOAM simulations that train the classifier."},{"cited_title":"Openfoam documentation","cited_arxiv_id":null,"evidence_quote":"Documents the OpenFOAM toolchain that produces the spatter-resolving simulation data."},{"cited_title":"Extremely randomized trees","cited_arxiv_id":null,"evidence_quote":"Defines the Extremely Randomized Trees algorithm that achieves the best spatter $R^2$ values."}],"review_version":1}