{"id":"c55e3023-913e-44ea-92f7-046ee9a2a37f","arxiv_id":"2608.01903","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"An element-independent random-forest/XGBoost model can interpolate composition from PVD cathode powers and invert powers from target composition on a 7-element, 82-point dataset, with RMSE of about 0.1 in composition and 70 W in power.","lead":"This paper reports a machine-learning model trained on 82 thin-film deposition experiments that predicts coating composition from cathode powers and, in reverse, the powers needed for a target composition, using element-agnostic descriptors such as sputtering yield. It also describes the French DIADEM initiative and proposed workflows for corrosion-resistant coatings for nuclear energy.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Element-independence is asserted but only tested by random splits within seven elements; no held-out element test, so the central transferability claim is unsupported.","rationale":"The reader's weakest assumption correctly identifies the element-independence claim as the load-bearing issue. My analysis refines it: the random 80/20 split means every element appears in training, so the reported RMSE cannot distinguish genuine descriptor-based transfer from element-label lookup. The paper is otherwise honest about its limitations and the narrow interpolation claim is plausible, so the appropriate verdict remains CONDITIONAL. The proposed leave-one-element-out test directly settles whether the central claim holds.","tokens_in":11001,"tokens_out":2142,"duration_ms":27682,"concrete_test":"Leave-one-element-out: remove all rows for one of the seven elements (e.g. W or Nb), retrain the same Random Forest/XGBoost pipeline on the remaining ~70 points, and predict the held-out element's composition and required powers. Report held-out RMSE and R² per element against the random-split metrics. If held-out error is drastically higher or no better than a simple baseline that predicts composition from power weighted by published sputtering yields, the element-independence claim fails and the model is only interpolating known elements.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is transferability: Section 4.2 says the model uses cathode operating mode and normalized physical descriptors such as sputtering yield instead of chemical identity, so the same model should apply to new elements. The load-bearing weakness is that this is asserted but never tested. The dataset contains 7 elements and 82 points, and the evaluation (Section 4.3) uses a random 80/20 split with 5-fold repetition. Under a random split, every element appears in training, so the test set only checks interpolation among known elements. With only 7 elements, sputtering yield is nearly a unique elemental label, so the 'element-independent' descriptor set may function as an implicit lookup table rather than a physically transferable representation. The sentence in Section 4.4 claiming that the methodology can be applied to elements not in the dataset is therefore unsupported. No leave-one-element-out experiment, no comparison to a simple sputter-yield scaling baseline, and no per-element errors are reported. If the model fails on a held-out element, the value proposition collapses to interpolation on one small seven-element system, not a generic framework.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a combination of a programmatic overview of the French DIADEM materials-data-science initiative and a machine-learning study on the DIADEM-2D combinatorial magnetron sputtering platform. The authors describe the selection of candidate alloy systems for molten-salt corrosion (Ni-Cr-Mo-Al) and nuclear cladding diffusion barriers (V-Nb-Mo-W), then train Random Forest, XGBoost, Bayesian Ridge, and a fuzzy inference system on a dataset of seven elements (Ni, Cr, Mo, Al, W, Nb, V; 82 points). The forward task predicts coating composition from cathode powers; the inverse task predicts powers from a target composition. The authors report RMSE ≈ 0.1 (forward) and RMSE ≈ 70 W (backward) and claim that, because the model uses physical descriptors such as sputtering yield rather than element identity, it is element-independent and transferable to new elements without retraining. The central conclusion is that this is a first prototype of a generic machine-learning framework for inverse design of PVD process parameters.","tokens_in":11229,"tokens_out":3766,"duration_ms":42034,"significance":"If the element-transferability claim were established, this would be a valuable contribution to high-throughput combinatorial sputtering, substantially reducing experimental effort when exploring new alloy systems. The paper has clear strengths: it describes a concrete closed-loop platform concept, checks model behavior through correlation matrices, and explores interpretable fuzzy systems in addition to black-box ensembles. However, the evidence presented is an in-sample proof of concept only. The evaluation uses random splits in which every element appears in training, no element is held out, there is no comparison against a simple physical scaling baseline, and no uncertainty quantification or per-element errors are reported. The central 'generic, element-independent' claim is therefore not yet supported by the data. The work could become a meaningful demonstration if the authors add leave-one-element-out cross-validation, per-element and test-only metrics, a descriptor-based baseline, and public data/code.","major_comments":[{"comment":"The load-bearing claim is element independence and transferability, but it is never tested. Section 4.3 describes a random 80/20 split repeated five times; in such a split every element appears in the training set, so the test set only measures interpolation among already seen elements. Section 4.4 then states that 'This methodology can be applied to dataset with elements which are not present in the dataset', a claim not supported by any experiment. With only seven elements, descriptors such as sputtering yield are nearly unique per element and may serve as an implicit element label. Please add leave-one-element-out cross-validation, report per-element errors, and compare against a simple sputtering-yield normalization baseline (e.g., composition approximately proportional to yield × power normalized over cathodes). Without such a test, the Conclusion's transferability statement is unsu","section":"§4.2, §4.3, §4.4"},{"comment":"The reported accuracy lacks the quantitative detail needed to assess the claims. The parity plots appear to show train and test points together, and the RMSE values (≈0.1 and ≈70 W) are given without normalization or scale context: what is the composition unit (mole fraction?) and what is the range of powers in the dataset? The repeated 5-fold split provides only an average; no error bars, confidence intervals, or per-element residuals are shown. Please report separate train/test RMSE with standard deviations, describe the composition and power ranges, and plot test points separately from training points. This is necessary for the 'accuracy' claim in the abstract and conclusion.","section":"§4.3, §4.4, Figures 10, 11"},{"comment":"The element-independent descriptor set is under-specified. The text names 'cathode operating mode' and 'sputtering yield' but does not give the full feature vector, the energy at which sputtering yield is evaluated, or an explicit statement that categorical element identity variables are excluded. Without this information the model is not reproducible, and the claim of element independence cannot be checked. Please provide a complete descriptor table, the normalization procedure, and a code/data release. A helpful additional check would be to train on a subset of elements and show that the learned mapping remains physically sensible for a held-out element, even if the eventual target is one of the seven studied elements.","section":"§4.2"}],"minor_comments":[{"comment":"Typo: 'Thises ways' should be 'These ways'. Also 'reliance' in §4.3 should be 'reliability'.","section":"§II"},{"comment":"The phrase 'test/train graph' or 'test/train batch' is misleading. Please distinguish training and test points clearly, for example with different markers and a separate test-only panel, and state the number of points in each set.","section":"§4.4, Figures 10, 11"},{"comment":"Axes of the learning-curve figure are unlabeled. Define CV RMSE in the caption and indicate which line corresponds to which model.","section":"Figure 14"},{"comment":"'Data are available from the corresponding author upon reasonable request' is not sufficient for reproducibility of the ML results. Please deposit the dataset and analysis code in a public repository, or specify any legal/ethical restrictions preventing release.","section":"Data availability"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is partly a project/program update and partly a research paper; the ML evaluation is currently thin relative to the strength of the claims. If leave-one-element-out testing and a simple physical baseline cannot be added, the transferability claim should be explicitly downgraded to within-dataset interpolation, and the paper might be more appropriate as a perspective or platform description. The fit with cond-mat.mtrl-sci is acceptable if the scientific content is strengthened along the lines above."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, the short version: there is a real, reproducible experimental dataset here and a sensible first prototype, but the paper's headline claim of a generic element-independent framework is not supported by the evaluation. The paper deserves a serious referee, but the authors need to add a held-out element test and a baseline before I'd believe the transferability.\n\nWhat's new: a 7-element, 82-point dataset from a hybrid HiPIMS/pulsed-DC 4-cathode combinatorial PVD system, plus forward and inverse models (RF, XGBoost, Bayesian ridge, fuzzy) that map powers to composition and back. The inverse problem is handled with bounded power windows. The in-sample test/train plots show reasonable scatter, RMSE ~0.1 in composition and ~70 W in power, and the learning curves converge. That's a legitimate proof of concept on their own equipment, and the paper is honest that the model currently focuses on composition, not the corrosion or mechanical properties named in the title.\n\nThe soft spot is the element-independence claim. Section 4.2 says the model uses cathode mode and sputtering yield instead of element identity, so it should transfer to new elements. But the evaluation uses a random 80/20 split with 5 repeats, which means every element appears in training. With seven elements, sputtering yield is close to a unique tag; the model may just be interpolating among known elements. The sentence in 4.4 that the methodology can be applied to elements not in the dataset is unsupported. The stress-test note is right: no leave-one-element-out, no per-element errors, no comparison to a simple sputter-yield scaling baseline. This is the load-bearing claim and it is untested. Also, no code or data are released, no error bars on the metrics, and the correlation-matrix comparison is qualitative.\n\nNone of this is disqualifying; all are addressable. The narrow claim, that a first prototype can guide composition on this platform, holds. The generic-framework claim does not yet.\n\nWho gets value: process engineers working on combinatorial sputtering and anyone teaching how to frame ML claims in materials science. The paper would be a good reading-group example of overclaim vs evidence.\n\nRecommendation: send to peer review with a request for major revision. Ask for a held-out element test, a baseline, error bars, and code/data deposit.","headline":"A plausible in-sample proof of concept for ML-guided composition control in multicathode PVD; the element-independence claim outruns the evidence.","tokens_in":11733,"tokens_out":2276,"would_cite":false,"duration_ms":24702,"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 single machine-learning model predicts the sputtering parameters needed to deposit a target alloy composition, and recovers composition from process settings.","keywords":["high-entropy alloys","complex concentrated alloys","combinatorial magnetron sputtering","inverse design","machine learning","element-independent model","sputtering yield","nuclear coatings"],"falsifier":"Train the element-independent model on any six of the seven elements, then use it to predict the cathode powers for a target composition containing only the held-out seventh element, deposit that coating, and measure the actual composition; if the error is comparable to or larger than the reported ~0.1 RMSE rather than the in-distribution error, the element-independence claim fails.","tokens_in":10886,"feed_emoji":"🎯","tokens_out":3784,"duration_ms":38434,"temperature":0.7,"pith_summary":"The paper claims that a generic, element-independent machine learning model can replace the usual trial-and-error loop in combinatorial magnetron sputtering: given the desired chemical composition of a complex or high-entropy alloy coating, it predicts the power to apply on each cathode, and given the powers it predicts the resulting composition. The model uses physical descriptors such as sputtering yield and cathode operating mode instead of element identity, so it should transfer to new alloy systems without retraining on a new dataset. The authors demonstrate this on a 7-element, 82-point dataset with forward prediction error around 0.1 atomic fraction and backward prediction error around 70 W. If correct, this would sharply cut the experimental effort in exploring complex compositional spaces for nuclear and low-carbon energy coatings.","feed_headline":"AI predicts sputter power for target alloy coatings","feed_subtitle":"One model maps composition to process settings and back, so new alloy systems need far fewer experiments.","key_machinery":"The element-independent descriptor encoding: each deposition event is described by the cathode operating mode (HiPIMS or pulsed-DC) and normalized physical parameters of the sputtered element, chiefly the sputtering yield, instead of chemical identity. This is what lets a single Random Forest/XGBoost model serve any alloy system; the inverse problem is made tractable by predicting power within user-specified windows, which bounds the otherwise infinite set of power combinations giving the same composition.","core_discovery":"The central claim is the feasibility of a first prototype generic machine learning framework for inverse design in hybrid HiPIMS/pulsed-DC PVD. Rather than training one model per alloy system, the authors encode each cathode by its operating mode and each element by normalized physical parameters, notably sputtering yield. On this basis, Random Forest and XGBoost models learn both the forward map (power -> composition) and the backward map (composition -> power), the latter regularized by bounding the predicted power windows to handle its non-injective nature. The model is validated on a dataset of 82 depositions spanning seven elements (Cr, Al, Mo, Ni, V, W, Nb), with correlations between p","pith_inferences":["The strongest test of element-independence would be a leave-one-element-out experiment: train on six elements and ask the model to predict the power for a seventh held-out element; the paper does not report this, so transferability to truly unseen elements remains an open claim.","Sputtering yield alone may not separate elements whose yields are similar but whose plasma chemistry differs (e.g., target poisoning, secondary electron emission, gas rarefaction); adding plasma diagnostics as descriptors could make the model robust to such cases.","The same descriptor-based inverse-design approach could be carried over to reactive sputtering of oxides or nitrides, where composition targets become stoichiometry targets and the descriptor set would need to include reactive gas flow and target state.","Because the backward map is surjective, the model as presented gives one recipe among many; a human or optimisation loop must impose additional constraints (cost, rate, stress) to select the best recipe."],"forward_implications":["New alloy systems can be explored by swapping target elements and reusing the same model, requiring only a handful of calibration depositions rather than a full retraining campaign.","The inverse map gives experimentalists a concrete starting recipe for a target composition, reducing the number of deposition runs needed to land on a specified high-entropy alloy.","The correlation matrices show the model captures cross-cathode plasma interactions, so it can be used to anticipate how changing one cathode's power perturbs the whole film composition.","The same descriptor logic can be extended to predict microstructural and functional properties once grain size, texture, stress, and deposition rate are added as targets.","The approach supports the long-term goal of closed-loop autonomous deposition, where in-situ diagnostics feed corrections back into the model in real time."],"fun_headline_variants":["AI predicts sputter power to design high-entropy alloys","Machine learning inverts alloy coating recipe for extreme service","Model learns alloy-to-process map, slashing PVD trial runs","From composition to cathode power: AI steers alloy deposition","AI reverse-engineers PVD parameters for new alloy coatings"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The model's transferability rests on the assumption that cathode operating mode plus normalized sputtering yield capture enough of the deposition physics to predict any element combination, yet no element is held out from the training set to verify this.","fun_headline_variants_meta":{"raw":{"variants":["AI predicts sputter power to design high-entropy alloys","Machine learning inverts alloy coating recipe for extreme service","Model learns alloy-to-process map, slashing PVD trial runs","From composition to cathode power: AI steers alloy deposition","AI reverse-engineers PVD parameters for new alloy coatings"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001001,"raw_usage":{"total_tokens":4082,"prompt_tokens":761,"completion_tokens":3321,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":3239}},"tokens_in":505,"tokens_out":3321,"duration_ms":28746,"temperature":1.0,"reasoning_tokens":3239,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T18:31:28.474644+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the element-independent model on any six of the seven elements, then use it to predict the cathode powers for a target composition containing only the held-out seventh element, deposit that coating, and measure the actual composition; if the error is comparable to or larger than the reported ~0.1 RMSE rather than the in-distribution error, the element-independence claim fails.","supporting_citations":[],"review_version":1}