{"id":"89c4f1e3-e1da-4367-af56-1ac9c9ea48f0","arxiv_id":"2507.10983","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":0.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"This review organizes machine learning and physics-informed neural network applications to semiconductor film deposition into four categories and outlines future research directions.","lead":"Semiconductor chip makers need precise control over the thin films deposited onto wafers, and this paper reviews how machine learning and physics-informed neural networks are being used for that task. It groups existing work into four application areas and lists open problems and future research directions.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'ALD precursor finding' category—one of the four pillars of the review's taxonomy—rests on a single cited study [28] that is about NLP opinion mining in nanomaterials, not about ALD precursors, so the central 'comprehensive review' claim is not supported as written.","rationale":"The reader identified the general weakness correctly: the review claims comprehensiveness without a systematic search strategy, and the cited studies might not fairly represent the literature. My stress-test sharpens that worry into a specific, verifiable instance: Section 3.4's 'ALD Precursor Finding' category is anchored by a single reference that, by the paper's own description, is an NLP opinion-mining study about energy materials, not an ALD precursor discovery study. This is more damaging than a mere citation error because it directly affects the structure of the central taxonomy: if the only exemplar in a category is misclassified, that category is not a literature-derived theme. The duplicate reference [37] in Section 3.1 and the missing 'Table 4' reinforce the pattern of careless support. I do not see this as grounds for rejection: a review's taxonomy can be repaired by adding a genuinely systematic search, re-labeling or removing unsupported categories, and correcting the citation errors. The paper's qualitative explanation of PINN loss-weighting and sampling methods in Section 4 remains plausible and could be useful as an introduction, even though none of the film-deposition studies reviewed actually uses PINNs. Therefore the verdict stays CONDITIONAL, matching the reader's assessment; my concrete check is designed to determine whether the ALD category can be rescued by additional literature or must be dropped.","tokens_in":14002,"tokens_out":6112,"duration_ms":72763,"concrete_test":"Run a targeted literature search in Scopus or Web of Science using variants of TITLE-ABS-KEY((('atomic layer deposition' OR ALD) AND ('precursor' OR 'precursor screening') AND ('machine learning' OR 'deep learning' OR 'neural network'))), and compare the retrieved studies with those cited in Section 3.4 and Table 3. Then retrieve the full text of reference [28] and search for 'ALD', 'atomic layer deposition', and 'precursor'. If the search returns multiple ALD-precursor ML studies that are not cited, or if the full text of [28] contains no ALD-precursor content, then Section 3.4's category is unsupported and the 'comprehensive' claim needs revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central contribution is the four-category taxonomy in Section 3, which is claimed to emerge from 'a comprehensive review of the literature.' Category 4, 'ALD Precursor Finding,' is supported by exactly one entry: Table 3, row [28], summarized as a 'novel knowledge-mining method using opinion-mining from scientific literature.' Reference [28] (Xie et al., J. Chem. Inf. Model. 64(7):2746–2759) is titled 'Opinion Mining by Convolutional Neural Networks for Maximizing Discoverability of Nanomaterials'; the paper's own description says it extracts and classifies expert opinions about challenges and opportunities in energy materials. Nothing in the title, reference, or summary connects it to identifying ALD precursors. No other study is cited in Section 3.4, so the fourth thematic area is effectively an empty label. This is not a matter of disagreeing with an outside consensus; it is an internal failure of support. A second concrete instance appears in Section 3.1, where reference [37] is presented as a separate CFD+ANN ALD study, but [37] is a duplicate of [32] (Cho, Shao, and Mesbah, the RI-TVBO paper), so that paragraph also contains one fewer independent supporting study than advertised. Additionally, the text says 'Tables 1–4' but only three tables appear, and Table 3 merges the maintenance and ALD-precursor categories. Without a stated search strategy, inclusion criteria, or quality assessment, these specific mischaracterizations mean the taxonomy cannot be trusted as a representative map of the literature; a reader cannot tell which categories are robust and which are artifacts of a small, non-systematic sample. The central claim of comprehensiveness therefore does not survive the evidence currently in the manuscript.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript presents a review of machine learning and deep learning applications in semiconductor film deposition, with a focus on the potential of physics-informed neural networks (PINNs). The authors propose a four-category taxonomy of existing work: (1) process control and optimization, (2) defect image recognition and classification, (3) tool preventative maintenance and anomaly detection, and (4) atomic layer deposition (ALD) precursor finding. They also survey PINN loss-weighting and adaptive sampling techniques and propose future research directions that integrate PINNs with graph neural networks and system-level decision making.","tokens_in":14184,"tokens_out":1730,"duration_ms":19533,"significance":"If the taxonomy and literature survey are reliable, the paper would provide a useful structured entry point for researchers seeking to apply physics-informed machine learning to film deposition. The PINN methodology summary (Section 4) is generally accurate and correctly cites foundational works on loss reweighting (e.g., Wang et al., Maddu et al., McClenny et al.) and adaptive sampling (Nabian et al., Katharopoulos et al., Tang et al.), with equations that match the original sources. The proposed future directions (Section 5) are plausible and align with current trends in physics-informed machine learning. The paper does not introduce new numerical results, and its claims are explicitly descriptive rather than predictive, so there is no circularity concern.","major_comments":[{"comment":"The 'ALD Precursor Finding' category is supported by exactly one cited study, [28] (Xie et al., J. Chem. Inf. Model. 64(7):2746–2759). The paper's own description of [28] says it develops SSNet, a model for extracting and classifying expert opinions about challenges and opportunities in energy materials from scientific literature. Nothing in the reference's title, venue, or the paper's summary connects it to ALD precursor identification. Since this is the only entry in the fourth category, the category is effectively empty, and the claim in Section 1 and the abstract of a 'comprehensive review' organized into four thematic areas is not supported as written.","section":"Section 3.4 and Table 3"},{"comment":"Reference [37] is a duplicate of reference [32]: both are Cho, Shao, and Mesbah, 'Run-indexed time-varying Bayesian optimization with positional encoding for auto-tuning of controllers: Application to a plasma-assisted deposition process with run-to-run drifts', Computers & Chemical Engineering, vol. 185, p. 108653 (2024). The sentence citing '[30, 37]' as 'novel data-driven modeling frameworks combining Computational Fluid Dynamics (CFD) and ANNs' therefore cites the same study twice, so the paragraph contains one fewer independent supporting study than advertised. This duplication undermines the credibility of the survey's reference handling.","section":"Section 3.1 and Table 1"},{"comment":"The paper repeatedly describes the review as 'comprehensive' (Abstract, Section 1, Section 3) but provides no search protocol, inclusion criteria, database list, or quality assessment. Without a stated methodology, the reader cannot assess whether the four-category taxonomy is representative of the broader literature or reflects a selective sample. The incorrect ALD category and the duplicate reference compound this concern. The authors should either add a transparent methodology section or reposition the paper as a scoping review with explicit limitations.","section":"Section 1 and Section 3"}],"minor_comments":[{"comment":"The text says 'A summary of the reviewed DL studies in film deposition is provided in Tables 1–4,' but the manuscript contains only three tables (Tables 1, 2, and 3). Please correct the cross-reference and ensure the table numbering matches the narrative.","section":"Section 3.5"},{"comment":"Equation (8) defines the loss as L = J + w_f L_f + w_B L_B + w_h L_h, but the subsequent sentence says 'L_f, L_B, and L_d represent the PDE residual loss, boundary condition loss, and data loss, respectively.' The symbol L_d does not appear in the equation, and L_h is not described. Please align the notation.","section":"Section 4.1.1, Eq. (8)"},{"comment":"The reference list contains inconsistent formatting, including incomplete entries (e.g., Ref. [8] with an incomplete author list and odd punctuation) and at least one duplicate as noted in Major Comment 2. A thorough editorial pass is needed.","section":"References"},{"comment":"The future-directions section mentions 'Continuous Graph Neural Networks (CGNNs)' but the cited Ref. [62] is a survey on causal graph neural networks; the connection between CGNNs and causality is not elaborated enough to justify the claim that CGNNs 'introduce causal inference capabilities' in this context.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's core idea—surveying ML/DL for film deposition and connecting it to PINNs—is timely, and the PINN methodology summary is competent. However, the review's central organizing claim (a comprehensive four-category taxonomy) is undercut by a category with no genuine supporting evidence and by a duplicate reference. These are fixable: the authors could add genuinely relevant ALD precursor work, remove or reframe the fourth category, and add a methodology section making the review's scope explicit. If they do not, the 'comprehensive' claim should be abandoned. I would not reject the paper outright because the PINN tutorial portion has value, but the review's integrity needs substantial repair."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read this as an introduction to ML for film deposition, not as a research contribution. The useful part is the PINN section: the loss-reweighting and adaptive sampling descriptions are accurate summaries of known methods (Wang, Maddu, McClenny, etc.), and the equations are standard. The thematic grouping of process control, defect detection, and predictive maintenance is a reasonable way to organize the papers it does cite.\n\nBut the central claim of comprehensiveness does not hold. The fourth category, ALD precursor finding, rests on exactly one citation, [28], and that paper is about opinion mining of nanomaterials literature, not about identifying ALD precursors. The authors' own description confirms it—SSNet classifies expert opinions; nothing in the title or summary connects it to precursor discovery. So that category is an empty label. Table 3 awkwardly merges maintenance and ALD, which hints the authors felt the thinness themselves. On top of that, [37] is a duplicate of [32], leaving the process-control section with one fewer independent study than advertised, and the text references 'Tables 1–4' when only three tables exist.\n\nThese are not nitpicks for a review whose only product is a structured map of the literature. If the map mislabels a major quadrant and cites papers inaccurately, readers cannot trust the categories or the abstracts. The paper also reports no search strategy or inclusion criteria, so the 'comprehensive' descriptor is unsupported regardless of the specific errors. That said, the PINN portion is sound and would be useful to someone new to the field.\n\nMy verdict: conditional pass at best. If the authors fix the citation errors, either fill or remove the ALD category, and add a methodology section describing how studies were selected, the review could serve as a serviceable introduction. As written, it needs major revision before I'd rely on it. I would not cite it in its current form. A serious referee could help the authors get there, so I'd send it out, but with the expectation of a demanding report.","headline":"A well-organized review with a solid PINN overview, but the taxonomy has a hollow category and citation errors that undercut the 'comprehensive' claim.","tokens_in":14825,"tokens_out":2251,"would_cite":false,"duration_ms":26571,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Machine learning in semiconductor film deposition clusters into four application areas, and the paper argues physics-informed neural networks are the route to more data-efficient, interpretable process control.","keywords":["Physics-informed neural networks","semiconductor manufacturing","film deposition","chemical vapor deposition","atomic layer deposition","defect recognition","process control","predictive maintenance"],"falsifier":"A systematic search with explicit inclusion criteria over the same period that finds a substantial number of deposition-ML papers outside the four categories would falsify the review's comprehensiveness claim; re-reading the cited papers and finding that the tabulated accuracies or model descriptions do not match the originals would falsify the summary layer.","tokens_in":1770,"feed_emoji":"🧠","tokens_out":10183,"duration_ms":159028,"temperature":0.7,"pith_summary":"This review examines machine-learning and deep-learning studies aimed at semiconductor film deposition (CVD, PVD, and ALD) and organizes them into four application areas: process control and optimization, defect image recognition, tool maintenance and anomaly detection, and ALD precursor finding. It argues that across these areas, current models share structural weaknesses: simplified reaction physics, heavy reliance on large labeled datasets, limited interpretability, and poor generalization beyond training conditions. The paper's constructive claim is that physics-informed neural networks (PINNs), which train networks to respect governing equations as well as data, fit film deposition naturally because the underlying transport and reaction physics are known. It surveys the main technical levers for making PINNs practical (loss reweighting and adaptive resampling) and proposes future directions including combining PINNs with graph networks and treating machine learning as a system-level decision tool.","feed_headline":"Review proposes physics-guided AI for chip film deposition","feed_subtitle":"It sorts the surveyed studies into process control, defect detection, tool health, and ALD precursor search.","key_machinery":"The load-bearing object is the PINN loss function, written in the paper as $L = w_f L_f + w_B L_B + w_d L_d$, where $L_f$ penalizes violation of the governing PDE, $L_B$ enforces boundary or initial conditions, and $L_d$ fits observed data, with weights $w_f$, $w_B$, and $w_d$ balancing the terms. The paper's argument that PINNs can help film deposition rests on this object: with physics embedded as a loss term, a network can be trained on sparse or noisy fab data while still respecting known transport and reaction laws. The second piece of machinery is the four-category taxonomy in Section 3, which gives the review its claim to being a structured map rather than an annotated bibliography. The named class of objects central to the review is the Physics-Informed Neural Network, a neural network trained to satisfy a differential equation, its boundary conditions, and available data simultaneously; the review also catalogues the main methods for balancing those objectives.","core_discovery":"On its own terms, the paper establishes a taxonomy: the machine-learning literature on film deposition in semiconductor manufacturing falls into four clusters: process control and optimization, defect image recognition and classification, tool preventive maintenance and anomaly detection, and ALD precursor finding. It then argues that the dominant limitations across clusters are not purely algorithmic but structural: PDE-based models oversimplify surface chemistry, process parameters resist interpretation, data are sparse and expensive, and models do not generalize beyond the conditions they were trained on. The proposed resolution is to embed physical laws into neural networks so that predictions are constrained by known physics rather than learned purely from data. For PINNs to work in this setting, the paper contends, the central technical problem is balancing the PDE residual loss against boundary and data losses; it reviews the two families of remedies, loss reweighting (gradient annealing, inverse Dirichlet weighting, self-adaptive masks, minimax formulations) and adaptive collocation-point resampling (quasi-random sequences, importance sampling, deep generative samplers). If the argument is right, the payoff is process control that needs less labeled data and is more trustworthy when extrapolating to new process conditions.","pith_inferences":["The authors stop short of saying it, but the framework implies the need for a standard benchmark: a public deposition PDE model, such as a CVD boundary-layer or ALD surface-reaction model, with shared data so that loss-weighting and sampling methods can be compared on equal footing.","A testable extension of the taxonomy is to run a formal systematic search with explicit inclusion criteria; if most papers still fall into the four categories, the review's structure is validated, and if not, the map needs revision.","The physics-constrained agenda could apply to adjacent fab processes the paper mentions only in passing (etching, lithography, and thermal processing), where the same data-sparsity and interpretability problems occur.","For practitioners, the strongest implicit message is that the bottleneck is not network architecture but loss engineering: choosing how to balance PDE, boundary, and data terms may matter more than network depth."],"forward_implications":["A researcher entering the area gets a ready-made map: process control and optimization, defect image recognition, tool maintenance and anomaly detection, and ALD precursor finding, each with representative model families and reported results.","If PINN-based physics constraints work as argued, deposition models can be trained with far fewer wafer runs, since the PDE residual supplies information that labeled data would otherwise have to provide.","The loss-reweighting and adaptive-sampling techniques surveyed for benchmark PDE problems are the practical entry point for building deposition-specific PINNs.","Combining PINNs with graph neural networks and continuous graph networks is proposed as the route to spatial-temporal process models that fuse multimodal fab data with physical constraints.","The review implies that machine learning should be repositioned from a single-task predictor to a system-level decision integrator for routine fab operations, an underexplored direction."],"supporting_citations":[{"why":"Defines physics-informed neural networks as the framework the review's proposals build on.","marker":"[10]"},{"why":"Supplies the DeepSEM-Net defect-segmentation result that anchors the defect recognition category.","marker":"[25]"},{"why":"Provides the transfer-learning wafer-map defect model used as a representative of the defect category.","marker":"[27]"},{"why":"Supplies the SSNet opinion-mining model that defines the ALD precursor finding category.","marker":"[28]"},{"why":"Provides the CFD-plus-neural-network SiO2 ALD model that anchors the process control and optimization category.","marker":"[30]"},{"why":"Supplies the BRANN PEALD HfO2 modeling study used to show physics-simulation-plus-ML coupling.","marker":"[31]"},{"why":"Provides the run-indexed time-varying Bayesian optimization method for controller auto-tuning under drift.","marker":"[32]"},{"why":"Supplies the ACO/BPNN CVD optimization study with reported deposition-rate and uniformity gains.","marker":"[38]"},{"why":"Provides the DBSCAN/LSTM anomaly detection study for the tool maintenance category.","marker":"[39]"},{"why":"Supplies the gradient-statistics learning-rate annealing method that is a key loss-reweighting technique in the PINN review.","marker":"[48]"}],"fun_headline_variants":["Physics-informed AI review maps chip film deposition ML","PINN review: physics-guided AI for semiconductor deposition","Review: embedding physics in nets to fix film deposition gaps","Four ML niches in film deposition, one physics-based fix","Chip film deposition ML: taxonomy and PINN path forward"],"cache_read_input_tokens":16896,"weakest_assumption_plain":"The review's conclusions depend on the assumption that the cited studies in Sections 3.1-3.4 and Tables 1-3 fairly represent the broader literature and are summarized accurately, since no search strategy, inclusion criteria, or quality assessment is reported.","fun_headline_variants_meta":{"raw":{"variants":["Physics-informed AI review maps chip film deposition ML","PINN review: physics-guided AI for semiconductor deposition","Review: embedding physics in nets to fix film deposition gaps","Four ML niches in film deposition, one physics-based fix","Chip film deposition ML: taxonomy and PINN path forward"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000212,"raw_usage":{"total_tokens":1447,"prompt_tokens":1005,"completion_tokens":442,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":621,"completion_tokens_details":{"reasoning_tokens":363}},"tokens_in":621,"tokens_out":442,"duration_ms":5179,"temperature":1.0,"reasoning_tokens":363,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:19:22.806669+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A systematic search with explicit inclusion criteria over the same period that finds a substantial number of deposition-ML papers outside the four categories would falsify the review's comprehensiveness claim; re-reading the cited papers and finding that the tabulated accuracies or model descriptions do not match the originals would falsify the summary layer.","supporting_citations":[{"cited_title":"Physics-informed neural networks: A deep learning framework for solving forward and inverse problems in- volving nonlinear partial differential equations","cited_arxiv_id":null,"evidence_quote":"Defines physics-informed neural networks as the framework the review's proposals build on."},{"cited_title":"Deepsem- net: Enhancing sem defect analysis in semiconductor manufacturing with a dual-branch cnn-transformer archi- tecture","cited_arxiv_id":null,"evidence_quote":"Supplies the DeepSEM-Net defect-segmentation result that anchors the defect recognition category."},{"cited_title":"Deep transfer wasserstein adversarial network for wafer map defect recognition","cited_arxiv_id":null,"evidence_quote":"Provides the transfer-learning wafer-map defect model used as a representative of the defect category."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the SSNet opinion-mining model that defines the ALD precursor finding category."},{"cited_title":"Machine learning-based model- ing and operation for ald of sio2 thin-films using data from a multiscale cfd simulation","cited_arxiv_id":null,"evidence_quote":"Provides the CFD-plus-neural-network SiO2 ALD model that anchors the process control and optimization category."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the BRANN PEALD HfO2 modeling study used to show physics-simulation-plus-ML coupling."},{"cited_title":"Optimizing the chemical vapor deposition process of 4h–sic epitaxial layer growth with machine-learning-assisted multiphysics simulations","cited_arxiv_id":null,"evidence_quote":"Supplies the ACO/BPNN CVD optimization study with reported deposition-rate and uniformity gains."},{"cited_title":"When magnetron sput- tering deposition meets machine learning: Application to process anomaly detection","cited_arxiv_id":null,"evidence_quote":"Provides the DBSCAN/LSTM anomaly detection study for the tool maintenance category."},{"cited_title":"Understand- ing and mitigating gradient flow pathologies in physics- informed neural networks","cited_arxiv_id":null,"evidence_quote":"Supplies the gradient-statistics learning-rate annealing method that is a key loss-reweighting technique in the PINN review."}],"review_version":1}