{"id":"b16430e0-a00f-4a70-b957-bc5183e54ce3","arxiv_id":"2505.00210","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A review proposes a four-part functional taxonomy of ML-enhanced adaptive manufacturing control and analyzes where generative models fit, identifying gaps and future research directions.","lead":"This review sorts machine-learning control systems for dynamic manufacturing into four functional types and maps generative AI models onto them. It argues that generative models are promising but currently generate predictions rather than control actions, and it names three gaps that must close.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The gap claim that generative ML in manufacturing is 'primarily predictive' rests on an unsampled, possibly skewed literature subset; a systematic corpus test is needed before the central future-work recommendation can be accepted.","rationale":"The reader's weakest assumption and my load-bearing concern coincide: the review's gap analysis depends on an unstated and possibly non-representative sample of the literature. This is the single most load-bearing issue because the paper's contribution is a classification plus a research-gap narrative, not a new algorithm or theorem. If the literature actually contains a substantial number of generative models used directly as controllers in manufacturing, then the central claim that generation and control remain separate collapses, and the proposed future direction becomes a description of existing work rather than a gap. If the literature does not contain such work, the taxonomy and gap analysis are useful but still need a disclosed methodology to be convincing as a review. The paper also includes internal tension: the robotics transfer examples are direct-policy generative controllers, so the claim that this capability is missing in manufacturing needs a precise statement of what is manufacturing-specific. The concern does not warrant rejection because the taxonomy itself is coherent, the standard equations in Section 4 are presented correctly, and the missing methodology can be supplied by the authors. A conditional acceptance with a request to add a systematic search protocol, inclusion criteria, and a quantitative classification of the surveyed generative-control papers would resolve the issue. Therefore the reader's CONDITIONAL verdict remains appropriate, and I recommend no change to it.","tokens_in":17057,"tokens_out":3552,"duration_ms":37910,"concrete_test":"Conduct a PRISMA-style systematic review: search Scopus, Web of Science, and arXiv from 2018 through 2025 using variants of (generative model OR GAN OR VAE OR diffusion model OR transformer) AND (manufacturing OR additive manufacturing OR process control) AND (control OR MPC OR policy OR controller); apply explicit inclusion/exclusion criteria after title/abstract/full-text screening; and classify each included paper into Prediction-Based, Direct Policy, Quality Inference, or Knowledge-Integrated categories using the Section 3.2 definitions. Report the proportion of screened papers in which the generative model directly produces control actions or policies rather than outputs consumed by a separate controller.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The review's central recommendation, to build integrated frameworks where generative models output control policies (Section 6, direction 1), is motivated by the assertion in Section 5.4 that current integrations 'primarily produce predictive outputs that serve as inputs to separate control systems rather than directly producing control strategies themselves.' That empirical generalization is not supported by a disclosed corpus. Section 3.2's taxonomy (Table 1) contains nine examples, none of which are generative methods in the distribution-modeling sense defined in Section 4 (they are CNNs, RL, SVR, and physics-informed NNs). Section 5 surveys only a handful of generative-integration cases: a GAN-GRU weld pool image predictor [74], a diffusion/VAE/GAN distortion predictor [76], a cGAN surface morphology predictor [77], plus a transformer-DRL scheduling system [75] that is not generative in the paper's own sense. A different sample, e.g., diffusion policies or generative RL controllers applied directly to laser path or process-parameter control, could contain direct-policy generative control, which would undermine the paper's primary gap. Moreover, the paper's own transfer examples from robotics, such as Diffusion Policy [78] and RT-1 [79], are direct-policy generative controllers; the paper therefore undercuts its claim unless it argues manufacturing-specific barriers with more precision. The absence of a search protocol, inclusion criteria, and time/venue scope makes Gap (1) a hypothesis rather than a finding.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a review paper that proposes a functional classification of ML-enhanced control in dynamic manufacturing processes—Prediction-Based, Direct Policy, Quality Inference, and Knowledge-Integrated—and then examines four generative ML architectures (VAEs, GANs, Transformers, Diffusion models) for their control-relevant properties. The central claim is that current integrations of generative ML in manufacturing are primarily predictive, feeding separate control systems rather than directly producing control policies, leading to three research gaps: separation of generation and control, insufficient physical understanding, and domain adaptation challenges. The paper concludes with four future research directions, the first being integrated frameworks in which generative models produce control policies. The standard equations for each generative architecture are presented correctly, and the illustrative applications, while few, are on-topic.","tokens_in":17290,"tokens_out":2445,"duration_ms":26348,"significance":"If the gap analysis were rigorously established, the review would provide a useful roadmap for a growing research area at the intersection of generative modeling and manufacturing control. The paper has strengths: the standard formulations of VAE, GAN, attention, and diffusion are correctly summarized; the proposed functional classification is a sensible organizing scheme; and the concrete applications cited—weld pool forecasting, distortion simulation, surface morphology prediction, and transformer-DRL scheduling—genuinely fit the categories. The authors also explicitly acknowledge limitations of current approaches, which is appropriate for a review. However, the review's load-bearing empirical generalization about the primacy of predictive integrations is not backed by a systematic corpus, and several control-relevant properties are asserted without direct evidence. The central future-work recommendation therefore rests on a claim that the paper's own cited transfer examples partially contradict. The significance of the review is conditional on substantially strengthening the evidence base and sharpening the claims.","major_comments":[{"comment":"The central claim that current generative ML integrations in manufacturing 'primarily produce predictive outputs that serve as inputs to separate control systems rather than directly producing control strategies themselves' is not supported by a disclosed systematic corpus, and the manuscript's own evidence undermines it. Section 5.3 cites Diffusion Policy [78] and RT-1 [79] as transferable approaches, and both are direct-policy generative controllers in the paper's own sense. The authors should either provide a systematic literature search with inclusion criteria demonstrating the predominance of predictive uses in manufacturing specifically, or articulate a precise manufacturing-specific barrier (e.g., safety constraints, sample efficiency, real-time latency) that prevents direct-policy generation in this domain. Without this, the primary future-work recommendation in Section 6, direction 1, is an unsupported empirical generalization.","section":"§5.4, Gap (1); §6, direction 1"},{"comment":"The functional classification is presented as a framework for 'incorporating generative ML,' but the nine examples in Table 1 are all non-generative methods (CNN, RL, SVR, physics-informed NN) in the distribution-modeling sense defined in Section 4. The claim at the end of Section 3.2 that generative ML 'offer solutions through their inherent probabilistic frameworks' is therefore not actually demonstrated by the taxonomy. The authors should either add generative examples to Table 1 (for instance, the generative applications surveyed in Section 5), or explicitly state that the taxonomy currently only covers conventional ML and that extending it to generative models is a hypothesis, not an observed regularity.","section":"§3.2, Table 1; end of §3.2"},{"comment":"There is a terminological inconsistency in the treatment of Transformers. Section 4.3 correctly states that 'Transformers are not inherently generative,' yet Section 5.1 describes the transformer-based scheduling system [75] as 'a generative ML model.' If the authors intend to count auto-regressive sequence generation as an instance of generative modeling, they should say so explicitly and reconcile this with the earlier caveat; otherwise the classification of [75] as a generative integration is misleading.","section":"§5.1 vs. §4.3"},{"comment":"The claim that GANs 'can produce synthetic data that adheres to physical constraints without requiring these constraints to be explicitly encoded' is asserted without citation or demonstration. This is not obvious, and it is in tension with Gap (2) in Section 5.4, which states that current approaches rely on pattern mimicry 'without deeper process understanding.' The authors should either provide evidence for the constraint-adherence claim or temper it to reflect that physical consistency is not guaranteed and must be evaluated.","section":"§4.2, property (1)"}],"minor_comments":[{"comment":"The abstract states that the review 'presents a functional classification' but the classification in Section 3.2 is for ML-enhanced control generally, not for generative ML specifically; a phrase clarifying this scope would prevent over-reading.","section":"Abstract"},{"comment":"The statement that existing studies 'primarily align with the Quality Inference control approaches' is not quantitatively supported; only three generative simulation/digital-twin examples are described. A sentence acknowledging the small sample would be more accurate.","section":"Section 5.2"},{"comment":"Figure 1 is referenced in the text but not present in the provided manuscript; if the figure is common to all architectures, a brief caption description would help readers interpret the properties diagram.","section":"Figure 1"},{"comment":"The phrase 'Even though, in recent times, conventional ML approaches...' is grammatically awkward and could be rewritten as 'Although conventional ML approaches...'.","section":"Section 3.1"}],"recommendation":"major_revision","confidential_remarks":"The review's topic is timely and the writing is generally clear, but the contribution is currently closer to a position paper than a systematic review. The central gap claim needs either a rigorous corpus-based justification or a substantial reframing that explicitly acknowledges the presence of direct-policy generative controllers in adjacent fields and identifies the manufacturing-specific obstacles with concrete evidence. The self-citations [15,16] are used only in the monitoring context and do not appear to bias the taxonomy, but the overall novelty is modest. The manuscript would be a better fit after the evidence-base issue is resolved, since the future-directions section is otherwise a reasonable agenda for the field."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This review does a legitimate service: it gives the field a functional vocabulary for ML-enhanced manufacturing control and maps generative architectures onto that vocabulary. The four categories (Prediction-Based, Direct Policy, Quality Inference, Knowledge-Integrated) are not deep conceptual breakthroughs, but they are reasonable organizing principles that could help researchers position their work. The presentation of VAE, GAN, attention, and diffusion equations is standard and correct, and the nine examples in Table 1 genuinely fit the taxonomy. The identified gaps—separation between generation and control, lack of physical understanding, domain adaptation difficulty—are plausible and worth discussing.\n\nThe soft spots are real but not fatal. Most importantly, the paper's central future-work claim (Section 5.4 and Gap 1) that current generative ML systems \"primarily produce predictive outputs\" rather than control policies is an empirical generalization about the literature, but the review gives no systematic search protocol, inclusion criteria, or time/venue scope. The actual generative-integration examples surveyed are only about four or five papers, and some of the robotics transfer examples cited—Diffusion Policy, RT-1, Janner et al.'s planning—are direct-policy generative controllers. So the claim undercuts itself unless the authors argue specific manufacturing-specific barriers, which they don't do in detail. This gap should be reframed as a hypothesis worth testing, not a finding.\n\nThere are also a couple of unsupported property claims: GANs producing physically constrained data \"without requiring these constraints to be explicitly encoded\" (Section 4.2) and attention providing interpretability (Section 4.3) are asserted without evidence or citation. These are minor but should be either cited or softened.\n\nThe self-citations [15, 16] are fine; they support the in-situ monitoring and transformer discussion and are not load-bearing for the taxonomy.\n\nOverall, the paper is honest and clearly written. It is a review, not a new result, and it does not deserve desk rejection. It deserves a serious referee, but the referee should push for disclosure of the literature selection method and either a broader corpus or a more tentative central claim. If revised, it could be a useful reference for researchers entering this area.\n\nI would send it to peer review with a request for major revision. I wouldn't cite it in its current form, but I might after revision.","headline":"A useful taxonomy and a mostly sound review, but the central gap claim rests on a thin, undisclosed sample and should be softened or substantiated before the paper steers the field.","tokens_in":17802,"tokens_out":1582,"would_cite":false,"duration_ms":17702,"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 review argues that generative machine learning in manufacturing should move from predicting process states to directly issuing control policies in integrated frameworks.","keywords":["generative machine learning","adaptive control","dynamic manufacturing","variational autoencoders","generative adversarial networks","transformers","diffusion models","digital twin"],"falsifier":"A systematic literature search with explicit inclusion criteria that uncovers multiple deployed manufacturing control loops in which a generative model directly outputs control actions in a closed loop—rather than feeding predictions to a separate controller—would undercut the paper's central separation gap. Finding no such cases would support it.","tokens_in":16832,"feed_emoji":"🏭","tokens_out":7560,"duration_ms":70517,"temperature":0.7,"pith_summary":"Generative machine learning—models that learn the probability distribution of data and sample new, realistic examples from it—has been applied to dynamic manufacturing almost entirely as a forecaster. This review claims that such models currently generate predictions, such as future weld-pool images or distortion fields, that are handed to a separate control system, rather than generating the control actions themselves. The paper's contribution is a four-way functional classification of ML-enhanced control—Prediction-Based, Direct Policy, Quality Inference, and Knowledge-Integrated—and a mapping of variational autoencoders, GANs, transformers, and diffusion models onto it. The central thesis is that the next step is to build integrated frameworks in which generative models directly produce control policies while respecting manufacturing constraints. If the thesis is right, adaptive manufacturing control could use the probabilistic reasoning and scenario generation of generative models to handle uncertainty and rare faults that conventional controllers miss.","feed_headline":"Generative models should drive manufacturing control, not just predict it","feed_subtitle":"A control-oriented review maps four routes from probabilistic forecasts to closed-loop policies in dynamic manufacturing.","key_machinery":"The carrying mechanism is the paper's four-way functional classification of ML-enhanced control—Prediction-Based, Direct Policy, Quality Inference, and Knowledge-Integrated—used as a lens for asking how information flows from sensor data to control decisions. The classification does the work of making the generation-control separation visible: when each generative architecture is placed in the taxonomy, its outputs land on the prediction or inference side, not the policy side. The paper pairs this lens with control-relevant properties of the four generative families—latent-space compression and uncertainty bounds for VAEs, implicit distribution learning and synthetic fault generation for GANs, long-range attention and interpretability for transformers, and iterative constraint-guided trajectory generation for diffusion models—to argue that the missing integration is technically plausible.","core_discovery":"The review's core discovery is a functional gap rather than a new algorithm: in current manufacturing practice, generative ML mostly acts as a predictive module inside a larger control loop. Surveying the field through its four functional categories, the paper shows that existing ML-enhanced control either forecasts future states (Prediction-Based), learns state-to-action mappings (Direct Policy), infers unmeasurable quality variables (Quality Inference), or embeds physical knowledge into the model (Knowledge-Integrated). Generative models map onto these categories mainly on the prediction and inference side—GANs and diffusion models synthesize future process images, distortion fields, and surface morphology—while direct policy generation remains rare. The authors therefore assert that the field's next step is to make generative models themselves the control policy generators, combining the uncertainty awareness of generative architectures with the direct-action capability of reinforcement learning, and to do so with physics-informed, purpose-built, computationally tractable models.","pith_inferences":["The review's own selection of nine illustrative examples and four architectures is not justified by a systematic search, so a broader literature scan could change the taxonomy or shift the claimed gap; this is my inference from the absence of inclusion criteria, not a finding the paper reports.","Robotics already demonstrates closed-loop generative policies (for example diffusion-based visuomotor policies), which suggests the transfer barrier to manufacturing may be less about the generative mechanism itself and more about physical process models, safety constraints, and real-time latency.","If integrated generative controllers mature, the four-way classification may need a fifth category for models that simultaneously infer quality and emit actions, since the proposed hybrid direction blurs the boundary between Prediction-Based and Direct Policy control.","A concrete benchmark would compare a diffusion-based or VAE-based policy against a deterministic neural-network policy under out-of-distribution disturbances in a simulated manufacturing process; the generative models' uncertainty accounting should show an advantage exactly where the distribution shifts."],"forward_implications":["If generative models directly emit control policies, adaptive manufacturing systems can respond to in-situ sensor feedback in real time without waiting for a separate controller's prediction-and-optimization step.","Embedding manufacturing physics as explicit constraints in generative architectures would shift the field from pattern mimicry to process understanding, improving reliability of generated trajectories.","Purpose-built generative models designed for manufacturing, rather than architectures adapted from image generation or language processing, would more naturally respect manufacturing-specific constraints and quality requirements.","Model compression and architectural improvements will be needed to reconcile the computational cost of iterative generative methods with real-time manufacturing requirements.","Hybrid frameworks that combine Prediction-Based or Quality Inference strengths with Direct Policy action generation could enable simultaneous optimization of quality, efficiency, and adaptability."],"supporting_citations":[{"why":"Supplies the Prediction-Based control category: a 3D CNN autoencoder predicts and compensates geometric deformation in polymer printing.","marker":"[42]"},{"why":"Exemplifies Direct Policy control: deep RL optimizes feed rate and heat input in wire arc additive manufacturing.","marker":"[44]"},{"why":"Anchors the Quality Inference category: virtual metrology estimates wafer thickness during semiconductor manufacturing.","marker":"[47]"},{"why":"Anchors the Knowledge-Integrated category: a physics-informed recurrent network couples physical laws with online parameter estimation.","marker":"[49]"},{"why":"Main current-integration example: a GAN-GRU generates future weld-pool images for a human-centered model predictive control system.","marker":"[74]"},{"why":"Limited example of direct policy generation: a transformer with deep RL maps production states to scheduling decisions.","marker":"[75]"},{"why":"Generative simulation/digital twin example: a diffusion-based framework with VQVAE and GAN predicts distortion fields in wire arc additive manufacturing.","marker":"[76]"},{"why":"Generative quality prediction example: a conditional GAN predicts surface morphology in directed energy deposition.","marker":"[77]"},{"why":"Transferable robotics approach: diffusion models learn real-time visuomotor policies, informing manufacturing parameter adjustment.","marker":"[78]"},{"why":"Transferable language-model approach: large language models handle real-time task planning and resource allocation in modular production.","marker":"[84]"}],"fun_headline_variants":["Generative ML should drive manufacturing control, not just predict","From prediction to policy: generative models in manufacturing control","Generative models as controllers: a new direction for manufacturing","Shift generative ML from prediction to control in dynamic manufacturing","When generative models become the controllers in manufacturing loops"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The review's taxonomy and gap analysis assume that the four generative architectures and the nine illustrative control examples in Table 1 faithfully represent the whole field of generative ML for manufacturing control, but no systematic search or inclusion criteria are given.","fun_headline_variants_meta":{"raw":{"variants":["Generative ML should drive manufacturing control, not just predict","From prediction to policy: generative models in manufacturing control","Generative models as controllers: a new direction for manufacturing","Shift generative ML from prediction to control in dynamic manufacturing","When generative models become the controllers in manufacturing loops"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000569,"raw_usage":{"total_tokens":2693,"prompt_tokens":944,"completion_tokens":1749,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":1672}},"tokens_in":560,"tokens_out":1749,"duration_ms":12810,"temperature":1.0,"reasoning_tokens":1672,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:47:25.073143+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A systematic literature search with explicit inclusion criteria that uncovers multiple deployed manufacturing control loops in which a generative model directly outputs control actions in a closed loop—rather than feeding predictions to a separate controller—would undercut the paper's central separation gap. Finding no such cases would support it.","supporting_citations":[{"cited_title":"Alearning-basedframe- workforerrorcompensationin3Dprinting","cited_arxiv_id":null,"evidence_quote":"Supplies the Prediction-Based control category: a 3D CNN autoencoder predicts and compensates geometric deformation in polymer printing."},{"cited_title":"Optimal data-driven control of manufacturing processes usingreinforcementlearning: anapplicationtowirearcad- ditivemanufacturing","cited_arxiv_id":null,"evidence_quote":"Exemplifies Direct Policy control: deep RL optimizes feed rate and heat input in wire arc additive manufacturing."},{"cited_title":"Avirtualmetrologysystemforsemiconductormanu- facturing","cited_arxiv_id":null,"evidence_quote":"Anchors the Quality Inference category: virtual metrology estimates wafer thickness during semiconductor manufacturing."},{"cited_title":"Physics-informed online machine learning and predictive control of nonlinear pro- cesseswithparameteruncertainty","cited_arxiv_id":null,"evidence_quote":"Anchors the Knowledge-Integrated category: a physics-informed recurrent network couples physical laws with online parameter estimation."},{"cited_title":"Gen- erative adversarial networks (GAN) model for dynamically adjusted weld pool image toward human-based model pre- dictive control (MPC)","cited_arxiv_id":null,"evidence_quote":"Main current-integration example: a GAN-GRU generates future weld-pool images for a human-centered model predictive control system."},{"cited_title":"Atransformer-baseddeepreinforcement learningapproachfordynamicparallelmachinescheduling problem with family setups","cited_arxiv_id":null,"evidence_quote":"Limited example of direct policy generation: a transformer with deep RL maps production states to scheduling decisions."},{"cited_title":"Onlinedistor- tion simulation using generative machine learning models: A step toward digital twin of metallic additive manufactur- ing","cited_arxiv_id":null,"evidence_quote":"Generative simulation/digital twin example: a diffusion-based framework with VQVAE and GAN predicts distortion fields in wire arc additive manufacturing."},{"cited_title":"Virtual surface morphology generation of Ti-6Al-4V di- rected energy deposition via conditional generative adver- sarial network","cited_arxiv_id":null,"evidence_quote":"Generative quality prediction example: a conditional GAN predicts surface morphology in directed energy deposition."},{"cited_title":"Diffusion policy: Visuomotor policy learning via action diffusion","cited_arxiv_id":null,"evidence_quote":"Transferable robotics approach: diffusion models learn real-time visuomotor policies, informing manufacturing parameter adjustment."},{"cited_title":"Towards autonomous system: flexible modu- lar production system enhanced with large language model agents","cited_arxiv_id":null,"evidence_quote":"Transferable language-model approach: large language models handle real-time task planning and resource allocation in modular production."}],"review_version":1}