{"id":"f8ec798b-ad43-4496-a6da-20111cd5fd77","arxiv_id":"2507.01194","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An autonomous STEM workflow with machine-learning atom decoding and FPGA-controlled beam targeting reliably creates metallic MoS nanowire edge structures in MoS2 monolayers.","lead":"This paper demonstrates a fully autonomous scanning transmission electron microscopy (STEM) system that uses machine learning to identify atoms and a programmable electron beam to selectively remove sulfur atoms, creating tailored edge structures in monolayer MoS2. It matters because it shows a path toward atom-scale automated manufacturing of defect structures that could eventually serve quantum and electronic devices.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The small-data ML classifiers in the autonomous feedback loop are never quantitatively validated; without precision/recall metrics across fabrication stages and pixel sizes, the 'reliable' targeting claim is unsupported.","rationale":"The reader's weakest-assumption analysis identifies exactly the same load-bearing concern: the unquantified generalization of the small-data ML classifiers. My reading of the manuscript confirms that no quantitative accuracy, precision, recall, or confusion-matrix data are provided anywhere, despite the persistent claim of 'reliable' classification. The paper does provide code and detailed methods, which is commendable, but the central claim of fully autonomous fabrication cannot be supported without evidence that the ML models perform correctly on the actual data distributions encountered during fabrication. The proposed concrete test would settle this by directly measuring classification performance on held-out experimental images from different stages and by tracking live targeting decisions against manual labels. If the metrics are strong, the conditional verdict may be upgraded to accept; if they are weak, the central claim fails. Until such evidence is provided, the conditional verdict is appropriate, and I do not see a reason to change it. I considered whether the absence of overall fabrication success-rate statistics is an even more load-bearing concern, but the ML generalization issue is more fundamental: if the classifiers cannot reliably identify the correct atoms, the feedback loop fails regardless of the success-rate statistics of the physical process. Thus I agree with the reader's choice of weakest assumption.","tokens_in":12780,"tokens_out":5410,"duration_ms":65453,"concrete_test":"Hold out a set of at least 5 HAADF images from each fabrication stage (pristine, seeded nanopore, MoS-NW growth) with manually labeled atom identities and NW regions. Run the published ELIT + U-net + random forest pipeline on these held-out images and report per-class precision, recall, and confusion matrices for Mo, S, SVL, and NW, broken down by pixel size. Additionally, during one full autonomous run, log the model's prediction for every targeted site and verify by post-hoc manual inspection whether the beam hit the intended sulfur atom; compute the targeting precision. If S-site recall falls below ~90% or targeting precision below ~80%, the feedback loop's decisions are not sufficiently reliable for the central claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of fully autonomous, reliable fabrication depends entirely on the ML decoding chain: ELIT atom localization, U-net MoS-NW segmentation, and random forest Mo/S/SVL classification. The random forest is trained on only 55 hand-labeled atom neighborhoods (21 Mo, 24 S, 10 SVL) and the U-net on about 30 images, yet the paper asserts 'reliable atom classification at various pixel sizes' without reporting any confusion matrices, precision, recall, or held-out evaluation. This is load-bearing because each beam-targeting decision is made directly from these classifications (Figure 2a). A single S-site misclassified as Mo would cause the beam to avoid the intended sulfur atom or strike molybdenum, potentially creating uncontrolled neighboring vacancies as the paper itself notes for Mo targeting. The U-net segmentation errors also propagate: atoms in a wrongly segmented NW region are automatically labeled NW and skipped by the random forest, masking their true identity. Furthermore, the random forest's nearest-neighbor vectorization depends on ELIT's atom positions; if ELIT mislocates atoms in distorted regions near nanopores, the random forest input is corrupted. Without quantitative validation on images representative of actual fabrication stages (pristine, seeded nanopore, NW growth, different pixel sizes), the 'reliable' fabrication claim rests on anecdotal image sequences rather than measured classifier performance.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript reports an autonomous scanning transmission electron microscopy (STEM) workflow for fabricating tailored MoS-nanowire (MoS-NW) edge structures in monolayer MoS2. The system iteratively acquires HAADF images, decodes them using a machine learning pipeline (ELIT for atom localization, a U-net for MoS-NW segmentation, and a random forest for Mo/S/SVL classification), selects sulfur sites for removal according to predefined fabrication strategies, and exposes those sites with an FPGA-controlled beam. The authors demonstrate three fabrication outcomes: MoS-NW growth along selected nanopore edges, freestanding MoS-NWs between two nanopores, and directional growth of long MoS-NWs across the field of view. The code and training data are made available in an open-source repository.","tokens_in":13171,"tokens_out":3915,"duration_ms":45070,"significance":"If the central claims hold, the work is a meaningful step toward closed-loop atomic-scale manufacturing: it integrates real-time ML-based image decoding with programmable beam control and demonstrates the formation of complex 1D-2D heterostructures. The open-source release of code and data is a strength that supports reproducibility. However, the demonstrations are essentially qualitative. The authors assert 'reliable' and 'fully autonomous' operation, but no quantitative ML validation metrics, yield statistics, or reproducibility data are provided. Because the ML classifications directly determine every beam-targeting decision, the lack of validation of these classifiers is a load-bearing gap. The framework's material-agnostic ambition is plausible but remains a promise rather than a demonstrated transfer.","major_comments":[{"comment":"No quantitative evaluation of the random forest (trained on 55 hand-labeled instances) or the U-net (trained on ~30 images) is reported. There are no confusion matrices, precision/recall values, or held-out validation results across fabrication stages, defect morphologies, or pixel sizes. The Conclusion's statement that the ML framework 'consistently achieved reliable atom classification at various pixel sizes' is therefore unsupported. This is load-bearing because every beam-targeting decision in Figure 2a is made directly from these classifications; a misclassified S site as Mo, or a wrongly segmented NW region, would produce incorrect targeting. The authors should report classification metrics on held-out images representative of pristine, seeded-nanopore, and NW-growth stages at the pixel sizes used, and ideally quantify how classification errors propagate into targeting decisions.","section":"Machine learning models and training; Conclusions"},{"comment":"The random forest uses a vectorization based on the five nearest atoms identified by ELIT, and the Methods section explicitly notes that its performance depends on ELIT accurately locating atom sites. Yet no localization error statistics for ELIT on experimental images are provided, particularly in distorted regions near nanopores and MoS-NWs where atom positions are most challenging. If ELIT mislocates atoms, the random forest input is corrupted and the entire decoding chain degrades. The authors should quantify ELIT detection precision/recall and localization error on representative experimental images, or otherwise demonstrate that the downstream classification is robust to realistic localization noise.","section":"Machine learning models and training"},{"comment":"The central claim of 'reliably' forming tailored MoS-NW structures is supported only by a small number of image sequences, with no statistics on the number of attempts, success rates, or variability across independent runs. Terms such as 'reliably formed' and 'we found this strategy to work relatively well' appear without quantitative support. For a paper whose contribution is an autonomous fabrication method, the authors should report the number of experiments performed for each strategy (targeted growth, freestanding NW, directional growth), the success/failure criteria, and quantitative descriptors of the resulting structures (e.g., NW length, continuity, alignment, or deviation from the intended geometry). This is needed to substantiate the reliability claim and to allow readers to judge the method's practical utility.","section":"Results, Section II; Figure 5"},{"comment":"The paper repeatedly describes the approach as 'fully autonomous,' but the workflow depends on human-predefined fabrication strategies (e.g., the triangular mask geometry, the DBSCAN cluster selection, the ~0.2-0.3 nm distance threshold for targeting S sites) and on the human operator choosing which strategy to apply. The autonomy lies in the execution of a pre-specified recipe, not in the discovery or selection of the strategy. This should be stated precisely and the title/abstract claim qualified accordingly, because the scope of the claim is part of the paper's contribution and current wording risks overstating the advance relative to prior automated feedback systems.","section":"Abstract; Results; Conclusions"}],"minor_comments":[{"comment":"The number of training images is inconsistent: the Conclusions state '~20 images with labeled MoS-NW regions and ~40 labeled atoms,' while the Methods state '~30 labeled MoS2 HAADF images' and a support set of '55 hand-labeled instances' (21 Mo, 24 S, 10 SVL). Please reconcile these numbers.","section":"Conclusions vs. Methods"},{"comment":"The phrase 'In all three subfigures, (i-v) represent raw ADF acquisition and decoded images' is unclear; consider rewriting as 'In each row, panels (i)-(v) show raw ADF images and decoded images with beam targets overlaid, from the beginning to the end of the iterative workflow.'","section":"Figure 5 caption"},{"comment":"The statement 'while S and Mo can potentially be classified by e.g. a simple pixel intensity threshold' is made without a citation or a quantitative comparison; either add a reference or provide a brief justification.","section":"Results, Section I"},{"comment":"The abbreviation 'MoS-NW' is inconsistent with the abstract's 'MoS-nanowire' and appears as 'MoS-NW' elsewhere; standardize the usage.","section":"Throughout"},{"comment":"The repository link is given as a URL only; for archival purposes, please provide a DOI or versioned release, since URLs can change.","section":"Machine learning models and training"},{"comment":"Reference [38] is a conference abstract (Microscopy and Microanalysis) that may not provide sufficient methodological detail; consider citing a full peer-reviewed article if one exists.","section":"References"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe real news here is that the loop is closed: HAADF image in, ML decoding, autonomous beam targeting, and out comes a freestanding MoS-NW between two pores or a MoS-NW grown across the field of view. That is a genuine capability advance over their earlier heuristic-driven work (refs 32, 33). Credit is due: the code and data are on GitHub, the methods are detailed enough to reproduce the workflow, and the demonstrations are compelling as qualitative proof-of-concept.\n\nWhat is new is the ML-based segmentation of SVL and MoS-NW features and the use of that segmentation to drive the decision-making, plus the FPGA spiral scans that give a radially symmetric dose. The paper argues convincingly that small-data models are appropriate for STEM atomic environments, and the uncertainty quantification in Fig. 3 is a nice touch.\n\nThe soft spots are real but not fatal. The central reliability claim is unsupported by statistics. The random forest is trained on 55 hand-labeled atom neighborhoods and the U-net on ~30 images, yet the paper asserts reliable classification at various pixel sizes without reporting precision, recall, or confusion matrices. That matters because the loop makes targeting decisions directly from these classifications. If ELIT mislocates atoms near a distorted nanopore edge, the random forest input is corrupted; if the U-net skips a region, those atoms are forced into the NW class. The 'fully autonomous' framing is also a bit strong: the system runs on predefined strategies and thresholds (triangular mask geometry, distance thresholds), and the paper itself notes the clustering can fluctuate before locking on. These are acceptable for a proof-of-concept, but they need to be stated as limitations and backed by metrics.\n\nThe paper is honest about some of this; it mentions the threshold choices and the lack of temporal tracking. I would send it to a serious referee. The demonstrations are compelling enough to warrant the field's attention, but the revision needs to add quantitative classifier validation and, ideally, a yield/success table for the different fabrication strategies. Worth a reading group discussion on what counts as 'autonomous' in microscopy.\n\nRecommendation: accept for peer review with major revisions.\n\nBest","headline":"A credible proof-of-concept for closed-loop autonomous STEM fabrication, but the 'reliability' claim needs quantitative classifier and yield metrics.","tokens_in":13604,"tokens_out":2690,"would_cite":true,"duration_ms":33307,"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 closed-loop electron microscope can fabricate designed atomic-scale defect structures on its own: it decodes images, selects sulfur atoms, and aims the beam to grow MoS-nanowire edges in MoS2.","keywords":["autonomous fabrication","scanning transmission electron microscopy","machine learning","MoS2","MoS nanowire","electron beam manipulation","defect engineering","HAADF imaging"],"falsifier":"Run the closed loop on a fresh pristine $\\mathrm{MoS_2}$ monolayer at a pixel size and defect morphology not seen in training, compare every targeted site against a high-quality post-fabrication image, and count how many beam exposures remove the intended sulfur atom; if mislabeled molybdenum sites or missed nanowire regions divert the beam often enough to prevent the intended $\\mathrm{MoS}$-NW structure from appearing, the claim of reliable autonomous fabrication is falsified.","tokens_in":12588,"feed_emoji":"⚛️","tokens_out":10153,"duration_ms":89019,"temperature":0.7,"pith_summary":"This paper establishes that a closed-loop scanning transmission electron microscope can fabricate designed atomic-scale defect structures without human intervention: machine-learning models decode each high-angle annular dark-field (HAADF) image into atom identities, an automated planner chooses sulfur sites to remove, and programmable beam control exposes exactly those sites. Using a monolayer of $\\mathrm{MoS_2}$ as the testbed, the authors report reliable production of $\\mathrm{MoS}$-nanowire ($\\mathrm{MoS}$-NW) edge structures on pre-defined nanopore edges, free-standing nanowires between two pores, and nanowires grown directionally across the field of view. The reason this matters is practical: if the loop works generally, atomically precise defect structures with tailored electronic and mechanical properties can be manufactured by specifying a target morphology and letting the microscope execute it, rather than by manual operation and favorable irradiation statistics. The authors show image sequences as evidence and note that the machine-learning models were trained on only a few dozen hand-labeled examples.","feed_headline":"Closed-loop electron beam writes MoS nanowires atom by atom","feed_subtitle":"Machine learning decodes each image, picks sulfur sites, and guides the beam to grow metallic nanowire edges on demand.","key_machinery":"The load-bearing machinery is the closed feedback cycle of image decoding, site selection, and beam delivery. Atom positions come from an ELIT ensemble-learning iterative-training model, pretrained on multislice simulations of $\\mathrm{MoS_2}$ and augmented with experimental images. Nanowire regions are segmented by a lightweight U-net, a convolutional network architecture for image segmentation. The remaining atomic sites are classified by a random forest that takes a hand-crafted vector of the intensity differences and distances to each atom's five nearest neighbors, with class probabilities averaged over trees as an uncertainty estimate. The fabrication side is an FPGA-controlled scan routine that can draw scan paths such as an Archimedean spiral of 2--5~\\AA{} size, giving a radially symmetric dose profile with controllable dwell time. The argument rests on these components working together: drift correction between HAADF snapshots keeps the field of view aligned, so every fabrication step is planned from a freshly decoded image.","core_discovery":"On its own terms, the central discovery is that a small-data machine-learning stack, coupled to a flexible beam scanner, is sufficient for autonomous feedback-controlled fabrication of complex defect structures. The workflow cycles through four stages: acquire a HAADF image; decode it by locating every atomic column with an ELIT ensemble of convolutional networks, segmenting $\\mathrm{MoS}$-NW regions with a lightweight U-net, and classifying each remaining atom as $\\mathrm{Mo}$, $\\mathrm{S}$, or sulfur-vacancy-line with a random forest; choose which sulfur sites to expose according to a pre-defined growth strategy; and deliver a localized 60~kV electron beam dose to each site, using either the microscope's native scan or an FPGA-controlled Archimedean spiral for radially symmetric exposure. The authors report that targeting sulfur sites is controllable while targeting molybdenum sites is not, consistent with sulfur's lower displacement threshold, and that this distinction is what makes the three demonstrated outcomes possible: nanopore-edge $\\mathrm{MoS}$-NWs, free-standing $\\mathrm{MoS}$-NWs spanning two nanopores, and directional growth of a single nanowire across the full field of view.","pith_inferences":["An extension the authors leave implicit is to use the random forest's per-class probabilities as a gating signal: when classification confidence drops, the system could re-image or expand the labeled support set instead of continuing to target, which would reduce misdirected beam exposure.","Because the paper's directional-growth strategy treats each frame independently and clusters nanowire sites afresh, adding temporal tracking of individual atoms and defect clusters across frames would likely stabilize target selection in the early iterations, where the authors note the algorithm can fluctuate between small clusters.","A natural transfer test would be to run the same loop on another 2D material with a known electron-beam removal site; the small training-set sizes suggest the bottleneck is not model capacity but whether the hand-crafted neighbor-distance and intensity features separate the relevant site classes."],"forward_implications":["Sulfur-site targeting, rather than molybdenum-site irradiation, is the controllable route for electron-beam defect fabrication in monolayer $\\mathrm{MoS_2}$ at 60 kV.","The workflow reliably produces $\\mathrm{MoS}$-NW structures on pre-defined nanopore edges by repeatedly removing sulfur atoms along the pore edge.","Free-standing $\\mathrm{MoS}$-NWs can be grown between two seeded nanopores without straining the surrounding lattice, by targeting atoms along the edge between the pores.","Directional growth of a single $\\mathrm{MoS}$-NW across the full field of view is achieved by clustering detected nanowire sites, fitting the largest cluster to a line, and removing nearby sulfur atoms at both ends of the line.","The authors claim the framework is material-agnostic, so the same closed loop could be applied to other 2D materials and other defect or heterostructure motifs."],"supporting_citations":[{"why":"Supplies the ELIT ensemble-learning atom-localization model that locates every atomic column in the HAADF image.","marker":"[35]"},{"why":"Established the earlier feedback-controlled beam fabrication of MoS2 defect structures whose ejection parameters this workflow automates.","marker":"[32]"},{"why":"Demonstrated the prior automated single-defect-level STEM platform and machine-learning postprocessing that this work extends to complex nanowire structures.","marker":"[33]"},{"why":"Provides the FPGA-based scan-control framework that enables custom beam paths such as the Archimedean spiral used for site targeting.","marker":"[50]"},{"why":"Identifies the metallic $\\mathrm{Mo_6S_6}$ nanowire edge structure with self-adaptive contacts to semiconducting MoS2, the target motif being fabricated.","marker":"[17]"},{"why":"Supports the mechanism by which sulfur atoms can be displaced below the knock-on threshold through localized electronic excitations, explaining why sulfur targeting is controllable.","marker":"[51]"},{"why":"Supplies the U-net convolutional architecture used to segment nanowire regions from HAADF images.","marker":"[58]"},{"why":"Provides the DBSCAN clustering step that selects the largest nanowire cluster for directional growth.","marker":"[53]"}],"fun_headline_variants":["AI-guided electron beam builds atomic-scale wires on command","Machine learning steers electron beam to write MoS nanowires atom by atom","Autonomous STEM crafts tailored defects with AI feedback loop","Self-driving microscope fabricates custom nanowires atom by atom"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The demonstration rests on the machine-learning classifiers continuing to label sulfur, molybdenum, and nanowire sites correctly as the lattice distorts and the pixel size changes, even though they were trained on only 55 hand-labeled atom neighborhoods and roughly 30 images and the paper reports no measured accuracy for them.","fun_headline_variants_meta":{"raw":{"variants":["AI-guided electron beam builds atomic-scale wires on command","Machine learning steers electron beam to write MoS nanowires atom by atom","Autonomous STEM crafts tailored defects with AI feedback loop","Self-driving microscope fabricates custom nanowires atom by atom"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00145,"raw_usage":{"total_tokens":5888,"prompt_tokens":1044,"completion_tokens":4844,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":660,"completion_tokens_details":{"reasoning_tokens":4775}},"tokens_in":660,"tokens_out":4844,"duration_ms":86233,"temperature":1.0,"reasoning_tokens":4775,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:56:41.298679+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the closed loop on a fresh pristine $\\mathrm{MoS_2}$ monolayer at a pixel size and defect morphology not seen in training, compare every targeted site against a high-quality post-fabrication image, and count how many beam exposures remove the intended sulfur atom; if mislabeled molybdenum sites or missed nanowire regions divert the beam often enough to prevent the intended $\\mathrm{MoS}$-NW structure from appearing, the claim of reliable autonomous fabrication is falsified.","supporting_citations":[{"cited_title":"npj Computational Materials, 2021","cited_arxiv_id":null,"evidence_quote":"Supplies the ELIT ensemble-learning atom-localization model that locates every atomic column in the HAADF image."},{"cited_title":"Advanced Materials, 2023","cited_arxiv_id":null,"evidence_quote":"Established the earlier feedback-controlled beam fabrication of MoS2 defect structures whose ejection parameters this workflow automates."},{"cited_title":"ACS Nano, 2022","cited_arxiv_id":null,"evidence_quote":"Demonstrated the prior automated single-defect-level STEM platform and machine-learning postprocessing that this work extends to complex nanowire structures."},{"cited_title":"Small Methods, 2024","cited_arxiv_id":null,"evidence_quote":"Provides the FPGA-based scan-control framework that enables custom beam paths such as the Archimedean spiral used for site targeting."},{"cited_title":"Nature Nanotechnology, 2014","cited_arxiv_id":null,"evidence_quote":"Identifies the metallic $\\mathrm{Mo_6S_6}$ nanowire edge structure with self-adaptive contacts to semiconducting MoS2, the target motif being fabricated."},{"cited_title":"Nano Letters, 2020","cited_arxiv_id":null,"evidence_quote":"Supports the mechanism by which sulfur atoms can be displaced below the knock-on threshold through localized electronic excitations, explaining why sulfur targeting is controllable."},{"cited_title":"Fischer, and T","cited_arxiv_id":null,"evidence_quote":"Supplies the U-net convolutional architecture used to segment nanowire regions from HAADF images."},{"cited_title":"1996, AAAI Press: Portland, Oregon","cited_arxiv_id":null,"evidence_quote":"Provides the DBSCAN clustering step that selects the largest nanowire cluster for directional growth."}],"review_version":1}