{"id":"b7fe10ec-16ef-483e-b35d-30dfc4dc9a0f","arxiv_id":"2505.03826","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An ANN and a BNN predict SiO2 etch depth from process parameters and from RGB color values, with held-out MSEs of 7.33 and 10.38 nm^2 respectively.","lead":"This paper trains neural networks to predict etch depth in plasma etching from process parameters or from the surface color of the etched wafer, using a smartphone photo as the input. The color-based model works without knowing the process settings, which could make real-time monitoring in semiconductor manufacturing cheaper and faster.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'in-situ' claim is unsupported: RGB→thickness mapping was learned from ex-situ white-background smartphone photos yet is asserted to work during the RIE process (Fig. 1 vs §2/§3.1); no through-viewport or in-chamber RGB was collected.","rationale":"The reader identified the same weakest assumption: the ex-situ training conditions do not match the claimed in-situ deployment scenario. My read of the full text confirms this is the most load-bearing concern. The paper provides a plausible proof-of-concept that a smartphone photo taken on a white background after etching contains enough RGB information to predict remaining SiO2 thickness on a small coupon dataset, and the numerical results are internally consistent with the reported table values. It also gives some supporting evidence for the ANN over the linear baseline and reports BNN coverage fractions close to nominal values. However, none of that support addresses the transfer from ex-situ images to in-chamber acquisition. Figure 1 explicitly labels the RGB data as collected 'during the RIE process,' while Section 2's experimental description and Section 3.1's title make clear the images were acquired after etching on a white background. This is an internal inconsistency, not merely a question of consensus. Because the abstract's central claim is specifically about in-situ, real-time monitoring, the missing through-viewport validation is decisive for the claim as stated. A conditional verdict is appropriate: the paper should either provide in-situ RGB validation or substantially revise the claim to ex-situ proof-of-concept. Since the reader's CONDITIONAL verdict already represents that position, I do not recommend changing the verdict.","tokens_in":10009,"tokens_out":3375,"duration_ms":37660,"concrete_test":"Install a transparent viewport on the etch chamber and, for a subset of the same 84 process conditions, capture RGB through the viewport while the coupon is in the chamber under process lighting (plasma off or on, as the claim requires), without unloading the wafer. Then measure the same coupons ex-situ with the spectroscopic ellipsometer. Apply the already-trained ex-situ ANN to the through-viewport RGB values and compare predicted versus measured thickness. If through-viewport RGB for a given measured thickness deviates enough to shift the prediction by more than the ex-situ RMSE (~3.2 nm), the in-situ transfer claim fails; if predictions stay within the BNN uncertainty intervals, the claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that DIC+ML is viable for real-time, in-situ, non-contact monitoring of etch depth. The data used to learn the RGB-to-thickness mapping were acquired ex-situ: Section 2 states that 'coupon wafers were placed on a white background during image acquisition' with an iPhone 12 after etching, and Section 3.1 is titled 'Training set generation with ex-situ analysis.' Yet Figure 1 and the abstract assert that RGB data are collected 'during the RIE process.' In an actual chamber, image formation differs materially: viewport window transmission, chamber lighting and plasma emission, viewing angle through the viewport, and possible window deposition all change RGB values for the same physical thickness. The model has no illumination normalization, no spectral calibration, and no multi-angle information to separate these nuisance factors from thickness. The held-out MSE of 10.38 nm^2 on seven validation conditions therefore demonstrates interpolation among ex-situ smartphone photos of coupons on a white background, not that the same RGB-to-thickness mapping holds in-situ. This is the load-bearing gap in the abstract's headline claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript reports machine-learning models for predicting the remaining SiO2 thickness after plasma etching. Eighty-four coupon wafers were etched in an ICP-RIE tool over a grid of chamber pressure, CF4 flow, and top power; thickness was measured by spectroscopic ellipsometry and surface images were captured with an iPhone 12 on a white background. The authors train an ANN to predict etch depth from the three process parameters, obtaining a held-out MSE of 7.33 nm^2 versus 33.94 nm^2 for a linear model, and an ANN trained on RGB values from the digital images, obtaining an MSE of 10.38 nm^2 versus 113.0 nm^2. They also apply a Monte Carlo Dropout BNN to the repeated measurements and report coverage statistics for predictive intervals. The paper concludes that this integration offers a viable, cost-effective alternative for real-time, in-situ, and non-invasive monitoring of plasma etching.","tokens_in":10213,"tokens_out":5015,"duration_ms":52598,"significance":"If the claims are supported, the RGB-based ANN is a genuinely interesting low-cost alternative to ellipsometric thickness measurement, and the BNN uncertainty quantification would strengthen its practical value for process monitoring. The experimental design is a clear strength: 84 process conditions with ellipsometry mapping, OES-based plasma stability checks, held-out validation, and direct comparison against a linear baseline. The held-out MSEs genuinely support the narrower claim that etch depth can be predicted from ex-situ smartphone photographs of coupons on a white background. However, the paper's headline claim of in-situ monitoring is not supported by the experimental protocol, and the validation statistics rest on a single small split with an unspecified BNN data-splitting procedure. The central feasibility result is defensible if reframed as ex-situ DIC-based prediction, but the current manuscript overclaims.","major_comments":[{"comment":"The central 'in-situ' claim is not supported by the experimental protocol. The abstract and Figure 1 state that RGB data are collected 'during the RIE process,' but Section 3.1 is titled 'Training set generation with ex-situ analysis' and Section 2 reports that 'coupon wafers were placed on a white background during image acquisition' with an iPhone 12 after etching. Plasma emission, viewport window transmission, chamber lighting, and viewing geometry all change RGB values for a fixed thickness, and the paper reports no illumination normalization, spectral calibration, or through-viewport images. Consequently, the held-out MSE of 10.38 nm^2 in Table 4 demonstrates interpolation among ex-situ white-background photos, not that the mapping transfers in-situ. Please either add in-chamber validation data or revise the abstract, title, and Figure 1 to claim ex-situ feasibility and explicitly discuss the transfer assumptions.","section":"Abstract, Fig. 1, §2, §3.1"},{"comment":"The validation protocol rests on a single random split of only seven samples for the process-parameter ANN and for the RGB ANN. With n=7, the reported MSEs (7.33 versus 33.94 and 10.38 versus 113.0) are sensitive to the particular split, training seed, and network initialization. No confidence intervals, repeated-seed results, or cross-validation results are given. Please report results over multiple splits or with a leave-one-condition-out scheme, including the mean and standard deviation of the MSE.","section":"§3.4.1, Table 3, Table 4"},{"comment":"For the BNN experiments, 756 samples are formed by treating nine repeated ellipsometer measurements as independent, and 152 are held out, but the text does not specify whether the split is at the coupon level or the individual-measurement level. If measurements from the same coupon appear in both training and validation, the quoted coverage (68.25%/23.81%/7.94% and 63.16%/34.87%/1.97%) is inflated by leakage. The DIC-BNN section similarly does not state how repeated RGB measurements were obtained or how the 152 validation samples were selected. Please specify the split, use coupon-level grouping (for example, GroupKFold), and report the average predictive standard deviation alongside the coverage percentages.","section":"§3.4.1 and §3.4.2, BNN coverage analysis"},{"comment":"The statement that the RGB-based model can generalize across plasma processes because it does not use explicit process parameters is too strong. The training data cover one tool, one gas chemistry, and one 300-nm SiO2 film, and RGB values may also encode process-induced surface conditions. Please temper this generalization claim or support it with data from a different tool, chemistry, or film stack.","section":"§3.4.2, last paragraph"}],"minor_comments":[{"comment":"The manuscript jumps from Section 3.2 to Section 3.4; Section 3.3 is missing and should be renumbered.","section":"Section numbering"},{"comment":"In Section 3.4.2, the text refers to 'surface features shown in Figure 5' and 'thickness illustrated in Figure 6,' but the relevant panels appear to be Figures 2 and 3; the cross-references should be corrected.","section":"§3.4.2, figure references"},{"comment":"Table 2 has inconsistent formatting: rows for the 30 mTorr conditions contain only eight entries and the column alignment is unclear; please reformat so all nine thickness columns are visible.","section":"Table 2"},{"comment":"Equation (4) defines the linear model without a bias term even though the surrounding text introduces a bias b_l; this should be reconciled.","section":"Eq. (4)"},{"comment":"The sentence describing 'in-situ RGB data obtained from DIC' conflicts with the ex-situ image acquisition described earlier in the same section; use 'ex-situ' consistently unless in-chamber images are actually acquired.","section":"§2, last paragraph"},{"comment":"The dropout probability and the number of hidden units for the BNN/MC Dropout model are not specified; these settings affect the reported coverage and should be stated.","section":"§3.4.1, BNN implementation"}],"recommendation":"major_revision","confidential_remarks":"The main gap is between the promotion of the method as in-situ and the ex-situ data actually collected. I would advise the editor that acceptance should be contingent on either adding through-viewport validation data or substantially revising the title, abstract, and figure captions to describe an ex-situ feasibility study with a clear discussion of the conditions under which in-situ transfer would need to be re-established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nQuick take: this is a decent proof-of-concept that a smartphone photo of an etched coupon, taken on a white background in fixed lighting, contains enough color information to predict remaining SiO2 thickness with an RMSE around 3 nm. The ANN/RGB model beats a linear baseline by a wide margin (10.38 vs 113.0 MSE), and the BNN coverage numbers are in the right ballpark. That part is real and worth acknowledging.\n\nWhat's actually new: applying the established DIC+ML pipeline from food and chemistry to plasma etch depth monitoring. The transfer is straightforward, but the empirical result is new, and the idea of replacing ellipsometry with a camera is genuinely attractive for cost and contamination reasons.\n\nThe soft spots are equally clear. The biggest problem is the word \"in-situ.\" The data were collected after etching, ex-situ, on a white background (Section 2 and Section 3.1 say so explicitly), yet the abstract and Figure 1 claim RGB collection \"during the RIE process.\" Through a viewport, with plasma emission, window transmission, and different viewing angles, RGB values will shift in ways this model has never seen. As a proof-of-concept for ex-situ monitoring it's fine; as an in-situ method it's unsupported.\n\nThe statistics are also thin. The held-out set is seven conditions from a single split. The BNN coverage analysis treats the nine repeated measurements as independent, which inflates the effective sample size. No code, data, or hyperparameters are provided, so the exact numbers are hard to verify. There are also some section and figure numbering inconsistencies (3.4 appears twice, Figure 7 is described before Figure 8, etc.) that suggest a hurried submission.\n\nNone of these are fatal to the core idea. The color-to-thickness relationship is physically sensible (thin-film interference), and the ANN is a reasonable tool to invert it. But the paper needs to either provide a true in-situ demonstration or drop the in-situ language, and it needs multiple splits or cross-validation before I'd trust the error bars.\n\nBottom line: it deserves a serious referee, because the application is timely and the negative result would be interesting even if the in-situ transfer fails. I'd send it to review, but the authors should be told that the in-situ claim won't survive without real through-viewport data.\n\nBest","headline":"A plausible proof-of-concept that RGB photos can predict SiO2 etch depth, undercut by an unsupported 'in-situ' claim and a very thin statistical basis.","tokens_in":10810,"tokens_out":1690,"would_cite":false,"duration_ms":17220,"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":"The paper claims that plasma etch depth can be read from the RGB values of a digital photograph with a validation RMSE around 3.2 nm, and that a Bayesian version of the network supplies trustworthy uncertainty intervals for each prediction.","keywords":["plasma etching","digital image colorimetry","etch depth prediction","artificial neural network","Bayesian neural network","uncertainty quantification","silicon dioxide","optical emission spectroscopy"],"falsifier":"Photograph a wafer of known etch depth both on a white background and through a plasma-chamber viewport under representative process illumination, and compare the RGB triplets for the same physical state. If the two RGB distributions differ by more than the pixel-to-pixel variation in the training images, the current model cannot support the claimed in-situ use; alternatively, capture the 84-condition image set through the viewport and check whether the validation MSE stays near 10.38 nm².","tokens_in":9808,"feed_emoji":"📸","tokens_out":8436,"duration_ms":85167,"temperature":0.7,"pith_summary":"This paper tries to establish that plasma etch depth can be monitored without touching the wafer or stopping the chamber, using only the color of the wafer's surface. The authors train an ANN to map RGB values extracted from photographs of etched SiO2 coupons to the etch depth recorded by ellipsometry, and report a held-out MSE of 10.38 nm² (about 3.2 nm RMSE) on seven validation conditions. They also train a Bayesian neural network with Monte Carlo dropout so that each prediction comes with an uncertainty interval; for the image-based model, true values fall inside the ±1σ interval 63.16% of the time and outside ±2σ only 1.97% of the time. If the claim holds, thickness monitoring in plasma etching could be done with a camera and a trained network instead of a separate ex-situ measurement step.","feed_headline":"A photo's RGB values predict etch depth to about 3 nm","feed_subtitle":"A neural net trained on 84 etched wafers maps surface color to SiO2 etch depth and flags its own uncertainty.","key_machinery":"Digital image colorimetry (DIC) is the central object: a photograph of the etched coupon on a white background is reduced to average red, green, and blue intensities between 0 and 255 over a representative rectangle, and those three numbers are the ANN's input features. The RGB-to-etch-depth mapping is carried by a feedforward ANN with one 32-neuron ReLU hidden layer and 20% dropout, trained with the Adam optimizer and weight decay. The BNN variant is the same network with Monte Carlo dropout: 50 stochastic forward passes per input produce a predictive distribution whose mean is the point prediction and whose standard deviation defines the uncertainty interval. Coverage analysis, which counts how often true etch depths fall in the ±1σ and ±2σ bands, is the mechanism the paper uses to argue that the uncertainty estimates are reliable.","core_discovery":"The central claim is that the RGB triplet of a digital image of a dielectric surface carries enough information about ellipsometric etch depth that an ANN can predict that depth without knowing pressure, gas flow, or RF power. With 84 etched coupon conditions and a seven-condition validation split, the image-based ANN achieves a validation MSE of 10.38 nm², versus 113.0 nm² for a linear RGB model, while the process-parameter ANN achieves 7.33 nm² versus 33.94 nm² for a linear parameter model. The BNN extension, implemented by treating each of nine measurement points as a separate sample and applying Monte Carlo dropout at inference, gives coverage of 63.16% within ±1σ and 34.87% within ±2σ for image inputs, and 68.25% and 23.81% for process-parameter inputs, with 7.94% outside ±2σ in the latter case. The paper reads these numbers as evidence that non-contact RGB monitoring is a viable, cost-effective alternative to conventional ex-situ analysis.","pith_inferences":["The claimed independence of the RGB-to-etch-depth mapping from pressure, CF4 flow, and RF power rests on one 84-condition grid; a stronger test would train on some pressure/flow/power combinations and test on others, since the seven validation points all come from the same grid.","Inside a real chamber, plasma optical emission contaminates the image signal; a natural extension the paper does not test is to subtract or normalize plasma emission lines before extracting RGB values.","If the color signal is dominated by thin-film interference, the same mapping should work for other transparent dielectrics, and could be checked on a staircase etch-depth calibration coupon with known thickness steps."],"forward_implications":["A camera-based monitor could flag etch depth in real time without transferring the wafer out of vacuum, reducing measurement delay and contamination risk.","Because RGB inputs contain no explicit process parameters, a trained network may transfer across plasma recipes that produce the same surface colors.","The BNN's uncertainty intervals give an operator a criterion for trusting or rejecting a prediction, such as flagging drift when coverage degrades.","The same ANN/BNN pipeline could be retrained for other insulating films, such as silicon nitride, using the same ellipsometer-plus-photograph data collection."],"supporting_citations":[{"why":"Establishes digital image colorimetry as the image-analysis basis, where digitized images from consumer cameras supply color data for quantitative measurement.","marker":"[16]"},{"why":"Provides precedent for coupling DIC with machine learning to replace direct color matching with data-driven prediction.","marker":"[28]"},{"why":"Recent application of machine learning with digital image data, motivating the color-based etch-depth prediction.","marker":"[29]"},{"why":"Further demonstration that machine learning with DIC improves precision and supports real-time prediction.","marker":"[30]"},{"why":"Identifies the difficulty of in-situ nanometer-scale measurement and the limited number of available samples, motivating the Bayesian uncertainty approach.","marker":"[31]"},{"why":"Supplies the Monte Carlo dropout method used to turn the ANN into a Bayesian network with predictive uncertainty.","marker":"[32]"},{"why":"Supports using a Bayesian neural network to estimate prediction reliability in physical measurement tasks.","marker":"[33]"},{"why":"Gives the physical explanation for why etch depth tracks power and pressure through ion mean free path and kinetic energy, grounding the color-to-thickness relationship.","marker":"[36]"}],"fun_headline_variants":["RGB photos plus ML predict etch depth to ~3 nm","Neural net maps wafer color to etch depth without process params","Bayesian NN adds uncertainty to in-situ etch depth from images","Non-contact etch depth from digital color and ANN","Photo-based neural net forecasts SiO2 etch depth and its error"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the color-to-etch-depth relationship learned from ex-situ photos taken on a white background under fixed lighting still holds when the method is called in-situ, meaning images taken through a chamber viewport during plasma processing; if viewport transmission, plasma emission, or viewing geometry changes the RGB values for the same physical etch state, the reported accuracy and uncertainty coverage will not transfer.","fun_headline_variants_meta":{"raw":{"variants":["RGB photos plus ML predict etch depth to ~3 nm","Neural net maps wafer color to etch depth without process params","Bayesian NN adds uncertainty to in-situ etch depth from images","Non-contact etch depth from digital color and ANN","Photo-based neural net forecasts SiO2 etch depth and its error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000383,"raw_usage":{"total_tokens":2057,"prompt_tokens":1004,"completion_tokens":1053,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":971}},"tokens_in":620,"tokens_out":1053,"duration_ms":11001,"temperature":1.0,"reasoning_tokens":971,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:09:06.499148+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Photograph a wafer of known etch depth both on a white background and through a plasma-chamber viewport under representative process illumination, and compare the RGB triplets for the same physical state. If the two RGB distributions differ by more than the pixel-to-pixel variation in the training images, the current model cannot support the claimed in-situ use; alternatively, capture the 84-condition image set through the viewport and check whether the validation MSE stays near 10.38 nm².","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes digital image colorimetry as the image-analysis basis, where digitized images from consumer cameras supply color data for quantitative measurement."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides precedent for coupling DIC with machine learning to replace direct color matching with data-driven prediction."},{"cited_title":"K.; Kim, J.; Kim, S.; Kim, C.; Lee, H.-C.; Kang, W.; Choi, M","cited_arxiv_id":null,"evidence_quote":"Recent application of machine learning with digital image data, motivating the color-based etch-depth prediction."},{"cited_title":"-U.; Kim, T., Sensors 2023, 23 (12), 5563","cited_arxiv_id":null,"evidence_quote":"Further demonstration that machine learning with DIC improves precision and supports real-time prediction."},{"cited_title":"R.; LeSar, R., Annu","cited_arxiv_id":null,"evidence_quote":"Identifies the difficulty of in-situ nanometer-scale measurement and the limited number of available samples, motivating the Bayesian uncertainty approach."},{"cited_title":"In Dropout as a bayesian approximation: Representing model uncertainty in deep learning, international conference on machine learning, PMLR: 2016; pp 1050-1059","cited_arxiv_id":null,"evidence_quote":"Supplies the Monte Carlo dropout method used to turn the ANN into a Bayesian network with predictive uncertainty."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supports using a Bayesian neural network to estimate prediction reliability in physical measurement tasks."},{"cited_title":"A.; Lichtenberg, A","cited_arxiv_id":null,"evidence_quote":"Gives the physical explanation for why etch depth tracks power and pressure through ion mean free path and kinetic energy, grounding the color-to-thickness relationship."}],"review_version":1}