{"id":"792b2fb5-fbac-49f2-92ab-7f68f9415d30","arxiv_id":"2501.14813","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"This thesis demonstrates neural-network prediction, simulation-based inference, and GFlowNet generation for CNTFET devices, claiming all three approaches can assist CNT device modeling and design.","lead":"Using three machine learning techniques, this dissertation tries to model carbon nanotube transistors, extract electrical parameters from messy device measurements, and generate fabrication recipes for a target current-voltage curve. A smart generalist might read it as a test of whether ML tools can replace or assist traditional device modeling and parameter extraction in materials science.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SBI parameter recovery is unvalidated: posterior may be broad or degenerate, so the claimed resistance values are not yet established.","rationale":"The reader's weakest assumption focused on structural correctness of the compact network model. My concern is adjacent but more specific: even if the model structure is correct, the inference procedure has not been shown to recover known parameters, so the reported numerical values (Rm ≈ 10 kΩ, Rex ≈ 120 kΩ) could be artifacts of degeneracy or prior choice. This is a correctness risk in the central claim of SBI as a parameter-extraction tool. I agree with the reader's conditional verdict: the thesis is promising and self-aware about limitations, but the absence of code/data and the missing calibration study prevent independent verification. My proposed SBC test is a concrete step that would settle whether the worry lands. I do not see a reason to move the verdict to ACCEPT or REJECT; keeping it CONDITIONAL is appropriate.","tokens_in":40656,"tokens_out":2025,"duration_ms":27264,"concrete_test":"Run a simulation-based calibration (SBC) study: fix parameters to a point in the reported posterior (e.g., Rm = 10 kΩ, Rin = 120 kΩ, k = 0.8), generate 100 synthetic datasets each containing 100 simulated devices, then run the exact SNPE pipeline from Section 5.5 (same simulator, same priors, 600 training simulations) on each dataset. Compute the coverage of 90% posterior intervals for Rm, Rin, and k. If the true value falls outside the interval in more than 10% of trials, or if the posterior is multimodal with near-equal probability across the prior range, then the extracted parameters are not identifiable from the observed distributions and the claimed physical correspondence is unsupported. Additionally, use parameters inferred from wafer 1 to predict wafer 2's I_off distribution; a factor-of-ten error would indicate overfitting to the training distribution.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Chapter 5 is that simulation-based inference (SNPE) recovers physically meaningful CNT contact, junction, and section resistances from measured I_on/I_off distributions, because the inferred values are 'close to previous experimental studies.' The load-bearing issue is that no parameter-recovery or identifiability test is reported. The simulator outputs only two gamma-distribution parameters (alpha, beta) for I_on and I_off, while the inference targets three parameters (k, Rm, Rin) after fixing the V_t-from-SS relation with a constant b that is never listed in the prior (Section 5.5). If b is fixed rather than inferred, or if the prior is too narrow, the posterior may concentrate on values that fit the observed gamma statistics without being the physical resistances. The text itself hints at non-identifiability: Section 5.6 reports k 'around 1 for I_on' but 'higher for I_off', and Rm 'seems to be higher for I_off', which is physically odd if these are universal device constants. No held-out wafer is used to check whether the inferred parameters predict a second wafer's I_on/I_off distributions. The self-description on page v, 'The dissertation is uploaded here just to protect my ideas,' explains the absence of code and digitized data but does not substitute for a simulation-based calibration study.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This dissertation applies machine learning to carbon nanotube field-effect transistors (CNTFETs) in three parts: neural network surrogate modeling of I-V characteristics, simulation-based inference (SBI) to extract resistance parameters in a compact network model, and GFlowNet-based generative models for device design. The central claim is in Chapter 5: SBI can recover CNT-metal contact resistance, CNT-CNT intersection resistance, and CNT section resistivity from measured on/off current distributions of randomly deposited CNTFETs, yielding values close to literature.","tokens_in":40941,"tokens_out":5763,"duration_ms":57346,"significance":"If validated, the SBI approach would be a useful practical tool for parameter extraction when a model's output is a distribution rather than a single value, a genuinely valuable methodological contribution. The thesis also presents a transparent network compact model and includes a negative control (failed SBI when the model is wrong), which is a strength. However, the current evidence is insufficient to establish the central extraction claim because of circularity in the Vt construction, the absence of any parameter-recovery test, and the lack of quantitative evaluation in the neural-network chapter.","major_comments":[{"comment":"The threshold voltage distribution is generated from the measured SS distribution through V_t = k*(1/n_ss)+b, with k and b inferred from the same wafer data. The agreement between the generated and measured V_t distributions shown in Fig 5.5 is therefore guaranteed by construction, and V_t is an input to the simulator rather than an independent prediction. Consequently, the resistance posterior is conditional on a fitted and uncertain V_t-SS relation; the paper should report the k,b fit uncertainty, propagate it through the inference, and validate the V_t relation on a held-out wafer.","section":"Section 5.5"},{"comment":"No parameter-recovery or identifiability test is reported. The simulator output is summarized by gamma distribution parameters (alpha, beta) for I_on and I_off, while the inference targets at least three parameters (k, Rm, Rin). The text does not demonstrate that a known parameter set can be recovered from simulated data. Given that the priors for Rm and Rin are set to [1,20] kOhm and [1,500] kOhm, which bracket the literature values cited in Section 5.2, the reported posteriors 'close to previous experimental studies' may largely reflect the prior. A simulation-based calibration study with ground-truth parameters and a prior-sensitivity analysis is needed.","section":"Section 5.6 and inference setup in Section 5.5"},{"comment":"The reported posterior behavior is physically problematic: k is 'around 1 for I_on' but 'higher for I_off', and Rm 'seems to be higher for I_off'. If these are intrinsic device constants, the fits to I_on and I_off should produce compatible posteriors. The discrepancy suggests model misspecification or non-identifiability, and the text does not address this tension. The authors should test whether the I_on and I_off posteriors overlap and discuss the implications.","section":"Section 5.6"},{"comment":"The claim that the neural network 'can provide reasonable predictions' is supported only by visual inspection of a few sample I-V curves (Figs 4.14-4.17). No quantitative error metrics, comparison to a baseline, or device-level held-out split are provided, so the predictive accuracy and generalization of the model are not established.","section":"Section 4.4"},{"comment":"The text is unclear whether k is fixed ('k is 1') or inferred with prior [0.3,1], and the parameter b in V_t = k*(1/n_ss)+b is never assigned a prior or a fixed value. This ambiguity affects the reproducibility of the inference and should be resolved.","section":"Section 5.5, prior specification"}],"minor_comments":[{"comment":"The section number 5.6 is used twice, first for 'Results' and then for 'Conclusion and Future Research'; this should be renumbered.","section":"Section 5.6"},{"comment":"The concluding chapter refers to 'chapter 3' for the neural-network work, 'chapter 4' for SBI, and 'chapter 5' for GFlowNet; the actual chapters are 4, 5, and 6 respectively.","section":"Chapter 7"},{"comment":"The text states 'we developed a method to tackle this problem in chapter 4 using simulation-based inference', but the SBI work appears in Chapter 5.","section":"Section 2.3"},{"comment":"The source of experimental data is given as 'experimental data from [ ]' with an empty citation; a specific reference is needed.","section":"Section 4.3"},{"comment":"The CNT diameter formula is garbled: it should involve a square root sqrt(m^2 + m*n + n^2) and a factor (sqrt(3)/pi)*a_CC; the printed equation is missing these elements.","section":"Section 2.1"},{"comment":"The reported inferred values are internally inconsistent: the text first says Rm is around 10 kOhm and intersection resistance around 120 kOhm, then later states 'CNT-metal contact resistance is around 150 kOhm'; the latter appears to be a typo but should be corrected.","section":"Section 5.6"}],"recommendation":"major_revision","confidential_remarks":"The dissertation is self-archived with the note 'The dissertation is uploaded here just to protect my ideas', which suggests that code and data are not made available. For a methods-oriented manuscript, the absence of released code and digitized data is a reproducibility concern, though not a correctness flaw. The manuscript is written as a raw dissertation rather than a journal article and would require substantial restructuring and additional validation experiments (especially a parameter-recovery study) to meet the standards of the journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thesis is a case study of three ML tools for CNTFETs: a neural-network surrogate, SBI for parameter extraction in a non-aligned CNT network, and GFlowNet for inverse design. The SBI chapter is the genuine contribution; the other two chapters are exploratory fills.\n\nThe most original piece is the combination of a resistor-network virtual-source compact model with simulation-based inference to estimate CNT-metal contact resistance, CNT-CNT junction resistance, and CNT section resistivity from the measured I_on/I_off distributions. That is a legitimate new use case, and the author is unusually candid about what did not work (e.g., the GFlowNet processing-information model only partially meeting targets). The chapter also includes a useful negative control: a deliberately wrong model that SBI fails to fit.\n\nThe soft spots are real, and they center on Chapter 5. The inferred resistance values are not validated. There is no simulation-based calibration or identifiability test, so the posterior might be broad or degenerate. The priors are set deliberately close to literature values, so the subsequent agreement with literature is partly circular. Worse, the threshold-voltage distribution is generated from the subthreshold-swing distribution using k and b inferred from the same wafer data, which guarantees the Vt match by construction. And the extracted parameters differ between I_on and I_off fits (k around 1 vs higher; Rm higher), which is physically odd for supposedly universal device constants. No held-out wafer is used to test predictive power. Chapter 4 lacks any error metrics, device-level held-out splits, or baseline comparisons; Chapter 6 has only one appendix plot comparing against random generation. The page-v note, 'uploaded here just to protect my ideas,' explains the absence of code and data but does not substitute for them.\n\nWho is this for? Researchers applying ML to device modeling, especially those interested in SBI for extract-parameters-from-distributions problems. It deserves a serious referee because the SBI application is plausible and novel, but the current form is not publishable without major revision: add parameter-recovery and identifiability checks, test on withheld wafers, report error bars, compare GFlowNet against random baselines, and release code and data.\n\nMy recommendation: send it to peer review, but flag that the SBI chapter's central quantitative claims are conditional on the missing validation.","headline":"A promising but unvalidated case study: the SBI-based resistance extraction is the real contribution, and it needs a calibration check, held-out data, and code before the numbers can be trusted.","tokens_in":41468,"tokens_out":1512,"would_cite":false,"duration_ms":20265,"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":"Simulation-based inference recovers the resistances inside random nanotube transistor networks from measured current spreads, matching earlier experiments.","keywords":["carbon nanotube field-effect transistors","simulation-based inference","compact model","random carbon nanotube networks","parameter extraction","neural network modeling","GFlowNet","junction resistance"],"falsifier":"Fabricate wafers at a clearly different channel length (for example 140 nm or 600 nm) or CNT density (for example 20 or 80 CNTs per micrometer) under the same nominal recipe, and compare the measured $I_{on}$ and $I_{off}$ distributions against the model's predictions evaluated at the inferred resistances; because the model predicts saturation in both directions, a systematic mismatch would indict the structural assumptions rather than the parameters. Separately, measure CNT-CNT junction resistance on the same sorted-CNT material with conducting AFM and check whether it lands near the inferred ~120 kΩ or closer to the ~700 kΩ DFT value.","tokens_in":40383,"feed_emoji":"⚡","tokens_out":14671,"duration_ms":130413,"temperature":0.7,"pith_summary":"This dissertation tries to show that machine learning can take over three jobs in carbon-nanotube transistor research that traditional methods handle poorly: predicting device curves from scattered experimental data, extracting physical parameters when the model output is a distribution rather than a single number, and generating fabrication recipes for a target performance. The load-bearing result is the middle one. Transistors built on randomly deposited carbon nanotube networks produce a spread of on- and off-currents, and the resistances at the CNT-metal contacts and at the CNT-CNT junctions inside such a random network cannot be pulled out by standard extraction techniques, since hundreds of junctions are wired together and no closed-form expression exists. The author builds a compact network model that turns every CNT segment into a virtual-source resistor, then uses simulation-based inference to find the resistances whose simulated current spread best matches three measured wafers. The recovered values, about 10 kΩ for the metal contact, about 120 kΩ for the CNT-CNT junction, and roughly 77 kΩ per micrometer of CNT section, sit close to earlier atomic-force-microscopy and engineering estimates, which is what one would expect if the extraction is finding physical parameters and not just fitting noise.","feed_headline":"Simulation-based inference extracts nanotube-transistor resistances","feed_subtitle":"It recovers contact and junction resistances from the measured spread of on/off currents, matching earlier experiments.","key_machinery":"The carrying mechanism is the pairing of a netlist-based compact model with simulation-based inference. For each sampled random network, the model builds a resistor netlist: each CNT section contributes a length-dependent virtual-source resistance, each crossing contributes a fixed intersection resistance $R_{int}$, and each metal contact contributes a fixed $R_m$; threshold voltage is drawn through $V_t = k/n_{ss} + b$ from the measured subthreshold-swing distribution. The simulator repeats this over 100 random devices and summarizes the resulting currents as a gamma distribution $f(x;\\alpha,\\beta)=\\beta^\\alpha x^{\\alpha-1} e^{-\\beta x}/\\Gamma(\\alpha)$, which is the model output. Sequential neural posterior estimation with truncated proposals learns the posterior over the triple $(k, R_m, R_{int})$ by matching simulated to observed gamma parameters. A structurally wrong model, for instance inverting the length dependence of resistance, produces no converging parameter set and no distributional fit, showing that the netlist structure carries the physical content and the inference algorithm does not manufacture agreement by itself.","core_discovery":"The central claim is that simulation-based inference can recover the resistance parameters of a random network transistor from the measured distribution of its performance, a setting in which classical extraction is impossible because the model output is a probability distribution over many devices rather than a single current value. The compact model treats each CNT segment between two junctions or contacts as an independent virtual-source resistor, adds a constant scalar resistance $R_{int}$ at every CNT-CNT intersection and $R_m$ at every metal contact, and folds threshold-voltage variation in through the measured subthreshold-swing distribution using the fitted relation $V_t = k/n_{ss} + b$. Sampling random networks of 285 nm channel length at 45 CNTs per micrometer, the simulator generates 100 devices per run and summarizes the on- and off-currents as a gamma distribution; sequential neural posterior estimation then finds the parameter triple whose simulated gamma parameters match the observed ones. With the inferred parameters the model reproduces the $I_{on}$ and $I_{off}$ distributions of all three wafers, with $k$ near 1 for the on-state and somewhat higher for the off-state, and a deliberately mistuned model fails to converge to any satisfactory parameter set, which the paper reads as evidence that the inference tracks the model structure rather than fitting anything arbitrarily.","pith_inferences":["The same netlist-plus-SBI pipeline should transfer to other random-network devices, such as sensors, transparent conductors, or printed electronics, where junction resistance dominates and the measured quantity is a spread of device outputs rather than a single value.","Summarizing each simulation run by a fitted gamma distribution discards shape information; conditioning the posterior on the full empirical distribution of simulated currents could tighten the posterior and reveal whether the three-parameter model is over- or under-determined.","Because the inferred junction resistance (~120 kΩ) sits close to the ~200 kΩ AFM-based and ~150 kΩ engineering estimates but well below the ~700 kΩ DFT value, an independent junction-resistance measurement on the same sorted-CNT material would separate model error from material-to-material variation.","The two fitted constants in $V_t = k/n_{ss} + b$ absorb any mismatch between the threshold-voltage shortcut and reality; checking the relation against directly measured threshold voltages on a fresh wafer would localize error in the network model versus error in that shortcut."],"forward_implications":["For compact models whose output is a distribution of device behavior, simulation-based inference supplies a parameter-extraction route where linearization and exponential-transformation methods are simply unavailable.","The same extracted parameters reproduce the measured on- and off-current spreads across three separately fabricated wafers, so the calibration transfers between wafers made under the same nominal recipe.","Using the inferred parameters, the model predicts that on/off current saturates as CNT density increases and falls asymptotically as gate length grows, matching observed scaling behavior.","The failed run with a deliberately mistuned model shows that simulation-based inference can signal a wrong model structure rather than silently returning a best-fit parameter set, giving modelers a diagnostic check."],"supporting_citations":[{"why":"Supplies the SNPE variant with truncated proposals that the chapter uses for sequential posterior estimation.","marker":"[56]"},{"why":"Provides the DFT-calculated CNT-CNT junction resistance (~700 kΩ) that the inferred intersection resistance is compared against.","marker":"[78]"},{"why":"The simulation-based-inference toolkit that provides the posterior estimator and sampler used in the inference pipeline.","marker":"[83]"},{"why":"Establishes the SNPE approach of training a conditional density estimator on simulated parameter-data pairs, which the inference relies on.","marker":"[84]"},{"why":"Conducting-AFM measurement of SWNT-SWNT junction resistance (~200 kΩ), an experimental baseline for the inferred value.","marker":"[88]"},{"why":"The aligned-CNTFET virtual-source compact model that the non-aligned network model adapts into per-section resistors.","marker":"[91][92]"},{"why":"Measured on/off current, threshold-voltage, and subthreshold-swing distributions from three wafers that serve as the inference target.","marker":"[94]"}],"fun_headline_variants":["Simulation-based inference recovers nanotube transistor resistances","AI extracts nanotube resistances from current spread","Nanotube resistances recovered via machine learning inference","Inference matches nanotube currents to reveal resistances"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the compact network model is structurally correct: each CNT segment behaves as an independent virtual-source resistor, junction and metal-contact resistances are constant scalars, threshold voltage follows the fixed two-constant function of subthreshold swing, and the CNT length and diameter distributions borrowed from a datasheet and similar studies match the fabricated wafers; if any of these fails, simulation-based inference will return parameters that fit the measured gamma distributions without being the physical resistances.","fun_headline_variants_meta":{"raw":{"variants":["Simulation-based inference recovers nanotube transistor resistances","AI extracts nanotube resistances from current spread","Nanotube resistances recovered via machine learning inference","Inference matches nanotube currents to reveal resistances"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000443,"raw_usage":{"total_tokens":2191,"prompt_tokens":839,"completion_tokens":1352,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":455,"completion_tokens_details":{"reasoning_tokens":1289}},"tokens_in":455,"tokens_out":1352,"duration_ms":12663,"temperature":1.0,"reasoning_tokens":1289,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:51:11.488023+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fabricate wafers at a clearly different channel length (for example 140 nm or 600 nm) or CNT density (for example 20 or 80 CNTs per micrometer) under the same nominal recipe, and compare the measured $I_{on}$ and $I_{off}$ distributions against the model's predictions evaluated at the inferred resistances; because the model predicts saturation in both directions, a systematic mismatch would indict the structural assumptions rather than the parameters. Separately, measure CNT-CNT junction resistance on the same sorted-CNT material with conducting AFM and check whether it lands near the inferred ~120 kΩ or closer to the ~700 kΩ DFT value.","supporting_citations":[{"cited_title":"Truncated proposals for scalable and hassle-free simulation-based inference","cited_arxiv_id":null,"evidence_quote":"Supplies the SNPE variant with truncated proposals that the chapter uses for sequential posterior estimation."},{"cited_title":"Contact resistance between carbon nanotubes","cited_arxiv_id":null,"evidence_quote":"Provides the DFT-calculated CNT-CNT junction resistance (~700 kΩ) that the inferred intersection resistance is compared against."},{"cited_title":"Fast ε-free inference of simulation models with bayesian conditional density estimation","cited_arxiv_id":null,"evidence_quote":"Establishes the SNPE approach of training a conditional density estimator on simulated parameter-data pairs, which the inference relies on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Conducting-AFM measurement of SWNT-SWNT junction resistance (~200 kΩ), an experimental baseline for the inferred value."},{"cited_title":"Fabrication of carbon nanotube field-effect transistors in commercial silicon manufacturing facilities","cited_arxiv_id":null,"evidence_quote":"Measured on/off current, threshold-voltage, and subthreshold-swing distributions from three wafers that serve as the inference target."}],"review_version":1}