{"id":"8c5d36d7-283c-4f32-b064-27bb898ab138","arxiv_id":"2507.05134","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A tandem neural network with surrogate-generated pretraining extracts eight TCAD parameters from I-V curves using ~500 physics simulations, achieving median R2 = 0.99 on experimental WS2 transistors and scaling to 35-parameter HEMT fits.","lead":"Tiny two-dimensional transistors are usually characterized by manually fitting models to electrical curves, which is slow and expert-dependent. This paper shows a neural network can do the fitting automatically using roughly 500 physics-based simulated devices, a claimed 40x reduction over prior machine-learning efforts, and it fits measured monolayer WS2 transistors with median R2 = 0.99.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Simulated-test R2 is computed through the same surrogate used to train the inverse network; a TCAD-based re-evaluation is needed before the >40x data-efficiency claim is secure.","rationale":"The paper's strongest claim is about data efficiency: a >40x reduction in physics-based training data enabled by surrogate-based pretraining. The quantitative evidence for that claim is the simulated-test R2 as a function of training size (Figure 4a), and that evidence is generated by feeding predicted parameters through a forward neural network rather than through Sentaurus Device. Because the same (or a very similar) forward network is used inside the tandem training loop, the simulated-test evaluation is not independent of the training procedure; it measures how well the inverse network can fool the surrogate, not necessarily how well the predicted parameters reproduce the true TCAD model. This is the most load-bearing soft spot because it directly affects the headline comparison to prior work. The experimental WS2 test is genuinely TCAD-based and shows excellent fits, which gives real support to the method's practical utility. However, that experimental test is a single configuration and does not probe the bootstrap size sweep or the pretraining-vs-no-pretraining difference, so it cannot by itself establish the quantitative data-efficiency claims. The reader's weakest assumption, the physical validity of the TCAD model for the measured WS2 transistors, is also important and is explicitly acknowledged by the authors as a general caveat of any fitting approach; it concerns the physical meaning of extracted parameters rather than the fitting performance. The surrogate-evaluation issue is more specific to the paper's central novelty and is testable. A TCAD-based re-evaluation of the 1,000-device simulated test set is a concrete, feasible check: it requires running Sentaurus on the predicted parameters for the held-out set and recomputing R2. If the re-evaluation reproduces the reported numbers, the concern is resolved and the conditional acceptance stands. If not, the data-efficiency claim would need to be restated with TCAD-based metrics, and the verdict would move toward conditional acceptance with stricter requirements or rejection of the headline comparison. Since the current CONDITIONAL verdict already flags the surrogate-based evaluation as a reason for conditionality, the verdict need not change on the basis of this stress-test; the condition should be made explicit and testable.","tokens_in":23952,"tokens_out":7456,"duration_ms":93500,"concrete_test":"For the inverse network trained on 500 physics-based simulations with pretraining, feed its predicted parameters for all 1,000 held-out test devices into the original Sentaurus Device model and recompute R2 between the original and regenerated Id-Vgs curves at Vds = 0.1 V and 1 V, using the same R2 definition as Supplementary Section 4. Compare the median and 5th-quantile R2 to Figure 4a. If they fall below the reported values by more than the forward network's own error (Supplementary Figure S3), or if the 5th-quantile R2 drops below approximately 0.94, the data-efficiency gain is partly an artifact of evaluating through the surrogate.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central data-efficiency claim rests on the simulated-test R2 values in Figure 4a, but those values are not obtained from the physics-based simulator. As described in the 'After training' paragraph, the inverse network's predicted parameters for the 1,000-device test set are fed into a well-trained forward neural network (trained on all 25,000 Sentaurus simulations) to estimate the corresponding Id-Vgs curves, and R2 is computed between those curves and the original simulations. This same forward surrogate is also used during training: the inverse network is trained in tandem with a forward network, and fine-tuning on the physics-based data still calls the forward network to estimate current. Thus, the surrogate appears on both sides of the evaluation: the inverse network can, in principle, exploit regions where the forward network is inaccurate, producing parameters that fool the surrogate without reproducing the true TCAD output. The forward network's own error at the relevant training size is not negligible: for 500-device training, the 5th-quantile R2 between forward-network output and Sentaurus simulation is approximately 0.94 (Supplementary Figure S2b). The reported median R2=0.995 and 5th-quantile R2=0.98 on the simulated test are therefore not an independent check of the data-efficiency claim. The experimental validation on 51 WS2 transistors does use actual Sentaurus simulations with predicted parameters and shows median R2=0.990, which is reassuring and partially mitigates this concern. However, that experimental test covers only one training size and one device geometry; it does not validate the bootstrap size sweep, the with/without-pretraining comparison, or the '>40x fewer than prior work' claim, all of which are quantified on the surrogate-evaluated simulated test. The comparison to prior efforts (20,000-1,000,000 devices) is similarly based on these surrogate-derived numbers.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a deep-learning pipeline for inverse parameter extraction in transistor TCAD modeling. A forward neural network is trained on physics-based Sentaurus Device simulations and is then used to generate a large augmented dataset for pretraining an inverse network that maps Id-Vgs curves to model parameters; the inverse network is fine-tuned on the original physics-based data, and during training the forward network is called in a tandem manner to evaluate current-mismatch losses. The authors report bootstrap experiments showing that pretraining yields high simulated-test R2 after training on roughly 500 physics-based simulations, claim a greater-than-40x reduction in required training data relative to prior work, and validate on 51 experimental monolayer WS2 transistors by re-simulating in Sentaurus with predicted parameters (median R2 = 0.990). They also demonstrate scaling to a 35-parameter ASM-HEMT compact model. The central data-efficiency claim rests on simulated-test R2 values that are computed through the surrogate forward network rather than through direct TCAD re-simulation, which is the main technical concern.","tokens_in":24252,"tokens_out":8868,"duration_ms":96553,"significance":"If the central claims hold, the paper makes a practical contribution: it shows that a surrogate-based pretraining strategy can substantially reduce the number of expensive physics-based simulations needed to train an inverse parameter-extraction network, and it provides a direct experimental check by re-simulating 51 measured WS2 devices in Sentaurus with extracted parameters. The bootstrapped experimental design, the held-out simulated test set, and the public release of code and sample data are clear strengths that support reproducibility. The significance would be strengthened if the simulated-test evaluation were verified by direct TCAD re-simulation, and if the wording distinguished 'physical parameters' from 'effective TCAD model parameters' given the acknowledged model assumptions.","major_comments":[{"comment":"The simulated-test R2 values in Figure 4, including the headline median R2 = 0.995 and the 5th-quantile R2 = 0.98 for 500-device training, are computed by feeding predicted parameters into a forward neural network surrogate rather than into Sentaurus Device. The paper says this allows accuracy to be estimated 'without having to re-run Sentaurus Device simulations,' but this is not an independent check: the inverse network is itself trained in tandem with a forward surrogate, so the evaluation can be optimistic if the inverse network exploits regions where the surrogate is inaccurate. The forward surrogate trained on 500 devices has a 5th-quantile R2 of only about 0.94 (Supplementary Figure S2b), so the surrogate error is non-negligible at the training sizes that anchor the greater-than-40x data-efficiency claim. I request that the authors re-run Sentaurus Device on the predicted parameters for at least the 500- and 1,000-device training settings, or on a random subset of the 1,000-device test set, and report direct TCAD R2 values. The main text should also state explicitly which forward network is used for evaluation at each training size.","section":"After training / Supplementary Section 5"},{"comment":"The paper claims to extract 'physical parameters' such as mobility, Schottky barrier height, and defect densities. The experimental validation establishes that Sentaurus simulations using the predicted parameters reproduce measured Id-Vgs curves with high R2, but high curve-fitting R2 does not by itself establish that the extracted values are the physically true parameters, especially given the acknowledged model assumptions (e.g., zero-mobility band-tail acceptor states, absence of a Fermi-level pinning model, effective-mass density of states, and fixed geometry and parameter ranges). The authors already caution about model validity in the 'Generating the training set' section, but the abstract and title make unqualified claims of physical parameter extraction. To make the central claim proportioned to the evidence, the authors should either validate a subset of extracted parameters against independent measurements (e.g., Hall mobility, Kelvin-probe barrier height, or defect spectroscopy) or explicitly reframe the results as extraction of effective TCAD model parameters whose physical interpretation is conditional on the validity of the Sentaurus model.","section":"Experimental validation and abstract"}],"minor_comments":[{"comment":"The sentence 'In Eq. (7) the main text' should refer to Eq. (6) in the main text, which is the loss function for the inverse network.","section":"Supplementary Section 4"},{"comment":"The standard deviation of the mobility error is quoted as 1.9 cm2 V-1 s-1 in the text, but Figure 4f reports 1.6 cm2 V-1 s-1; please reconcile these numbers.","section":"Figure 4 and main text"},{"comment":"The sentence 'Actual vs. predicted values for (e) mobility, (f) Schottky contact barrier height, and (g) peak donor density in Figures 4f-h' has panel labels inconsistent with the figure; panels f through h show mobility, barrier height, and ND0, respectively.","section":"Main text, parameter accuracy paragraph"},{"comment":"When claiming a greater-than-40x reduction in training devices, please specify in the main text that the comparison is to prior compact-model parameter-extraction efforts and that the factor refers to the number of physics-based simulations for a fixed geometry and parameter range, so that the claim is not misread as a general computational-cost reduction.","section":"Introduction and conclusion"},{"comment":"The exclusion of the one device that could not reach an on-state current of 1 uA/um at Vgs = 50 V is a selection criterion; please state this criterion in the main text when describing the 51-device test set, since it affects how the reported median R2 should be interpreted.","section":"Supplementary Section 7 / main text"}],"recommendation":"major_revision","confidential_remarks":"The surrogate-based evaluation is the key risk for the central data-efficiency claim. I recommend the revision include direct TCAD re-evaluation for at least the 500- and 1,000-device training sizes, and that the authors either add independent validation of the extracted parameters or soften the 'physical parameters' language in the abstract and title."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know: the central data-efficiency claim is real but is partly measured through a surrogate, and the experimental validation is the strongest part of the paper. The key idea is to train a fast forward network on a small TCAD training set, use it to generate 100k cheap curves for pretraining, then fine-tune on the real simulations. That is a simple, sensible use of established tandem-network training, and the combination applied to TCAD-based extraction for 2D FETs is new enough to be worth reporting.\n\nThe paper does several things well. The bootstrap study across training sizes is careful. The authors include a well-reasoned feature-engineering ablation. The experimental test is genuinely convincing: 51 WS2 transistors, with predicted parameters re-simulated in Sentaurus, median R2 = 0.990. They also show scaling to 35-parameter HEMT extraction, which is a useful stress test.\n\nNow the soft spots. The simulated-test R2 in Figure 4a is computed by passing predicted parameters through a well-trained forward network, not through Sentaurus. That is a surrogate-on-surrogate evaluation. The stress-test note worries the inverse network could exploit weaknesses in the forward network. That concern is muted by the fact that the evaluation surrogate is trained on all 25k simulations and reproduces Sentaurus with worst 0.5% R2 of 0.9915, and the experimental validation at the 500-device size independently confirms the headline. But the claim that pretraining roughly doubles effective training size, and the detailed scaling curve, rest only on surrogate evaluation. A determined referee should ask for a TCAD re-evaluation of at least a subset of the bootstrap conditions, or an explicit statement that Figure 4a is an estimate.\n\nA second, more important omission in my view: there is no baseline against conventional optimization-based fitting. For 51 experimental devices, how many Sentaurus calls would a standard Levenberg-Marquardt or derivative-free fit need, and how do the resulting R2 and parameter values compare? The paper sells \"automate parameter extraction,\" but without a traditional-fit comparison, the reader cannot judge whether the ML approach is actually faster or better for someone who only has dozens of devices. This is not a fatal flaw—the method's value for expensive simulators stands on the data-efficiency argument—but the missing baseline makes the engineering advantage less crisp.\n\nThe paper is transparent about the TCAD model assumptions and the need to retrain for new geometries. Code and sample data are public, which earns credit. The full training sets are not, but the experimental fits make the core results reproducible in spirit.\n\nBottom line: this is a solid, useful engineering paper. It deserves a serious referee. I would send it to review, and I would ask the authors to add the TCAD-based check on the bootstrap curve and a conventional-fitting baseline. This will be a useful reference for anyone doing ML-based transistor characterization.","headline":"Solid data-efficiency result for ML-based TCAD parameter extraction, anchored by strong experimental validation; the simulated-test scaling curves need a TCAD-based check before the >40x claim is fully secure.","tokens_in":24909,"tokens_out":3147,"would_cite":true,"duration_ms":34607,"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 paper claims that reverse-engineering transistor parameters from current-voltage curves can be done with about 500 physics-based simulations instead of 20,000 to 1,000,000, using a surrogate-trained inverse network that achieves…","keywords":["deep learning","inverse parameter extraction","TCAD","tandem neural network","surrogate model","two-dimensional transistors","monolayer WS2","high-electron-mobility transistors"],"falsifier":"Train the same pipeline on simulations from a deliberately incomplete forward model, for example omitting nonlocal tunneling through the Schottky barrier, fit the measured WS2 curves, and check whether the extracted barrier heights still land in the physically expected 170-400 meV range; if high $R^2$ persists while the parameters shift unphysically, the fit metric is not validating the physics. Likewise, applying the trained network without retraining to a transistor with a different channel length or oxide thickness should degrade markedly; failure to degrade would contradict the paper's stated limitation.","tokens_in":23758,"feed_emoji":"⚡","tokens_out":8899,"duration_ms":84713,"temperature":0.7,"pith_summary":"To fit a transistor model to measured current-voltage data, one normally must tune many physical parameters, such as mobility, contact barrier height, and defect densities, by hand or through optimizer loops that require expensive simulations. This paper claims that a neural network can learn the inverse map, from measured $I_d$-$V_{gs}$ curves straight to those parameters, using only about 500 physics-based simulations, more than 40 times fewer than earlier deep-learning extraction efforts. The enabling move is to train a cheap surrogate network that imitates the physics-based simulator, use the surrogate to mass-produce a pre-training set, and only then fine-tune the inverse network on the real simulation data. If this works as claimed, rigorous TCAD-based parameter extraction becomes practical for 2D semiconductors and other emerging devices where one simulated curve can cost minutes to hours.","feed_headline":"A neural net needs 40x fewer simulations to fit transistor I-V curves","feed_subtitle":"Surrogate-network pretraining lets TCAD-based parameter extraction work from just ~500 simulated devices.","key_machinery":"The load-bearing object is the tandem inverse network: the inverse network outputs a candidate set of device parameters, and those candidate parameters are fed into a pre-trained forward neural network that approximates the TCAD simulator. The training loss penalizes errors in both the output parameters and the predicted current curves, including log-current and derivatives, which is how the network copes with the fact that different parameter sets can produce nearly identical $I_d$-$V_{gs}$ curves. A second component is the pre-training scheme: the forward surrogate generates about 100,000 augmented curves in under 30 seconds, the inverse network is pre-trained on these, and then fine-tuned on the original set of about 500 physics-based simulations. The input representation also matters: each device enters as eight feature vectors at 32 gate voltages, namely current, log-current, and their derivatives at two drain biases.","core_discovery":"On the paper's own terms, the central discovery is that pre-training on surrogate-generated data makes high-quality inverse fits achievable with dramatically less expensive simulation data: training on simulated $I_d$-$V_{gs}$ data from only about 500 devices yields median $R^2=0.995$ on a simulated test set and $R^2=0.990$ when reverse-engineering measured monolayer WS2 transistors, with median absolute extraction errors of 0.88 cm$^2$ V$^{-1}$ s$^{-1}$ for mobility and 18.4 meV for Schottky barrier height. The authors show that the pre-training step is worth roughly a doubling of the physics-based training set, and that the same pipeline scales to 35-parameter fits of GaN high-electron-mobility transistors when the training set is enlarged to about 16,000 devices. They state the caveat that a fit is only physically meaningful if the forward simulator accurately describes the measured transistors.","pith_inferences":["Because the paper shows the surrogate pre-training is what breaks the data bottleneck, the same two-network recipe is a natural template for other inverse problems with expensive forward simulators, such as photonic nanostructures, batteries, or thermoelectrics, though that is an extension the paper does not make.","The non-uniqueness the paper acknowledges implies that parameter-error metrics understate the true ambiguity, especially for correlated defect parameters; downstream users should treat extracted defect densities as effective values rather than unique physical truths.","A strong test of physical meaningfulness would be to compare extracted parameters against independently measured values, such as Hall mobility or Kelvin-probe barrier heights, on the same devices; the paper does not do this.","Requiring retraining for each geometry is a real limitation; conditioning the network on channel length, oxide thickness, or gating configuration would remove that cost, and the scaling data here suggest it is worth trying."],"forward_implications":["TCAD model fitting for emerging semiconductors becomes feasible when a single simulated curve takes minutes to hours: a 500-device training set can be generated in about 30 minutes on 32 cores.","The method is forward-model-agnostic, so it can import drift-diffusion, Monte Carlo, or quantum-transport simulators rather than being restricted to cheap compact models.","Once trained, the network can rapidly screen large numbers of devices; the authors use it to map mobility and barrier-height distributions across 51 WS2 transistors.","For more complex models, the required training set grows roughly fourfold per ten additional fitted parameters, so 35-parameter fits remain achievable with about 16,000 simulations.","Pre-training provides an accuracy gain roughly equivalent to doubling the size of the physics-based training set."],"supporting_citations":[{"why":"Supplies the Sentaurus Device TCAD simulator used to generate the physics-based training data and the final validation simulations.","marker":"21"},{"why":"Supplies the tandem neural-network training technique that lets the inverse network be trained through errors in predicted current, not just parameters.","marker":"22"},{"why":"The prior deep-learning parameter-extraction studies that establish the 20,000 to 1,000,000-device training-set baseline the paper claims to beat by more than 40 times.","marker":"5-10"},{"why":"The ASM-HEMT compact model used in the scaling experiments to test fitting up to 35 simultaneous parameters.","marker":"11"},{"why":"The source of the experimental monolayer WS2 transistors used as the measured test set.","marker":"29"},{"why":"The derivative-free optimization parameter-extraction baseline whose HEMT parameter set and ranges are adapted for the high-dimensional scaling study.","marker":"4"}],"fun_headline_variants":["AI fits transistor curves with 40x less data","Deep learning fits transistor I-V curves from ~500 simulations","Neural network extracts transistor parameters with 40x fewer sims","Fit 2D transistor models with 40x fewer simulations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire extraction is only as sound as the TCAD model used to generate its training data; if that drift-diffusion, Schottky-contact, defect-profile model does not describe the measured WS2 transistors, the network can still achieve high $R^2$ fits while returning physically wrong parameters, a caveat the paper states in its training-set section.","fun_headline_variants_meta":{"raw":{"variants":["AI fits transistor curves with 40x less data","Deep learning fits transistor I-V curves from ~500 simulations","Neural network extracts transistor parameters with 40x fewer sims","Fit 2D transistor models with 40x fewer simulations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0002,"raw_usage":{"total_tokens":1388,"prompt_tokens":971,"completion_tokens":417,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":587,"completion_tokens_details":{"reasoning_tokens":348}},"tokens_in":587,"tokens_out":417,"duration_ms":5526,"temperature":1.0,"reasoning_tokens":348,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:31:57.485105+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same pipeline on simulations from a deliberately incomplete forward model, for example omitting nonlocal tunneling through the Schottky barrier, fit the measured WS2 curves, and check whether the extracted barrier heights still land in the physically expected 170-400 meV range; if high $R^2$ persists while the parameters shift unphysically, the fit metric is not validating the physics. Likewise, applying the trained network without retraining to a transistor with a different channel length or oxide thickness should degrade markedly; failure to degrade would contradict the paper's stated limitation.","supporting_citations":[],"review_version":1}