Pith. sign in

REVIEW 4 major objections 8 minor 37 references

Automated HEMT Model Construction from Datasheets via Multi-Modal Intelligence and Prior-Knowledge-Free Optimization

T0 review · 4 major / 8 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that a fully automated pipeline can turn a raw PDF datasheet into a simulation-ready ASM-HEMT SPICE model, with DC fit errors below 5 percent and RF S-parameter errors below 19 percent across 17 commercial devices.

desk verdict A genuine datasheet-to-SPICE pipeline that runs on 17 devices, but the accuracy claim rests on in-sample RMSEs and the promised independent check is never reported. read the letter →

arxiv 2507.21430 v1 pith:4U2AX25B submitted 2025-07-29 cs.AR

Yuang Peng , Jiarui Zhong , Yang Zhang , Hong Cai Chen This is my paper
classification cs.AR
keywords ASM-HEMTparameterextractiondatasheetdigitizationlargelanguagemodelcomputervisionBayesianoptimizationtree-structuredParzenestimatorSPICEmodeling
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that a complete, simulation-ready ASM-HEMT SPICE model can be generated from a manufacturer's PDF datasheet with no manual parameter extraction and no device-specific prior knowledge. It claims to be the first fully automated end-to-end workflow of this kind, and validates it on 17 commercial HEMTs from 10 manufacturers, reporting DC fitting errors of 1.17 to 4.85 percent RMSE and RF S-parameter errors of 9.06 to 18.7 percent RMSE. If true, circuit designers could skip both measurement equipment and manual digitization when manufacturers do not supply compact models.

What carries the argument

The load-bearing pieces are two. First, the parsing stack: a layout-analysis model splits each PDF page into text, table, and figure regions; a large language model classifies those regions; and a curve digitizer reconstructs I-V plots by detecting axes and labels with an object detector, reading calibration values with OCR, locating axes with a Hough transform, segmenting curve trajectories, then mapping pixel coordinates to physical coordinates through per-axis affine (or log-linear) transformations. Second, the IF-TPE optimizer: a nested loop in which an outer loop contracts and recenters the search hyperrectangle around the current best point with an exponentially decaying contraction rate, while an inner TPE loop — built from Gaussian-mixture kernel density estimates and expected-improvement acquisition — explores the focused window. Adaptive precision quantization shrinks the discretization step as the window contracts, converting broad global search into fine local refinement.

What would settle it

Take a device for which an independent, manufacturer-verified ASM-HEMT model exists, run the pipeline on its datasheet, and compare the generated model's simulated I-V and S-parameter behavior to the independent model at bias points and frequencies outside the datasheet's plotted curves. If multiple fitted parameter sets all achieve the reported RMSEs on the datasheet curves yet diverge sharply off-curve, the claim that the workflow produces accurate device models would be refuted; the same test is passed only if the generated model tracks the independent model off-curve. A second, simpler check is whether the generated models reproduce the datasheet's tabulated threshold voltage, saturation current, and on-resistance, which the paper announces but never reports.

Watch

Extended reading notes

Core claim

The central claim is that the two obstacles to datasheet-only modeling — heterogeneous document layouts and ill-posed high-dimensional parameter fitting under sparse data — can both be automated away. The paper reports that a multimodal pipeline (layout analysis, LLM-based element classification, and a curve digitizer built from object detection, OCR, and affine calibration) correctly parsed all 17 datasheets, and that the IF-TPE optimizer fit the extracted I-V curves and, for four RF devices, S-parameter tables with the stated RMSEs. The authors assert this makes the generated models agree closely with the datasheet characteristics and that the workflow outperforms traditional manual extraction in efficiency and accuracy.

Load-bearing premise

The paper assumes that roughly 40 core ASM-HEMT parameters plus the parasitic elements are uniquely and meaningfully pinned down by the few curves a datasheet provides, so that a good fit to those curves is the same as a good model of the device.

Editorial extensions

If this is right

  • Engineers can obtain simulation-ready ASM-HEMT models for parts whose manufacturers publish only datasheets, without owning measurement instruments.
  • The pipeline's document-processing success on 17 heterogeneous datasheets suggests it could generalize beyond HEMTs to other device families whose datasheets contain similar I-V and S-parameter curves.
  • Because the optimizer needs no device-specific priors, a new part can be modeled from a cold start in a few hundred simulator evaluations, typically under 500 iterations.
  • The reported S-RMSE range of 9 to 19 percent indicates RF models usable for first-pass design but noticeably less precise than the DC fits, marking where further work is needed.
  • If models built this way are trusted, circuit design cycles can shorten because manual extraction and measurement steps vanish.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The reported DC and RF fit errors on datasheet curves do not by themselves guarantee predictive accuracy outside those curves; a stronger test would be to benchmark the generated models against manufacturer-supplied ASM-HEMT models or independent measurements at bias points not in the datasheet.
  • The independent validation against tabulated values (threshold voltage, saturation current, on-resistance) promised in Section II.B is not reported in Section IV; adding such a table would be a cheap, high-value check on whether the fitted parameters have physical meaning rather than being purely curve-fitting artifacts.
  • Since only 4 of 17 devices had S-parameter data, the RF claims rest on a small sample; extending the dataset with more S-parameter-bearing datasheets would sharpen the estimate of RF reliability.
  • The IF-TPE idea of batch-wise search-window contraction plus adaptive quantization is general and could transfer to other compact-model parameter extraction problems, such as BSIM or PSP, that face similarly sparse datasheet data.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 8 minor

Summary. The paper presents an end-to-end framework that takes a PDF datasheet and produces a simulation-ready ASM-HEMT SPICE model without manual intervention. The pipeline combines a layout-analysis model (EDocNet), an LLM (Deepseek) for semantic classification of tables and figures, a custom curve-digitization module (EDocCurve), and a new Bayesian optimization algorithm called IF-TPE that iteratively contracts the parameter search space. The authors validate the workflow on 17 commercial HEMT devices from 10 manufacturers, reporting document-processing success for all 17, IV-RMSE values between 1.17% and 4.85%, and S-RMSE values between 9.06% and 18.7% for the four devices that provide S-parameter tables. The central claim is that this is the first fully automated datasheet-to-SPICE workflow and that the generated models show excellent agreement with published DC and RF characteristics.

Significance. If the claims are supported, the framework would be practically valuable: it addresses a real bottleneck in device modeling by removing manual digitization and parameter extraction, and the 17-device, 10-manufacturer dataset is a useful stress test for layout heterogeneity. The proposed IF-TPE algorithm is also a reasonable contribution to derivative-free optimization under sparse data. However, the current evidence is largely in-sample: the reported RMSE values are the same objectives that IF-TPE minimizes, and the independent validation that the paper itself promises in Section II.B is not reported. The lack of an out-of-sample check, combined with the high parameter count relative to datasheet data, means the headline accuracy claims are not yet established. The paper would be significantly strengthened by adding the promised datasheet-table comparisons, holdout evaluations, and an identifiability discussion.

major comments (4)
  1. [IV.C / III.B.1] The headline accuracy metrics in Table II are in-sample fit residuals, not validation errors. IV-RMSE is exactly the normalized root-mean-square objective defined in Eq. (2), evaluated on the digitized I-V curves that IF-TPE is given, and S-RMSE is the Eq. (3) objective on the same S-parameter tables used for fitting. Reporting these values as evidence of 'excellent agreement' (Abstract) and as 'dual validation' (Section IV.C) is circular. The manuscript itself identifies the correct remedy in Section II.B: datasheet-table 'indicative parameters' such as threshold voltage, saturation current, and on-resistance are intended to serve as 'critical, independent validation benchmarks after completion of the parameter extraction process.' Section IV never reports this comparison for any of the 17 devices. Please add (i) a table comparing simulated versus datasheet-table values for Vth, Idsat, and Ron across all devices; (ii) a holdout analysis, for example fitting on a subset of gate-voltage curves and reporting error on held-out curves; and (iii) a clear statement of which metrics in Table II are in-sample versus out-of-sample.
  2. [II.A / IV.A] The paper does not address the identifiability of roughly 40 core ASM-HEMT parameters plus the extrinsic parasitic network from datasheet-scale data. Section II.A says 'over 40 core parameters' are selected, and Section IV.A reports that only four of the seventeen devices provide S-parameters; the remaining devices are fitted from one or two I-V families and one transfer curve. With only hundreds of data points and a broad initial hyper-rectangle, many physically distinct parameter vectors can plausibly produce nearly identical RMSE on the fitted curves. If that is the case, the reported in-sample RMSEs say little about whether the generated model predicts behavior outside those curves, which is what 'usable SPICE model' requires. Please add an identifiability or sensitivity analysis: for example, report the spread of extracted parameter vectors across independent optimization runs, examine parameter correlations or the conditioning of the local optimum, and run a leave-one-curve-out test to show that extrapolation error remains bounded.
  3. [Table II / Fig. 12] The S-RMSE of 18.7% for FHX04LG is a substantial outlier relative to the other three RF devices (9.09%, 14.5%, 9.06%), yet the text in Section IV.C describes the S-parameter results as exhibiting 'very low S-RMSE' and 'excellent consistency.' Even if these are in-sample fits, an 18.7% error on the objective being minimized indicates that the RF fit for this device is not of the same quality as the others. The paper should either explain why this device is harder (for example, datasheet curve quality, bias conditions, or parameter bounds) or temper the uniform robustness claim.
  4. [IV.B / Fig. 9] The data-extraction stage is validated only qualitatively. Fig. 9(d) shows a visual overlay of the digitized curve on the original datasheet image, and Fig. 10 shows one successfully parsed S-parameter table, but no quantitative digitization error is reported. Since digitization error and calibration error propagate into the parameter-fitting objective, the end-to-end accuracy claim needs a numeric assessment of EDocCurve. Please report affine-calibration residuals, or compare EDocCurve output against manually digitized ground truth for a sample of curves spanning linear and logarithmic axes across the 17-device dataset.
minor comments (8)
  1. [Table II] The table header contains a typo: 'TABEL II' should be 'TABLE II'.
  2. [Table I] The parameter named 'RHT0' in Table I is inconsistent with the standard ASM-HEMT thermal-resistance parameter 'RTH0' used later in the text; please standardize the notation.
  3. [III.B.1] Equation (1) appears to be missing its actual formula, and the text jumps to Eqs. (2) and (3); please ensure all equation numbers are assigned and referenced consistently.
  4. [III.B.3 / Algorithm 1] The pseudo-code and surrounding text use inconsistent symbols for the candidate point ('x_next' vs. theta) and do not define all quantities appearing in Eqs. (4)-(7), such as the threshold y* and the kernel-density estimators l(x) and g(x); this makes the algorithm description harder to reproduce.
  5. [III.B.1] The term 'prior-knowledge-free' is misleading because the initial search space Omega is explicitly constructed using 'the physical significance of parameters.' The paper should clarify that only device-specific priors are excluded, not general semiconductor physics knowledge.
  6. [I] The introduction claims the framework's performance is 'significantly superior to traditional manual methods,' but no comparison against manual extraction or against prior automated tools such as ModelGen [21] is provided; either add a baseline comparison or soften the claim.
  7. [IV.C] The sentence 'the parameter optimization tasks for most devices converged in fewer than 500 iterations' is imprecise given that EPC2037 requires 900 iterations and SAV551 requires 700; please state the distribution or give a median.
  8. [Throughout] Device names are typeset inconsistently, for example 'SA V551' and 'TA V581' in Table II, which should likely be 'SAV-551+' and 'TAV-581' as in the text; please standardize.

Circularity Check

1 steps flagged · score 6.0 of 10

Table II's RMSE values are the same losses IF-TPE minimizes; the promised independent datasheet-table validation is never reported, so the headline 'excellent agreement' is an in-sample fit report.

  1. fitted input called prediction [Section IV.C, Table II; objectives in Section III.B.1, Eqs. (2)-(3)]
    "The I -V RMSE and S -RMSE columns present the final normalized root -mean-square errors for the DC and RF parameter optimization, respectively, demonstrating the high accuracy of the model."

    These RMSEs are the objective functions defined in Section III.B.1: Eq. (2) is the I-V RMSE and Eq. (3) is the S-parameter RMSE over the digitized datasheet points, and the optimizer minimizes them against exactly those curves. Reporting the post-optimization values as 'high accuracy' and later as 'dual validation results' means the evidence for accuracy is the value of the loss being minimized, so the agreement is in-sample by construction.

full rationale

The quantitative accuracy claim in the Abstract and Section V ('excellent agreement with published DC and RF characteristics', 'precise, usable SPICE models') is supported only by Table II's IV-RMSE and S-RMSE values. Those values are the values of the optimization objectives defined in Eqs. (2) and (3), evaluated on the same digitized datasheet curves that were fed to IF-TPE. Reporting the minimized objective as 'validation' is therefore an in-sample fit report, not an independent test. The paper itself identifies the only available out-of-sample check: Section II.B states that datasheet table values (threshold voltage, saturation current, on-resistance) should serve as 'critical, independent validation benchmarks after completion of the parameter extraction process,' but Section IV never reports this comparison. Without that check, the 40+ parameter model could fit the sparse I-V and (for most devices absent) S-parameter data while misrepresenting other operating points; the RMSEs do not establish a 'usable' SPICE model. The extraction pipeline (EDocNet/EDocCurve) is shown with visual overlays and 17/17 document-success, and the use of the authors' own EDocNet is a normal component citation, not by itself circular. The circularity score is 6 because one central quantitative validation step reduces by construction, though the pipeline components retain independent content.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central fitting operation is a 40+ parameter optimization per device, yet the paper reports only aggregate RMSEs, never the extracted values, their identifiability, or their proximity to physically expected ranges. The IF-TPE schedule constants and the per-parameter bounds are absent, so the claimed efficiency cannot be audited. All fitted quantities target the same sparse datasheet curves, which is the core well-posedness risk. The paper introduces no new physical entity, force, or conserved quantity, so the invented-entities ledger is empty; the contributions are software methods (EDocCurve, IF-TPE) and their integration.

free parameters (4)
  • 40+ ASM-HEMT core parameters (VOFF, U0, UA, UB, VSAT, LAMBDA, RTH0, RS, RD, etc.) = not reported per device
    These are the decision variables of the optimization (Section III.B); only aggregate RMSEs are reported, so identifiability and physical plausibility cannot be audited.
  • Extrinsic parasitic network parameters (Cpg, Cpd, Lg, Ld, Ls, Rg, Rd, Rs) = not reported
    Fitted in the S-parameter optimization stage (Section II.A.2, Fig. 1); values are absent, and the I-V and S-parameter fits are done separately without a unified revalidation.
  • IF-TPE schedule constants (initial contraction rate, decay constant, grid-point count, batch sizes, quantile… = not reported
    Algorithm 1 defines the structure but the text gives no numerical values, so the fewer-than-500-iterations claim cannot be reproduced or sensitivity-tested.
  • Per-parameter search bounds in the initial hyper-rectangle Omega = not tabulated
    The prior-knowledge-free setting still needs hard bounds (Section III.B.1); the actual bounds are not listed, and their breadth drives both convergence speed and solution quality.
assumptions (5)
  • domain assumption ASM-HEMT v101.4.0, as shipped in ngspice 42, correctly implements the industry-standard model that the extracted parameters target.
    All simulations use this public implementation (Section IV setup); any implementation discrepancy becomes a fitting bias that cannot be detected in-sample.
  • domain assumption Datasheet curves and tables describe one consistent device under consistent bias and temperature conditions.
    Datasheets publish typical values that may come from different samples or test conditions (Section II.B); the fit treats I-V curves and S-parameter tables as jointly consistent.
  • domain assumption The digitization stage (EDocCurve) introduces negligible error relative to fitting error.
    No quantitative digitization accuracy is reported (Section III.A.2, Fig. 9), yet the digitized curves are the exclusive optimization targets.
  • ad hoc to paper The reduced 40+ parameter subset, with broad physics-based bounds, is expressive enough to reproduce any of the 17 devices on the datasheet-observable curves.
    The subset is chosen 'based on physical significance and correlation with I-V and S-parameters' (Section II.A.1); no identifiability or coverage analysis is given.
  • ad hoc to paper The IF-TPE contraction schedule (Eqs. 10-11) and adaptive quantization (Eq. 12) preserve convergence to a good optimum within the reported iteration budgets.
    No convergence guarantee is proven for the contracted, quantized space; convergence is asserted empirically with one run per device (Table II, Section IV.C).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Automated HEMT Model Construction from Datasheets via Multi-Modal Intelligence and Prior-Knowledge-Free Optimization." pith.science (2026). https://pith.science/paper/4U2AX25B

@misc{pith2026250721430,
  author       = {Pith},
  title        = {Pith review of: Automated HEMT Model Construction from Datasheets via Multi-Modal Intelligence and Prior-Knowledge-Free Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4U2AX25B}},
  note         = {Machine review of arXiv:2507.21430}
}
read the original abstract

Parameter extraction for industry-standard device models like ASM-HEMT is crucial in circuit design workflows. However, many manufacturers do not provide such models, leaving users to build them using only datasheets. Unfortunately, datasheets lack sufficient information for standard step-by-step extraction. Moreover, manual data extraction from datasheets is highly time-consuming, and the absence of a fully automated method forces engineers to perform tedious manual work. To address this challenge, this paper introduces a novel, end-to-end framework that fully automates the generation of simulation-ready ASM-HEMT SPICE models directly from PDF datasheets. Our framework is founded on two core innovations: 1) a multi-modal AI pipeline that integrates computer vision with a large language model (LLM) to robustly parse heterogeneous datasheet layouts and digitize characteristic curves, and 2) a novel Iterative-Focusing Tree-structured Parzen Estimator (IF-TPE) optimization algorithm is specifically designed for device parameter extraction under the high-dimensional, sparse-data condition by adaptively refining the parameter search space. Experimental validation on a diverse set of 17 commercial HEMT devices from 10 manufacturers confirms the framework's accuracy and robustness. The generated models demonstrate excellent agreement with published DC and RF characteristics. As the first fully automated workflow of its kind, our proposed solution offers a transformative approach to device modeling, poised to significantly accelerate the circuit design cycle by eliminating the need for manual parameter extraction.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 32 canonical work pages

  1. [1]

    GaN -Based RF Power Devices and Amplifiers,

    U. K. Mishra, L. Shen, T. E. Kazior, and Y. -F. Wu, “GaN -Based RF Power Devices and Amplifiers,” Proceedings of the IEEE , vol. 96, no. 2, pp. 287–305, Feb. 2008, doi: 10.1109/JPROC.2007.911060

    work page arXiv 2008

  2. [2]

    Recent Developments, Reliability Issues, Challenges and Applications of GaN HEMT Technology,

    T. Garg and S. Kale, “Recent Developments, Reliability Issues, Challenges and Applications of GaN HEMT Technology,” IEEE Electron Devices Reviews , vol. 1, pp. 16 –30, Oct. 2024, doi: 10.1109/EDR.2024.3491716

    work page arXiv 2024

  3. [3]

    TCAD Modeling of GaN HEMT Output Admittance Dispersion through Trap Rate Equation Green’s Functions,

    E. Catoggio, S. Donati Guerrieri, and F. Bonani, “TCAD Modeling of GaN HEMT Output Admittance Dispersion through Trap Rate Equation Green’s Functions,” Electronics, vol. 12, no. 11, p. 2457, May 2023, doi: 10.3390/electronics12112457

    work page doi:10.3390/electronics12112457 2023

  4. [4]

    Multiphysics Modeling of RF and Microwave High - Power Transistors,

    P. H. Aaen et al., "Multiphysics Modeling of RF and Microwave High - Power Transistors," IEEE Transactions on Microwave Theory and Techniques, vol. 60, no. 12, pp. 4013 -4023, Dec. 2012, doi: 10.1109/TMTT.2012.2224366

    work page arXiv 2012

  5. [5]

    TCAD modeling of a lateral GaN HEMT using empirical data,

    M. R. Hontz, R. Chu, and R. Khanna, “TCAD modeling of a lateral GaN HEMT using empirical data,” in 2018 IEEE Applied Power Electronics Conference and Exposition (APEC) , Mar. 2018, pp. 244 –

    work page 2018

  6. [6]

    Extensions of the Chalmers nonlinear HEMT and MESFET model,

    I. Angelov, L. Bengtsson, and M. Garcia, “Extensions of the Chalmers nonlinear HEMT and MESFET model,” IEEE Transactions on Microwave Theory and Techniques , vol. 44, no. 10, pp. 1664 –1674, Oct. 1996, doi: 10.1109/22.538957

    work page doi:10.1109/22.538957 1996

  7. [7]

    Semi -physical nonlinear circuit model with device/physical parameters for HEMTs,

    H. Otsuka et al. , “Semi -physical nonlinear circuit model with device/physical parameters for HEMTs,” International Journal of Microwave and Wireless Technologies , vol. 3, no. 1, pp. 25 –33, Feb. 2011, doi: 10.1017/s1759078711000043

    work page doi:10.1017/s1759078711000043 2011

  8. [8]

    Assessment and Comparison of Measurement-Based Large-Signal FET Models for GaN HEMTs

    R. P. Martinez, M. Iwamoto, J. J. Xu, et al. "Assessment and Comparison of Measurement-Based Large-Signal FET Models for GaN HEMTs." IEEE Transactions on Microwave Theory and Techniques , vol. 72, no. 5, pp. 2692-2703, 2024

    work page 2024

Show all 37 references
  1. [9]

    A Scalable Knowledge -Based Neural Network Model for GaN HEMTs With Accurate Trapping and Self-Heating Effects Characterization

    M. Li, Q. -J. Zhang, F. Feng, et al. "A Scalable Knowledge -Based Neural Network Model for GaN HEMTs With Accurate Trapping and Self-Heating Effects Characterization." IEEE Transactions on Microwave Theory and Techniques, vol. 71, no. 9, pp. 3747-3760, 2023

    2023

  2. [10]

    A Hybrid GaN HEMT Model Merging Artificial Neural Networks and ASM -HEMT for Parameter Precision and Scalability,

    Z. Lu et al., “A Hybrid GaN HEMT Model Merging Artificial Neural Networks and ASM -HEMT for Parameter Precision and Scalability,” IEEE Trans. Electron Devices , vol. 71, no. 12, pp. 7334 –7342, Dec. 2024, doi: 10.1109/ted.2024.3478181

    2024

  3. [11]

    Artificial Neural Network Nonlinear Transistor Behavioral Models: Structure and Parameter Determination Process Based on the Cardiff Model,

    M. Tian, J. J. Bell, R. Quaglia, E. M. Azad, and P. J. Tasker, “Artificial Neural Network Nonlinear Transistor Behavioral Models: Structure and Parameter Determination Process Based on the Cardiff Model,” IEEE Transactions on Microwave Theory and Techniques , vol. 73, no. 2, p...

    2025

  4. [12]

    Surface-Potential-Based Compact Modeling of Gate Current in AlGaN/GaN HEMTs

    S. Ghosh, A. Dasgupta, S. Khandelwal, S. Agnihotri, and Y. S. Chauhan, "Surface-Potential-Based Compact Modeling of Gate Current in AlGaN/GaN HEMTs", IEEE Transactions on Electron Devices, vol. 62, no. 2, Feb. 2015

    2015

  5. [13]

    Device Modeling IC -CAP,

    Keysight, “Device Modeling IC -CAP,” Keysight. Accessed: Jul. 14,

  6. [14]

    Physics-Based Multi -Bias RF Large -Signal GaN HEMT Modeling and Parameter Extraction Flow,

    S. Aamir Ahsan, S. Ghosh, S. Khandelwal, and Y. S. Chauhan, “Physics-Based Multi -Bias RF Large -Signal GaN HEMT Modeling and Parameter Extraction Flow,” IEEE Journal of the Electron Devices Society, vol. 5, no. 5, pp. 310 –319, Sep. 2017, doi: 10.1109/JEDS.2017.2724839

    2017

  7. [15]

    ASM GaN: Industry Standard Model for GaN RF and Power Devices —Part-II: Modeling of Charge Trapping,

    S. A. Albahrani, D. Mahajan, J. Hodges, Y. S. Chauhan, and S. Khandelwal, “ASM GaN: Industry Standard Model for GaN RF and Power Devices —Part-II: Modeling of Charge Trapping,” IEEE Transactions on Electron Devices, vol. 66, no. 1, pp. 87 –94, Jan. 2019, doi: 10.1109/TED.2018.2868261

    2019

  8. [16]

    [Tool] Automatically Extracting Hardware Descriptions from PDF Technical Documentation ,

    N. Hauser and J. Pennekamp, "[Tool] Automatically Extracting Hardware Descriptions from PDF Technical Documentation ," Journal of Systems Research, vol. 3, no. 1, pp. 1-21, 2023

    2023

  9. [17]

    Automated Extraction of Data from MOSFET Datasheets for Power Converter Design Automation,

    F. Tian, Q. Sui, D. B. Cobaleda and W. Martinez , "Automated Extraction of Data from MOSFET Datasheets for Power Converter Design Automation," IEEE Journal of Emerging and Selected Topics in Power Electronics, vol. 12, no. 6, pp. 5648-5660, 2024

    2024

  10. [18]

    SpecLLM: Exploring Generation and Review of VLSI Design Specification with Large Language Model,

    M. Li, W. Fang, Q. Zhang, and Z. Xie, “SpecLLM: Exploring Generation and Review of VLSI Design Specification with Large Language Model,” Jan. 24, 2024, arXiv: arXiv:2401.13266. doi: 10.48550/arXiv.2401.13266

    2024 doi arXiv

  11. [19]

    TinyChart: Efficient Chart Understanding with Visual Token Merging and Program -of-Thoughts Learning,

    L. Zhang et al., “TinyChart: Efficient Chart Understanding with Visual Token Merging and Program -of-Thoughts Learning,” Apr. 25, 2024, arXiv: arXiv:2404.16635. doi: 10.48550/arXiv.2404.16635

    2024 doi arXiv

  12. [20]

    DocEDA: Automated Extraction and Design of Analog Circuits from Documents with Large Language Model,

    H. C. Chen, L. Wu, M. Gao, L. Shen, J. Zhong, and Y. Xu, “DocEDA: Automated Extraction and Design of Analog Circuits from Documents with Large Language Model,” Nov. 25, 2024, arXiv: arXiv:2412.05301. doi: 10.48550/arXiv.2412.05301

    2024 doi arXiv

  13. [21]

    ModelGen: Automating Semiconductor Parameter Extraction with Large Language Model Agents,

    Y. Wei et al. , “ModelGen: Automating Semiconductor Parameter Extraction with Large Language Model Agents,” ACM Trans. Des. Autom. Electron. Syst., May 2025, doi: 10.1145/3736165

    2025 doi

  14. [22]

    DocGenome: An Open Large-scale Scientific Document Benchmark for Training and Testing Multi -modal Large Language Models,

    R. Xia et al., “DocGenome: An Open Large-scale Scientific Document Benchmark for Training and Testing Multi -modal Large Language Models,” Jun. 17, 2024, arXiv: arXiv:2406.11633. doi: 10.48550/arXiv.2406.11633

    2024 doi

  15. [23]

    EDocNet: Efficient Datasheet Layout Analysis Based on Focus and Global Knowledge Distillation,

    H. C. Chen, L. Wu, and Y. Zhang, “EDocNet: Efficient Datasheet Layout Analysis Based on Focus and Global Knowledge Distillation,” Feb. 23, 2025, arXiv: arXiv:2502.16541. doi: 10.48550/arXiv.2502.16541

    2025 doi arXiv

  16. [24]

    Hybrid GWOCS Optimization Based Parameter Extraction Method Applied to GaN Devices,

    A. Abushawish and A. Jarndal, “Hybrid GWOCS Optimization Based Parameter Extraction Method Applied to GaN Devices,” in 2021 IEEE International Midwest Symposium on Circuits and Systems, MWSCAS , pp. 1035 –1039, Aug. 2021. doi: 10.1109/MWSCAS47672.2021.9531703

    2021

  17. [25]

    A parameter extraction method for GaN HEMT empirical large‐signal model including self‐heating and trapping effects,

    Z. Wen, Y. Xu, C. Wang, X. Zhao, Z. Chen, and R. Xu, “A parameter extraction method for GaN HEMT empirical large‐signal model including self‐heating and trapping effects,” Int J Numerical Modelling, vol. 30, no. 1, Jan. 2017, doi: 10.1002/jnm.2137

    2017 doi

  18. [26]

    Machine Learning-Based Large-Signal Parameter Extraction for ASM-HEMT

    F. Chavez and S. Khandelwal , "Machine Learning-Based Large-Signal Parameter Extraction for ASM-HEMT." IEEE Microwave and Wireless Technology Letters, vol. 34, no. 2, 147-150, 2024

    2024

  19. [27]

    Improving Semiconductor Device Modeling for Electronic Design Automation by Machine Learning Techniques,

    Z. Wang et al. , “Improving Semiconductor Device Modeling for Electronic Design Automation by Machine Learning Techniques,” IEEE Trans. Electron Devices , vol. 71, no. 1, pp. 263 –271, Jan. 2024, doi: 10.1109/TED.2023.3307051

    2024

  20. [28]

    iPREFER: An Intelligent Parameter Extractor based on Features for BSIM -CMG Models,

    Z. Peng, Y. Wang, Z. Yuan, and X. Wang, “iPREFER: An Intelligent Parameter Extractor based on Features for BSIM -CMG Models,” Apr. 11, 2024, arXiv: arXiv:2404.07827. doi: 10.48550/arXiv.2404.07827

    2024 doi arXiv

  21. [29]

    A Physics -Based Analytical Model for 2DEG Charge Density in AlGaN/GaN HEMT 12 > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) < Devices,

    S. Khandelwal, N. Goyal, and T. A. Fjeldly, “A Physics -Based Analytical Model for 2DEG Charge Density in AlGaN/GaN HEMT 12 > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) < Devices,” IEEE Transactions on Electron Devices , vol. 58, no. 10, pp. 3...

    2011

  22. [30]

    A physics based compact model of I–V and C–V characteristics in AlGaN/GaN HEMT devices,

    S. Khandelwal and T. A. Fjeldly, “A physics based compact model of I–V and C–V characteristics in AlGaN/GaN HEMT devices,” Solid- State Electronics , vol. 76, pp. 60 –66, Oct. 2012, doi: 10.1016/j.sse.2012.05.054

    2012 doi

  23. [31]

    ASM - HEMT: Compact model for GaN HEMTs,

    A. Dasgupta, S. Ghosh, Y. S. Chauhan, and S. Khandelwal, “ASM - HEMT: Compact model for GaN HEMTs,” in 2015 IEEE International Conference on Electron Devices and Solid -State Circuits (EDSSC), pp. 495–498, Jun. 2015

    2015

  24. [32]

    A new small -signal modeling and extraction method in AlGaN/GaN HEMTs ,

    J. Lu, Y. Wang, L. Ma, and Z. Yu, “A new small -signal modeling and extraction method in AlGaN/GaN HEMTs ,” Solid-State Electron ics, vol. 52, pp. 115–120, Jul. 2008

    2008

  25. [33]

    An Improved Small -Signal Equivalent Circuit for GaN High -Electron Mobility Transistors,

    F. Y. Huang, X. S. Tang, Z. N. Wei, L. H. Zhang and N. Jiang, "An Improved Small -Signal Equivalent Circuit for GaN High -Electron Mobility Transistors," IEEE Electron Device Letters, vol. 37, no. 11, pp. 1399-1402, Nov. 2016, doi: 10.1109/LED.2016.2609462

    2016

  26. [34]

    DeepSeek-V3 Technical Report,

    DeepSeek-AI et al., “DeepSeek-V3 Technical Report,” Feb. 18, 2025, arXiv: arXiv:2412.19437. doi: 10.48550/arXiv.2412.19437

    2025 doi arXiv

  27. [35]

    Singer -Vine and The pdfplumber contributors, pdfplumber

    J. Singer -Vine and The pdfplumber contributors, pdfplumber. (Jun. 2025). Python. Accessed: Jul. 18, 2025. [Online]. Available: https://github.com/jsvine/pdfplumber

    2025

  28. [248]

    doi: 10.1109/APEC.2018.8341017

    2018

  29. [2025]

    Available: https://www.keysight.com/us/en/products/software/pathwave-design- software/device-modeling-products/ic-cap-device-modeling- software.html

    [Online]. Available: https://www.keysight.com/us/en/products/software/pathwave-design- software/device-modeling-products/ic-cap-device-modeling- software.html

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.