Agentic TCAD Calibration Workflow for Oxide Semiconductor Transistors
Gyujun Jeong, Junmo Lee, Sungwon Cho, Woohyun Hwang, Kwangyou Seo, Suhwan Lim, Wanki Kim, Daewon Ha, et al.
physics.ins-det · cs.AI · cs.LG · physics.app-ph
Quick verdict
Claims first agentic TCAD calibration on a fabricated IWO transistor; rigorous transfer tests but no manual/optimizer baseline to prove agent value.
Summary
The paper reports the first experimental demonstration of an LLM-agent-orchestrated TCAD calibration workflow, applied to a fabricated bottom-gate In–W–O (IWO) transistor. The agent (GPT-5.5 backend, executed via OpenAI Codex and the NVIDIA NeMo Agent Toolkit) compares measured and simulated transfer curves, extracts residuals, and proposes bounded parameter corrections or activates additional physical models, using retrieval-augmented generation over the Sentaurus manual and device literature; a human reviewer approves each deck change, and Sentaurus W-2024.09-SP1 performs all device simulation. Updates are accepted only if they reduce a fixed multi-metric objective J (log/linear current RMSE, Vth, Ion, SS errors with pre-set weights). Starting from J=2.589, five accepted updates (effective S/D and gate work functions, fixed charge, effective donor density, Gaussian/tail DOS activation, mobility) yield J=0.181, a 14.3× reduction. The fixed calibrated model, fitted on VD∈{0.05, 1} V at L=1 µm, transfers without retuning to five drain biases (max errors 36.1 mV in Vth, 0.022 decade in Ion) and to a 2 µm channel-length device (46.2 mV, 0.062 decade). A separately recalibrated 4% W device maps the composition change mainly to a reduced effective donor-like density. Main limitations: single-device data with no repeatability statistics, no comparison against manual calibration or non-agentic automated optimization, no code or data release, and residual model ambiguity despite the transferability tests.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- This is the first experimental demonstration of an agentic TCAD calibration workflow, with an LLM agent orchestrating while Sentaurus governs device physics, on a fabricated BG-IWO transistor.
- Five agent-suggested, human-approved updates reduce the multi-metric objective J by 14.3× (2.589 → 0.181) for the 2%-W reference device.
- The fixed calibrated model transfers without retuning across five drain biases (max 36.1 mV Vth, 0.022 decade Ion) and to a 2 µm channel-length device (max 46.2 mV, 0.062 decade), demonstrating transferability rather than a local fit.
- Recalibration of the 4%-W device attributes the composition change mainly to a reduced effective donor-like background density ND,eff, providing process-sensitive insight consistent with prior InOx dopant studies.
Strengths
Pre-registered validation protocol targets the ambiguity problem head-on
Table IV fixes the data split before calibration: fitting uses only VD ∈ {0.05, 1} V at L = 1 µm; the fixed model must then pass five drain-bias tests and L = 2 µm tests without retuning (Sec. III-B), plus a separately recalibrated 4%-W split. This is far stronger than the local-fit-only practice the introduction criticizes.
Disciplined, auditable update machinery
Sec. II-C/D combines bounded OFAT sensitivity ranking with derivative-free Gauss–Newton updates (Eq. 3: bounds ℓ/u, trust region ΔP), acceptance only on convergence plus J reduction, a predefined numerical-recovery sequence, and mandatory human review of every deck change (Table II). This constrains the LLM far more than typical agentic pipelines.
Fixed, transparent objective with uniform metric extraction
Honest parameter semantics
Each accepted update is individually reported with its physics rationale
Weaknesses
No baseline isolates the agent's value; efficiency claim never quantified
Single-device evidence with no uncertainty treatment
Degree of automation is opaque
Model ambiguity is narrowed, not resolved
Reproducibility is low
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper reports the first experimental demonstration of an LLM-agent-orchestrated TCAD calibration workflow, applied to a fabricated bottom-gate In–W–O (BG-IWO) transistor (Sec. II-B, Table III). Claimed contributions: (i) residual-guided selection between bounded parameter corrections and physics-model activation; (ii) a fixed calibrated model that transfers across drain bias and channel length without retuning; (iii) process-sensitive insight from a 4%-W recalibration. The "first" claim is scoped in Table I against prior agentic TCAD work [5] (simulation-defined targets) and AI inverse calibration [4] (predefined physics), so it is carefully worded — but it remains contestable in a fast-moving area. The practical claim of a "faster route" (Abstract, Sec. V) is asserted and never quantified.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is agentic TCAD calibration?
It is an automated calibration loop in which a large-language-model agent compares simulated and measured transistor characteristics, extracts residuals, and proposes bounded parameter changes or activates additional physics models inside a TCAD simulator (here Synopsys Sentaurus). Each proposed change is human-reviewed and accepted only if it improves a predefined multi-metric error objective.
Which device was calibrated in this paper?
A fabricated bottom-gate amorphous In–W–O (IWO) transistor with 2% tungsten, 1 µm channel length, 5 µm width, 3.5 nm IWO channel, 7 nm HfO2 gate dielectric, and Pd contacts, measured with forward ID–VG sweeps at 50 mV and 1 V drain bias. A 2 µm channel-length device and a 4% W device served as held-out and process tests.
How accurate is the final calibrated model?
After five accepted updates the weighted objective J fell 14.3× (2.589 to 0.181). On held-out tests without retuning, maximum errors were 36.1 mV in threshold voltage and 0.022 decade in on-current across five drain biases, and 46.2 mV / 0.062 decade on the 2 µm channel-length device.
Is the code or data from this paper publicly available?
The paper does not report a public code, TCAD deck, or measurement-data release. Reproducing it requires a Synopsys Sentaurus W-2024.09-SP1 license, the fabricated devices, and the OpenAI GPT-5.5 via Codex plus NVIDIA NeMo Agent Toolkit toolchain.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Only figure captions and quoted numbers are present in the extraction; the actual curve overlays, residual heatmaps (Figs. 5–7), and SEM images could not be visually inspected, so qualitative fit quality is judged from reported metrics alone.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.