Source-linked AI summary
Agentic TCAD Calibration Workflow for Oxide Semiconductor Transistors
Gyujun Jeong, Junmo Lee, Sungwon Cho, Woohyun Hwang, Kwangyou Seo, Suhwan Lim, Wanki Kim, Daewon Ha, Rishi Ranade, Kihang Youn, Ram Cherukuri, Yiyi Wang, Asif Khan, Shimeng Yu
TL;DR
Experimental TCAD calibration is slow and expert-dependent because multiple models can fit the same transfer curve, while local fitting alone cannot identify the underlying device physics. The paper demonstrates an agent-orchestrated workflow that uses measurement–TCAD residuals to guide bounded model or parameter updates. For the 2%-W reference device, calibration reduces the objective and the fixed model transfers across bias and geometry without retuning.
Problem
Experimental TCAD calibration is time-consuming and expert-dependent, and multiple model and parameter choices can reproduce a measured transfer curve without local fitting identifying the underlying device physics.
Source: §I · p. 1Method
An AI-agent-orchestrated workflow uses measurement–TCAD residuals to guide bounded parameter corrections or physical-model evaluation and accepts updates only when predefined metrics improve.
Source: §I · pp. 1–2Results
The 2%-W reference device objective J fell 14.3×, and the fixed calibrated model transferred across drain bias and channel length without retuning.
Source: B. Fixed-Model Evaluation Across Bias… · p. 4, §V · p. 6Takeaways & Limitations
The 4%-W process-composition test provides process-sensitive insight by identifying model and parameter classes that reduce distinct composition-related residuals.
Source: §IV · p. 5
Abstract
from arXiv · showhide
Experimental TCAD calibration is essential for predictive technology modeling of emerging oxide semiconductor transistors. However, it remains time-consuming and expert dependent because of model ambiguity. Multiple physical models and parameter sets can reproduce the same measured transfer characteristics, while local fitting alone cannot uniquely identify the underlying device physics. We present the first demonstration of an agentic TCAD calibration workflow for a fabricated bottom-gate In--W--O (BG-IWO) transistor. Starting from the measured transfer curve and device information, the workflow uses measurement--TCAD residuals and local sensitivity tests to select bounded parameter corrections or evaluate additional physical models, and accept only updates that improve device metrics. The LLM agent orchestrates the workflow, while Sentaurus governs the device physics. For the 2\%-W reference device, five agent-suggested updates yield a fixed calibrated model, reducing the multi-metric device objective $J$ by 14.3$\times$. Maximum $V_{\mathrm{th}}$/$I_{\mathrm{on}}$ errors are 36.1~mV/0.022 decade for varying-drain-bias tests and 46.2~mV/0.062 decade for varying-channel-length tests, demonstrating model transferability across bias and geometry rather than a local parameter fit. W-composition tests provide process-sensitive insight. This agentic workflow provides a faster route to model development for emerging device technologies.
I. INTRODUCTION
Experimental TCAD calibration is an inverse-modeling bottleneck: multiple physics models and parameter choices can fit one measured transfer curve, while a local fit does not identify device physics or ensure generalization. This work presents an experimental agentic workflow that uses residuals to guide model and parameter searches and extract device-physics and process insight.
I. INTRODUCTION
- Multiple physics models and parameter choices can fit the same measured transfer curve, so local fitting alone cannot identify the underlying device physics or guarantee model generalization.Calibration is also time-consuming and expert-dependent because it requires model selection, parameter adjustment, deck implementation, simulation, and residual interpretation.Source: §I · p. 1
- Prior AI/ML-assisted TCAD studies addressed adjacent tasks within predefined physics, parameter, or design spaces, rather than experimental-data-driven model selection and parameter calibration.The cited tasks include variation and temperature analysis, inverse parameter calibration, and LLM-based code generation and device optimization.Source: §I · p. 1
- The work presents the first experimental demonstration of an agentic TCAD calibration workflow, using measurement–TCAD residuals to rank physical models and parameter groups and test bounded updates.Updates are accepted only when they improve predefined device metrics; the workflow also extracts device-physics and process insights.Source: §I · pp. 1–2, Figure 1 · p. 1
II. AGENTIC TCAD CALIBRATION WORKFLOW
The workflow uses measurement–TCAD residuals to choose between parameter adjustment and additional physics-model evaluation, accepting updates only when they reduce the device objective. The AI agent coordinates the process, Sentaurus solves the device response, and a human reviewer approves workflow constraints.
A. Workflow Overview and Responsibilities
- The loop compares measured and simulated transfer curves, extracts residuals and device metrics, then accepts a proposed parameter or model update only if it reduces J.Rejected updates lead to another bounded TCAD test.Source: A. Workflow Overview and… · p. 2
- When subthreshold swing is the dominant mismatch, the workflow can add Gaussian and tail DOS models after literature, parameter-range, syntax, and human-review checks.The subsequent Sentaurus run reduces J; bounded OFAT probes and derivative-free Gauss–Newton updates target residual-sensitive parameter groups.Source: A. Workflow Overview and… · p. 2, Figure 2 · p. 2
- The LLM agent orchestrates actions, Sentaurus TCAD computes the device response, and a human reviewer approves permitted deck modifications and workflow instructions.The workflow also specifies calibration-metric checks and numerical-recovery procedures.Source: A. Workflow Overview and… · p. 2, Table II · p. 3
- The device is represented with a fixed 2D TCAD geometry, with an SEM top view of an L = 2 µm transistor serving as a dimension reference.Source: Figure 3 · p. 3
B. Device Structure and Experimental Design
The study calibrates a fabricated bottom-gate In–W–O transistor using measured forward ID–VG sweeps, with geometry and basic material information fixed before fitting. Only the nominal 2%-W, L = 1 µm curves at two drain biases contribute to calibration; other bias and channel-length conditions evaluate the final model without retuning.
B. Device Structure and Experimental Design
- The experimental target is a fabricated bottom-gate In–W–O transistor measured using forward ID–VG sweeps, with geometry and basic material information fixed before fitting.The physical model parameters are calibrated; Table III summarizes device dimensions, materials, and measurement configuration.Source: B. Device Structure and Experimental… · p. 3, Table III · p. 3
- Only the nominal 2%-W, L = 1 µm curves at VD = 50 mV and 1 V contribute to J; other drain-bias and channel-length conditions are reserved for evaluation without retuning.The same current floor, gate-voltage windows, and metric-extraction method are applied at every parameter setting.Source: B. Device Structure and Experimental… · p. 3, Table IV · p. 3
- The calibration objective uses log- and linear-current, Vth, Ion, and SS residual windows.Fig. 4 shows the metric windows used to evaluate errors and extract Vth, Ion, and SS.Source: Figure 4 · p. 3
C. Calibration objective J and sensitivity-based parameter adjustment
The calibration objective scores transfer-curve fit using extracted device metrics, while residuals and local sensitivities guide bounded parameter updates. Proposed changes are checked in a full Sentaurus run before judging whether the fit improves.
C. Calibration objective J and sensitivity-based parameter adjustment
- The multi-metric objective compares simulated and measured transfer curves using weighted, normalized terms and fixed extraction rules.Vth is extracted at |ID| = 100 nA, SS from the average inverse slope between 0.1 and 100 nA, and Ion as maximum |ID|; currents below 10 pA are excluded from log-current error.Source: C. Calibration objective J and… · pp. 3–4
- Residual magnitude, sign, and VG dependence distinguish threshold-voltage, drain-current, and subthreshold-swing mismatches.J quantifies overall fit; each residual is normalized to be dimensionless.Source: C. Calibration objective J and… · p. 4
- OFAT Sentaurus probes vary one parameter at a time to estimate local sensitivities and rank parameter groups that could correct the dominant residual.The sensitivity matrix describes finite-difference changes in the residual vector.Source: C. Calibration objective J and… · p. 4
- DFO-GN computes a bounded trial change for one selected parameter group, then a full Sentaurus run determines whether J improves.The update respects parameter limits, a per-update change limit, and fixed residual weights.Source: C. Calibration objective J and… · p. 4
D. Physics-Model Activation and Update Verification
Additional physics is considered only when bounded adjustment of existing parameters leaves a systematic transfer-curve mismatch. Model changes require review and a converged run that reduces J.
D. Physics-Model Activation and Update Verification
- An additional physics model is considered only if systematic mismatch persists after bounded adjustment of existing parameters.Candidate models are retrieved for documented behavior consistent with the remaining mismatch.Source: D. Physics-Model Activation and Update… · p. 4
- A human reviewer verifies the proposed model’s applicability and parameter range before approving a deck modification.Candidates come from the Sentaurus manual and reviewed device literature.Source: D. Physics-Model Activation and Update… · p. 4
- An update is accepted only when the Sentaurus run converges and reduces J; unrecovered numerical failures are excluded from update decisions.Nonconverged trials undergo a predefined recovery sequence before they are recorded as failures.Source: D. Physics-Model Activation and Update… · p. 4
A. Residual-Guided Calibration of the Reference Device
Five agent-suggested updates calibrate the reference BG-IWO model by addressing residuals in sequence, from electrostatic and charge response to subthreshold swing and transport.
A. Residual-Guided Calibration of the Reference Device
- Five accepted updates reduce the multi-metric objective J from 2.589 to 0.181.The baseline objective is 2.589; the final calibrated model has J = 0.181.Source: A. Residual-Guided Calibration of the… · p. 4
- The first three updates lower source/drain and gate work functions, add Qf = +1.59 × 10^10 cm−2, and raise ND,eff to 2.56 × 10^18 cm−3 as residuals shift from low current to threshold offset and charge deficit.The updates reduce J to 1.537, 0.574, and 0.443, respectively.Source: A. Residual-Guided Calibration of the… · p. 4
- A remaining SS residual prompts Gaussian and tail DOS terms, followed by a mobility-prefactor change from 10 to 24 cm2/Vs, bringing J to 0.181.The fitted variables are calibration-effective descriptors, not direct extractions of defect density, oxygen-vacancy concentration, microscopic mobility, or trap parameters.Source: A. Residual-Guided Calibration of the… · p. 4
B. Fixed-Model Evaluation Across Bias and Geometry
The fixed calibrated 2%-W model transfers across drain bias and channel length without retuning, with the reported maximum errors remaining limited across both tests.
B. Fixed-Model Evaluation Across Bias and Geometry
- The 2%-W model uses the same TCAD physics and parameter set across VD = {0.05, 0.10, 0.20, 0.50, 1.00} V and L = {1, 2} µm, supporting transferability beyond nominal calibration conditions.The splits were defined before calibration, so this is not another local parameter fit.Source: B. Fixed-Model Evaluation Across Bias… · p. 4, Figure 6 · p. 5
- Fig. 5’s residual heatmap normalizes each error component to its Step 0 value, with Step 0 = 1, to visualize reductions across calibration steps.The figure also links residual components to suggested TCAD actions and model classes.Source: Figure 5 · p. 5
- Maximum errors were 36.1 mV in Vth and 0.022 decade in Ion for drain-bias tests, and 46.2 mV in Vth and 0.062 decade in Ion for channel-length tests.The tests use the same fixed calibrated model without retuning.Source: B. Fixed-Model Evaluation Across Bias… · p. 4, Figure 6 · p. 5
IV. PROCESS-SENSITIVE INSIGHT
The separately recalibrated 4%-W device has a positive Vth shift and lower current than the 2%-W reference. The agent maps these changes to model descriptors and uses the resulting calibration sequence to identify process-sensitive insight.
IV. PROCESS-SENSITIVE INSIGHT
- The 4%-W device shows a positive Vth shift and lower current than the 2%-W reference, represented in TCAD by reduced ND,eff rather than microscopic oxygen-vacancy extraction.The 4%-W fit and metrics use the measured range VG ≥ −0.7 V.Source: §IV · p. 5, Figure 7 · p. 6
- Step 1 reduced ND,eff from 2.56 × 10^18 to 6.39 × 10^16 cm−3; Step 2 changed gate WF from 4.00 to 4.45 eV and Qf from +1.59 × 10^10 to −1.00 × 10^12 cm−2.These updates address the dominant threshold and current shifts, then the residual threshold offset.Source: §IV · p. 5
- Step 3 reduced the remaining SS residual by broadening σtail from 0.133 to 0.238 eV and changing the donor-like Gaussian DOS center from EC−0.122 to EC−0.080 eV.The Gaussian DOS increased from 2.19 × 10^19 to 4.40 × 10^19 cm−3eV−1, and its width increased from σ = 0.080 to 0.224 eV.Source: §IV · p. 5
- The sequential TCAD curves move toward the 4%-W data, identifying model and parameter classes that reduce distinct residual components and can guide later process experiments and TCAD studies.The descriptor sequence provides a model-level inverse diagnostic of W-composition sensitivity.Source: §IV · p. 5, Figure 7 · p. 6
V. CONCLUSION
The workflow calibrates fabricated BG-IWO transistors from measured ID–VG data, with an LLM agent orchestrating the process and Sentaurus governing device physics. The 2%-W model’s objective falls by 14.3×, and the fixed calibrated model passes drain-bias and channel-length transferability tests.
V. CONCLUSION
- For the 2%-W reference model, the workflow reduces the device objective J by 14.3×, and the fixed calibrated model passes drain-bias and channel-length transferability tests.The 4%-W test provides process-sensitive insight by ranking TCAD model classes that reduce distinct residual components associated with the composition change.Source: §V · p. 6
- The LLM agent orchestrates the workflow, while reviewed model classes, live Sentaurus execution, and transferability tests establish validity.Sentaurus performs the device physics calculations.Source: §V · p. 6
- The methodology is extendable to other emerging device technology pathfinding to support a faster iterative loop between model generation and process optimization.Source: §V · p. 6