🔬 applied physics

Agentic TCAD Calibration Workflow for Oxide Semiconductor Transistors

This paper introduces an agentic TCAD calibration workflow that leverages an LLM to orchestrate iterative model refinement and physical model selection for oxide semiconductor transistors, successfully reducing multi-metric device errors and demonstrating superior model transferability across varying biases and geometries compared to traditional expert-dependent methods.

Original authors: 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

Published 2026-09-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: 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

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of modern electronics, the silicon chip has long been the undisputed king, but engineers are increasingly looking to new materials to build the next generation of devices. Among these contenders are transistors made from oxide semiconductors, a class of materials that can be processed at lower temperatures and integrated into the complex layers of advanced computer chips. These devices are crucial for everything from the memory in our phones to the power converters in electric vehicles. However, predicting how these new materials will behave before they are built is a notoriously difficult task. Scientists use sophisticated computer simulations to model these devices, but matching the simulation to the real world is like trying to tune a radio in a storm: there are many different settings that might produce a clear signal, but only one set of settings reflects the actual physics of the material. Without a precise match, engineers cannot reliably design the future circuits that will power our technology.

A team of researchers has now introduced a new way to solve this tuning problem, moving away from slow, manual guesswork toward an automated, intelligent process. In a recent study, they demonstrated a workflow where an artificial intelligence agent acts as a guide, helping to calibrate a computer model of a specific type of transistor made from indium, tungsten, and oxygen. The goal was to take the raw electrical measurements from a physical device and find the exact set of rules and numbers inside the simulation that would reproduce those measurements perfectly. The researchers found that by letting the AI agent analyze the differences between the real device and the simulation, it could systematically suggest changes to the model. After just five rounds of these intelligent suggestions, the simulation became a highly accurate representation of the real device, reducing the error between the two by more than fourteen times.

The process began with a physical transistor that had already been built and measured in a laboratory. The researchers fed the electrical data from this device into a computer program designed to simulate how the transistor should work. Initially, the simulation did not match the real data well; the curves representing the flow of electricity were in the wrong places. Instead of a human expert spending hours manually adjusting dozens of variables, an AI agent took over. This agent acted as a project manager, constantly comparing the simulation results with the real measurements. When it spotted a mismatch, it did not just guess a number to fix it. Instead, it consulted a vast library of technical manuals and scientific literature to understand what physical phenomenon might be causing the error.

If the mismatch suggested that the device was not turning on at the right voltage, the agent might suggest adjusting the work function, a property that describes how easily electrons can leave a material. If the mismatch appeared in the way the current leaked when the device was supposed to be off, the agent might propose adding a new physical model to account for trapped charges or specific energy states within the material. Crucially, the agent only accepted changes that actually improved the match. If a suggested adjustment made the simulation worse or failed to converge, the agent discarded it and tried a different path. This cycle of measuring, comparing, consulting, and adjusting continued until the simulation and the real-world data aligned.

The results of this automated calibration were striking. The final model, which was built using data from a device with a specific chemical composition, was able to predict the behavior of that same device under a wide variety of conditions without needing to be re-tuned. When the researchers tested the model with different electrical voltages and even with a transistor that was twice as long as the original, the simulation remained accurate. The error in predicting the voltage at which the device turns on was less than 46.2 mV, and the error in predicting the current flow was less than 0.062 decade. This proved that the model had learned the underlying physics of the device rather than just memorizing a single set of numbers.

The study also revealed how the model could be used to understand the manufacturing process itself. The researchers tested a second device made with a slightly different chemical mix, containing more tungsten. The AI agent successfully recalibrated the model for this new device, identifying that the change in chemistry primarily affected the density of charge carriers within the material. This provided a clear, quantitative link between the manufacturing recipe and the electrical performance, offering insights that could help engineers optimize the production process in the future.

This work demonstrates that artificial intelligence can serve as a powerful partner in the development of new electronic materials, turning a process that once required deep, specialized intuition into a more systematic and faster workflow. By combining the physical rigor of traditional simulation with the adaptive reasoning of an AI agent, the researchers have created a path toward more reliable and efficient design of the next generation of semiconductor devices. The approach does not replace human expertise but rather amplifies it, allowing engineers to focus on the big picture while the agent handles the complex, iterative task of finding the perfect match between theory and reality.

Technical Summary: Agentic TCAD Calibration Workflow for Oxide Semiconductor Transistors

Problem Statement
Experimental Technology Computer-Aided Design (TCAD) calibration is a critical bottleneck in predictive technology modeling for emerging oxide semiconductor transistors. The process is traditionally time-consuming and heavily dependent on expert intuition because it constitutes a nontrivial inverse problem. A primary challenge is "model ambiguity": multiple combinations of physical models and parameter sets can reproduce the same measured transfer characteristics, yet a local parameter fit alone fails to uniquely identify the underlying device physics or guarantee model generalization. Existing AI/ML-assisted TCAD studies typically operate within predefined physics or design spaces, lacking the capability to autonomously select physical models or provide insights based on experimental data residuals.

Methodology
The authors present an agentic TCAD calibration workflow designed to automate the selection of physics models and parameter adjustments for a fabricated bottom-gate In–W–O (BG-IWO) transistor. The workflow integrates four key components orchestrated by a Large Language Model (LLM) agent:

Retrieval-Augmented Generation (RAG): The agent queries a knowledge base (KB) containing the Sentaurus Device User Guide and reviewed device literature to identify candidate physical models consistent with observed measurement–TCAD residuals.
Residual-Guided Loop: The system compares measured transfer curves with current TCAD results to extract residuals and device metrics. It determines whether the next action requires adjusting an existing parameter or activating an additional physics model.
Bounded Optimization:
- Parameter Tuning: For existing parameters, the workflow uses One-Factor-At-a-Time (OFAT) probes to estimate local sensitivity, followed by Derivative-Free Gauss–Newton (DFO-GN) updates to calculate bounded parameter changes that minimize a multi-metric objective function ( $J$ ).
- Model Activation: If systematic mismatches persist after parameter bounds are exhausted, the agent proposes new physical models (e.g., Gaussian or tail density-of-states models) based on RAG evidence.
Human-in-the-Loop Validation: A human reviewer defines the workflow constraints (Agent Skills), approves deck modifications, and verifies the physical validity of proposed updates. The LLM acts as the orchestrator, while Sentaurus TCAD solves the device equations.

Updates are accepted only if the Sentaurus run converges and reduces the multi-metric objective $J$ , which aggregates errors in log/linear current, threshold voltage ( $V_{th}$ ), on-current ( $I_{on}$ ), and subthreshold swing ( $SS$ ).

Key Contributions

First Experimental Demonstration: This work presents the first experimental demonstration of an agentic TCAD calibration workflow applied to a fabricated oxide semiconductor device.
Active Physics Model Selection: Unlike prior works that fit parameters within a fixed model, this workflow actively selects and activates physical models (e.g., DOS terms) based on residual analysis and literature retrieval.
Process-Sensitive Insight: The workflow successfully differentiates between calibration-effective descriptors (e.g., effective work function, effective donor-like background density) and microscopic physical quantities, providing a model-level inverse diagnostic for process variations.

Results
The workflow was applied to a 2%-W reference BG-IWO device ( $L=1 \mu m$ ) calibrated against transfer curves at drain biases of 50 mV and 1 V.

Calibration Efficiency: Five agent-suggested updates reduced the multi-metric device objective $J$ by 14.3× (from 2.589 to 0.181).
Model Transferability: The fixed calibrated model was tested without retuning on varying drain biases (0.05–1.0 V) and varying channel lengths ( $L=2 \mu m$ $L = 2 μ m$ ).
- Drain-bias tests: Maximum errors were 36.1 mV for $V_{th}$ and 0.022 decade for $I_{on}$ .
- Channel-length tests: Maximum errors were 46.2 mV for $V_{th}$ and 0.062 decade for $I_{on}$ .
Process Sensitivity: A separate recalibration for a 4%-W device revealed that the composition change primarily affected the effective donor-like background density ( $N_{D,eff}$ ) and gate electrostatics, consistent with prior studies on dopant–oxygen bonding.

Significance and Claims
The paper claims that this agentic workflow provides a faster route to model development for emerging device technologies by reducing the active search effort and model/parameter ambiguity. The authors emphasize that the workflow produces a fixed calibrated model capable of generalizing across bias and geometry, rather than merely achieving a local parameter fit. By integrating RAG, live TCAD execution, and residual-guided decision-making, the framework offers physical and process insights that can guide follow-up TCAD studies and process optimization. The authors note that the methodology is extendable to other emerging device technology pathfinding, enabling a fast iterative loop between model generation and process optimization.

Technical Summary: Agentic TCAD Calibration Workflow for Oxide Semiconductor Transistors

More like this