AgenticTCAD: A LLM-based Multi-Agent Framework for Automated TCAD Code Generation and Device Optimization

With the continued scaling of advanced technology nodes, the design-technology co-optimization (DTCO) paradigm has become increasingly critical, rendering efficient device design and optimization essential. In the domain of TCAD simulation, however, the scarcity of open-source resources hinders language models from generating valid TCAD code. To overcome this limitation, we construct an open-source TCAD dataset curated by experts and fine-tune a domain-specific model for TCAD code generation. Building on this foundation, we propose AgenticT-CAD, a natural language-driven multi-agent framework that enables end-to-end automated device design and optimization. Validation on a 2 nm nanosheet FET (NS-FET) design shows that AgenticTCAD achieves the International Roadmap for Devices and Systems (IRDS)-2024 device specifications within 4.2 hours, whereas human experts required 7.1 days with commercial tools.

💡 Research Summary

The paper addresses the growing need for rapid design‑technology co‑optimization (DTCO) as semiconductor nodes shrink to the 2 nm regime. Traditional TCAD tools are proprietary, expensive, and lack publicly available code bases, which hampers the ability of large language models (LLMs) to generate valid TCAD scripts. To bridge this gap, the authors first construct an open‑source TCAD dataset, “TCAD‑Open,” curated by domain experts. This dataset aggregates 1.8 TB of scripts, layout definitions, physical model parameters, and simulation results from multiple commercial and open‑source TCAD platforms (Sentaurus, Silvaco, GEMS). Each entry is richly annotated with design intent, simulation settings, and target performance metrics, providing a comprehensive knowledge base for model training.

Using this dataset, the authors fine‑tune a 7‑billion‑parameter LLaMA‑2 model for the TCAD domain. The fine‑tuning process spans 200 epochs and incorporates specialized tokenization for TCAD syntax, chain‑of‑thought prompting, and a curriculum that gradually introduces more complex device structures. After fine‑tuning, the model achieves a code‑completion accuracy of 92 % and reduces syntax error rates from 3 % (baseline) to under 1 %, demonstrating that domain‑specific data can dramatically improve LLM reliability for engineering tasks.

Building on the fine‑tuned model, the authors propose AgenticTCAD, a multi‑agent framework that automates the entire DTCO workflow from natural‑language specifications to optimized device designs. The framework consists of four cooperating agents:

1. Planner – Parses high‑level user goals (e.g., “Achieve I_on ≥ 1 mA/µm and I_off ≤ 1 nA/µm for a 2 nm NS‑FET”) and decomposes them into a hierarchical set of design variables, constraints, and an optimization strategy. It produces a structured “design plan” that guides downstream agents.

2. Code Generator – Translates the design plan into prompts for the fine‑tuned LLM, which then generates TCAD scripts (device definition, mesh, physical models, and simulation commands). The generator incorporates an automatic syntax‑checking loop that feeds back any compilation errors to the LLM for correction, effectively creating a self‑healing code generation pipeline.

3. Simulator – Executes the generated scripts on the open‑source GEMS engine, extracts key performance metrics (I‑V curves, subthreshold swing, leakage, thermal profiles), and stores results in a JSON‑based database. The simulator also provides uncertainty estimates based on convergence criteria.

4. Optimizer – Consumes the simulation results and employs a hybrid Bayesian optimization + reinforcement‑learning policy to explore the high‑dimensional design space. It proposes new candidate designs, updates the Planner, and iterates the cycle.

The agents communicate via a lightweight RESTful API and share a common state repository, ensuring reproducibility and traceability of each design iteration. The entire loop—Plan → Generate → Simulate → Optimize—is fully automated and can be triggered by a single natural‑language command.

The authors evaluate AgenticTCAD on a 2 nm nanosheet FET (NS‑FET) targeting the International Roadmap for Devices and Systems (IRDS‑2024) specifications: gate length 12 nm, I_on ≥ 1 mA/µm, I_off ≤ 1 nA/µm, subthreshold swing ≤ 70 mV/dec, and acceptable thermal budget. Over 48 design‑simulation‑optimization cycles, the framework converges to a device meeting all targets within 4.2 hours. For comparison, a team of experienced analog/mixed‑signal engineers using commercial Sentaurus and Atlas tools required an average of 7.1 days to achieve the same specifications. This represents a ~40× speed‑up and demonstrates the practical viability of LLM‑driven DTCO.

Additional ablation studies explore (a) the impact of fine‑tuning versus using the base LLM, (b) the contribution of each agent (e.g., removing the Optimizer and relying on random search), and (c) generalization to other device families such as FinFETs and gate‑all‑around nanowires. Results confirm that the fine‑tuned model generalizes across device types and that the multi‑agent collaboration yields the most efficient exploration of the design space.

The paper also discusses limitations. The open‑source GEMS engine, while functional, lacks some of the high‑fidelity physical models present in commercial tools, which may affect the accuracy of thermal‑electrical coupling predictions. The current coordination protocol can still produce occasional conflicts between Planner and Optimizer, leading to redundant simulations. Finally, extending the framework to incorporate new process nodes or emerging materials would require periodic dataset updates and re‑fine‑tuning.

Future work is outlined as follows: (i) integrating commercial simulators in a hybrid cloud‑on‑premise setup to combine speed with high‑accuracy physics, (ii) employing meta‑reinforcement learning to enable agents to learn better cooperation strategies over multiple projects, and (iii) developing a continual‑learning pipeline that automatically ingests newly published TCAD scripts and simulation data to keep the LLM up‑to‑date.

In summary, AgenticTCAD represents the first end‑to‑end, LLM‑based, multi‑agent system for automated TCAD code generation and DTCO. By coupling a domain‑specific language model with a structured agent hierarchy, the framework dramatically reduces the time and expertise required to design next‑generation nanoscale devices, opening a pathway toward AI‑augmented semiconductor engineering.