Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Guangxi Fan, Tianliang Ma, Xuguang Sun, Xun Wang, Kain Lu Low, Leilai Shao

TL;DR

AgenticTCAD addresses the scarcity of open TCAD resources by building an expert-curated open TCAD dataset and fine-tuning a domain-specific language model to generate complete SDE/SDevice code from natural language. The authors then deploy a multi-agent NL-driven workflow that couples code generation with a reasoning agent to perform end-to-end NS-FET design, TCAD simulation, and iterative optimization, guided by IRDS-2024 metrics. On a 2 nm NS-FET, the framework achieves IRDS-2024 specifications in 4.2 hours, dramatically reducing the human effort from days to hours, while also providing energy-band diagrams to enhance interpretability. This work offers a scalable, data-driven pathway for automated DTCO device design and sets the stage for broader AI-assisted TCAD research and DTCO workflows.

Abstract

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 AgenticTCAD, 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.

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

TL;DR

AgenticTCAD addresses the scarcity of open TCAD resources by building an expert-curated open TCAD dataset and fine-tuning a domain-specific language model to generate complete SDE/SDevice code from natural language. The authors then deploy a multi-agent NL-driven workflow that couples code generation with a reasoning agent to perform end-to-end NS-FET design, TCAD simulation, and iterative optimization, guided by IRDS-2024 metrics. On a 2 nm NS-FET, the framework achieves IRDS-2024 specifications in 4.2 hours, dramatically reducing the human effort from days to hours, while also providing energy-band diagrams to enhance interpretability. This work offers a scalable, data-driven pathway for automated DTCO device design and sets the stage for broader AI-assisted TCAD research and DTCO workflows.

Abstract

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 AgenticTCAD, 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.
Paper Structure (14 sections, 7 figures, 2 tables)

This paper contains 14 sections, 7 figures, 2 tables.

Table of Contents

  1. Introduction
  2. preliminaries and related works
  3. TCAD Simulation
  4. EDA Code Generation with LLM
  5. Agent Flow in EDA
  6. methodology
  7. Dataset Construction
  8. Domain-Specific LLM Fine-Tuning
  9. Multi-Agent Workflow for Automated Device Optimization
  10. Experiments and results
  11. Experimental Setup
  12. Open TCAD Dataset and Model Evaluation
  13. Multi-Agent Workflow Evaluation
  14. conclusion

Figures (7)

  • Figure 1: Overall flow of TCADAgent. (a) Dataset construction flow; (b) Supervised fine-tuning (SFT) flow; (c) Multi-Agent flow for End-to-end device TCAD simulation and optimization based on natural language.
  • Figure 2: Prompt templates and examples used in our framework. (a) Data Construction, (b) Code Generation and (c) Device Optimization.
  • Figure 3: SFT process: (a) training and evaluation loss, and (b) gradient norm.
  • Figure 4: Case study results with our TCAD code generation LLM with tailored meshing strategies. (a) 2D n-type MOSFET with Gaussian doping profile and (d) the corresponding $C-V$ simulation. (b) P-type FinFET array and (e) the corresponding $I_d-V_d$ simulation. (c) InGaAs-based n-type FinFET with high-k/metal gate stack and (f) the corresponding $I_d-V_g$ simulation.
  • Figure 5: TCADAgent for NS-FET design and optimization: (a) device structure and cross-section, (b) $I_d-V_g$ before (blue) and after (red) optimization.
  • ...and 2 more figures