Skip to content

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

🕒 Published (v1): 2025-12-26 01:34 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

AgenticTCAD is a multi-agent LLM framework that automates end-to-end TCAD (Technology Computer-Aided Design) device simulation and optimization via natural language. It combines a fine-tuned domain-specific code generation model (Qwen2.5-14B) with a commercial reasoning LLM (DeepSeek V3.1) in a three-stage loop covering code generation, simulation execution, and iterative parameter optimization. Applied to a 2 nm nanosheet FET, it meets IRDS-2024 specs in 4.2 hours vs. 7.1 days for human experts.

Problem

TCAD simulation requires highly domain-specific scripting languages (Synopsys SDE/SDevice) for which no public datasets existed, making LLM-based code generation unreliable. Prior work either omitted mesh construction (making electrical simulation infeasible), required Python-intermediate translations, or relied on case-specific templates. No unified framework covered both structure generation (SDE) and electrical simulation (SDevice) with closed-loop device parameter optimization.

Method

Three integrated components:

  1. Dataset construction: TCAD experts handcrafted baseline SDE/SDevice scripts covering planar MOSFETs, FinFETs, nanowires, nanosheets in 2D/3D, silicon and III-V materials. A commercial LLM generated paired natural-language queries from annotated, JSON-packaged scripts; experts manually verified the augmented data. Final corpus: ~30,000 training samples.

  2. Domain-specific SFT: Qwen2.5-14B-Instruct fine-tuned for 5 epochs (batch size 8, bfloat16, AdamW with cosine LR decay, ~107 GPU-hours on 4×A6000) to directly map NL queries → SDE or SDevice scripts, with prompt tokens masked during loss computation so the model learns output structure, not copying.

  3. Multi-agent optimization loop (3 stages):

  4. Stage 1: Fine-tuned model converts JSON design parameters → SDE (geometry/doping/meshing) + SDevice (physical models/boundary conditions) scripts.
  5. Stage 2: Sentaurus executes scripts; post-processing extracts Ion, Ioff, SS, I–V curves, and energy band diagrams into structured JSON.
  6. Stage 3: DeepSeek V3.1 reasoning model ingests simulation results + quantitative IRDS targets + previous parameters → proposes updated parameters; checks convergence or max-iteration; a recovery mechanism handles non-convergent parameter sets by providing the last convergent design alongside the failed config.

Key Contributions

  • First open-source expert-crafted TCAD dataset (SDE + SDevice scripts, LLM-augmented to ~30k samples) covering diverse device architectures and material systems.
  • Fine-tuned Qwen2.5-14B that generates complete, simulation-ready SDE and SDevice code directly from natural language without intermediate Python translation.
  • AgenticTCAD multi-agent framework: first end-to-end NL-driven TCAD flow integrating code generation, simulation execution, physical post-processing, and closed-loop LLM-based optimization.
  • Automatic extraction of energy band diagrams (EC, EV, EFn, EFp) during optimization to provide physically interpretable insight.
  • Quantitative vs. qualitative prompting ablation demonstrating that explicit numerical targets are necessary for stable convergence.

Results

  • NS-FET optimization (2 nm, IRDS-2024 high-performance specs, Vdd = 0.65 V):
  • SS: 286.72 → 60.38 mV/dec (spec: ≤72 mV/dec) ✓
  • Ioff: 3.40×10⁻⁵ → 8.26×10⁻⁹ A/µm (spec: ≤1.0×10⁻⁸ A/µm) ✓
  • Ion: 4.23×10⁻³ → 2.31×10⁻³ A/µm (spec: ≥7.87×10⁻⁴ A/µm; 2.94× above spec) ✓
  • ON-OFF ratio: 2.10 → 5.45 (spec: ≥4.90) ✓
  • Efficiency: 25 optimization iterations, 4.2 hours total vs. 5,120 simulations / 7.1 days by human expert with five commercial licenses.
  • Qualitative vs. quantitative guidance: qualitative prompting failed to converge Ioff (still ~1 order above spec at best); quantitative guidance consistently converged all metrics.
  • Code generation breadth: demonstrated on 2D n-MOSFET, p-FinFET array, InGaAs n-FinFET with high-k/metal gate (Id–Vg, Id–Vd, C–V all valid).

Limitations

  • Validated on a single device type (NS-FET) at one technology node (2 nm); generalization to other architectures or process flows is unverified.
  • Depends on commercial Synopsys Sentaurus for simulation execution; not fully open-source end-to-end.
  • SFT dataset constructed semi-automatically; physical plausibility requires manual expert verification, limiting scalability.
  • Recovery mechanism for non-convergent parameters is heuristic; persistent non-convergence could still stall the loop.
  • Optimization uses DeepSeek V3.1 as a black-box reasoner; no formal guarantee of convergence or optimality over the nonlinear device physics space.
  • 107 GPU-hours of fine-tuning required; barrier for groups without A100/A6000-class hardware.

Relevance to Harnesses / Meta-Harnesses

AgenticTCAD is a textbook domain-specific meta-harness: it wraps a specialized fine-tuned model, a commercial reasoning LLM, a simulation engine, and a post-processing pipeline into a single closed-loop orchestration layer driven by natural language. The three-stage loop (generate → execute → optimize) directly parallels general agentic harness patterns — tool-calling, structured JSON handoffs between stages, error recovery sub-routines, and quantitative stopping criteria — instantiated in a high-stakes scientific domain. For harness researchers, the key finding is that prompt engineering (quantitative vs. qualitative guidance) is the dominant variable controlling loop convergence, reinforcing that the meta-harness control layer — not the underlying models — is the critical design surface. The non-convergence recovery mechanism is also a concrete example of harness-level fault tolerance that generalizes beyond TCAD.