Yao Lai1, Souradip Poddar2, Sungyoung Lee2, Guojin Chen3, Mengkang Hu1, Bei Yu3, Ping Luo1, David Z. Pan2.
1 The University of Hong Kong, 2 The University of Texas at Austin, 3 The Chinese University of Hong Kong.
[Paper]
This work is an extension of AnalogCoder (AAAI 2025) [repo].
- Our work appears in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems (TCAD), vol. 45, no. 11, pp. 5149–5162, November 2026. [Paper]
- Challenge: Analog front-end design still relies heavily on expert intuition and iterative simulations, with limited automation.
- Solution: AnalogCoder-Pro — a unified multimodal LLM-based framework for analog design automation.
- Key Features:
- Joint circuit topology generation and device sizing optimization
- Automatic generation of performance-specific schematic netlists
- Multimodal diagnosis & repair using specifications and waveform images
- Automated extraction of design parameters and parameter space formulation
- Outcome: Improves design success rate and circuit performance, enabling an end-to-end automated workflow.
- Update the LLM run scripts.
- Update the sample waveform figures.
- Update the BO optimization.
- Update all ablation study prompts.
- Task descriptions are in the file
problem_set.tsv. - Sample circuits are in the
sample_designdirectory. - Test benches are in the
problem_checkdirectory.
git clone [email protected]:laiyao1/AnalogCoderPro.git
cd AnalogCoderPro
conda create -n analog python==3.10
conda activate analog
pip install -r requirements-sizing.txt
conda install -c conda-forge ngspice -y
conda install -c conda-forge pyspice
python run.py --task_id=19 --num_per_task=3 --model=gpt-5-mini --api_key="[API_KEY]" --base_url="[BASE_URL]"
This script will attempt Mixer generation 3 times.
The mapping of task IDs can be found in problem_set.tsv.
Automatic optimizer: set --task_id to 51–66 (inclusive) in the current
benchmark. run.py generates a circuit topology, checks its operating point
and function, then parameterizes the successful circuit and runs the Optuna
optimizer. Sizing starts only after the generated circuit passes these checks.
Task IDs at or below 50 do not enable automatic sizing. There is no --optimize
CLI option.
--task_id |
Automatic optimization objective |
|---|---|
| 51–54 | Voltage gain |
| 55–58 | Gain-bandwidth product (GBW) |
| 59–62 | Figure of merit (FoM) |
| 63–66 | FoM (the default objective for these tasks) |
For example, after configuring OPENAI_API_KEY (and optionally
OPENAI_BASE_URL), generate and size one circuit for task 59:
python run.py --task_id=59 --num_per_task=1 --model=gpt-5-mini --no_vlm \
--sizing-trials=10 --sizing-jobs=1 --sizing-runs=1Use --sizing-trials, --sizing-jobs, and --sizing-runs to control the trials
per run, parallel jobs, and independent optimization runs. The command above
makes model API calls and uses a small sizing budget for an initial check.
To size an existing parameterized circuit directly, use the included amplifier or op-amp examples with the standalone optimizer:
python optimize/circuit_optimizer_helper.py \
optimize/examples/amplifier_optimize.py fom 1 \
--n-trials 10 --n-jobs 1 --output-dir results/amplifierSee the sizing guide for installation, objective definitions, module requirements, output descriptions, and integration test instructions.
This case study compares GPT-5-mini circuits generated for task 59 (with an explicit FoM objective) and task 63 (without that objective). Each fixed topology was sized with the same 10 optimizer seeds (42–51), 1000 trials per seed, and a 100 pF load. The main plot zooms into late-stage convergence; the inset retains the full scale.
Final mean FoM: 1876.371 with the spec-driven prompt and 1876.144 without. The mean paired difference is +0.0121% (95% descriptive interval: −0.0465% to +0.0708%). This single topology pair does not establish a stable benefit across independent model generations. Figure data and plotting script are included for reproducibility.
Here are example waveforms for different circuit types, demonstrating the appropriate analysis methods for each design.
| Mixer — Transient + FFT Spectrum | Schmitt Trigger — Transient + DC Transfer |
![]() |
![]() |
| Oscillator — Transient | Integrator — Transient |
![]() |
![]() |
| Differentiator — Transient | BandStop Filter — AC |
![]() |
![]() |
| Comparator — DC Sweep | |
![]() |
If you find our work useful, we would appreciate a citation of our paper.
@article{lai2025analogcoder,
title={AnalogCoder: Analog Circuit Design via Training-Free Code Generation},
volume={39},
DOI={10.1609/aaai.v39i1.32016},
number={1},
journal={Proceedings of the AAAI Conference on Artificial Intelligence},
author={Lai, Yao and Lee, Sungyoung and Chen, Guojin and Poddar, Souradip and Hu, Mengkang and Pan, David Z. and Luo, Ping},
year={2025},
pages={379-387}
}
@article{lai2026analogcoderpro,
author={Lai, Yao and Poddar, Souradip and Lee, Sungyoung and Chen, Guojin and Hu, Mengkang and Yu, Bei and Luo, Ping and Pan, David Z.},
journal={IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems},
title={AnalogCoder-Pro: Unifying Analog Circuit Generation and Optimization via Multimodal LLMs},
year={2026},
volume={45},
number={11},
pages={5149--5162},
month=nov,
doi={10.1109/TCAD.2026.3673493}
}







