Research map / Large language models / Agents, code and math
LaMAGIC: Language-Model-based Topology Generation for Analog Integrated Circuits
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Fine-tunes Flan-T5-base with supervised finetuning and novel circuit formulations (canonical, adjacency-matrix, float-input) to generate power converter topologies and duty cycles in one pass from voltage ratio and efficiency specs.
Problem
Automated analog circuit topology design for power converters currently relies on search-based methods that need hundreds of simulator queries per new specification; the task is to generate circuit topologies directly from specifications in a single pass.
Approach
Casts analog topology generation as sequence generation with SFT of an encoder-decoder LM, using canonical edge ordering, one-hot duty-cycle tokens, an adjacency-matrix circuit format, and a float-input embedding for numerical specs, plus random vertex-order permutation augmentation.
Key results
- Achieves a success rate of up to 0.96 under a strict tolerance of 0.01 for the proposed formulations.
- Float-input adjacency matrix (FM) gives the lowest topology-generation MSE (voltage 0.002, efficiency 0.001) and the best 6-component transfer results (voltage MSE 0.038 with 2000 samples).
- GPT-4 with few-shot prompting on 100 samples failed to produce circuits meeting the specifications; encoder-only T5 failed to converge.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the naive circuit format with a canonical formulation (ablation) | LaMAGIC, edge and topology generation · 3, 4, 5-component circuits | MSE (voltage conversion ratio): 0.054 -> 0.016 (edge); 0.031 -> 0.008 (topology) | Helped |
| use one-hot duty-cycle selection in the canonical formulation (ablation) | LaMAGIC, edge and topology generation · 3, 4, 5-component circuits | MSE (voltage conversion ratio): 0.016 -> 0.007 (edge); 0.008 -> 0.005 (topology) | Helped |
| use adjacency-matrix output with float inputs instead of text inputs (ablation) | LaMAGIC, edge generation · 3, 4, 5-component circuits | MSE (voltage conversion ratio): 0.007 -> 0.006 | Helped |
| use adjacency-matrix output with float inputs instead of text inputs (ablation) | LaMAGIC, edge generation · 3, 4, 5-component circuits | MSE (efficiency): 0.002 -> 0.013 | Hurt |
| use adjacency-matrix output with float inputs instead of text inputs (ablation) | LaMAGIC, topology generation · 3, 4, 5-component circuits | MSE (voltage conversion ratio): 0.009 -> 0.002 | Helped |
| use adjacency-matrix output with float inputs instead of text inputs (ablation) | LaMAGIC, topology generation · 3, 4, 5-component circuits | MSE (efficiency): 0.003 -> 0.001 | Helped |
| use adjacency-matrix output with float inputs instead of text inputs (ablation) | LaMAGIC, 6-component edge generation · 6-component circuits, 500, 1000, and 2000 samples | MSE (voltage conversion ratio): 0.092 -> 0.050; 0.096 -> 0.048; 0.051 -> 0.038 | Helped |
| add random vertex-order permutation during training (ablation) | PM, edge generation · 3, 4, 5-component circuits | MSE (voltage conversion ratio): 0.010 -> 0.007 | Helped |
| add random vertex-order permutation during training (ablation) | FM, edge generation · 3, 4, 5-component circuits | MSE (voltage conversion ratio): 0.040 -> 0.006 | Helped |
| use an encoder-only model instead of the encoder-decoder model (ablation) | LaMAGIC with PM · Circuit generation | valid circuit generation | Hurt |
| pretrain on simpler circuits before fine-tuning on 6-component circuits (ablation) | LaMAGIC · 6-component circuits | valid circuit generation | Helped |
| use the proposed circuit formulations instead of the naive formulation (ablation) | LaMAGIC, edge and topology generation · 3, 4, 5-component circuits | success rate | Helped |
Surprising findings
- FM performs worse than text-based formulations on edge generation at small tolerances, but better on topology generation and transfer to 6-component circuits.
- Training solely on 6-component circuits without pretraining on simpler circuits produced no valid circuits.
Limitations
- Limited to power converter circuits with a small set of components (3 to 6) and a fixed duty cycle set.
- Evaluated on a small model (Flan-T5-base, 248M parameters) and a simulated dataset; scalability to larger circuits is not established.
Experimental setup
Datasets: NGSPICE-simulated power converter dataset of 3-, 4-, 5-component circuits 6-component circuits (76k sampled, 9k evaluation)
Benchmarks: held-out power converter test set (13K requirements) 6-component transfer test set (9k circuits)
Metrics: success rate under tolerance 0.01 to 0.1 MSE of voltage conversion ratio and efficiency
Baselines: RL-based tree sampling power converter search GPT-4 few-shot prompting naive instruction formulation (NF) canonical formulation (CF) CFDC PM (pure text adjacency matrix) encoder-only T5
Builds on: Flan-T5 T5 masked language modeling NGSPICE
Closest papers by meaning
- LaMAGIC2: Advanced Circuit Formulations for Language Model-Based Analog Topology Generation (ICML 2025 · similarity 0.96)
- AnalogToBi: Device-Level Analog Circuit Topology Generation via Bipartite Graph and Grammar Guided Decoding (NeurIPS 2026 · similarity 0.91)
- AnalogGenie-Lite: Enhancing Scalability and Precision in Circuit Topology Discovery through Lightweight Graph Modeling (ICML 2025 · similarity 0.91)
- AnalogGenie: A Generative Engine for Automatic Discovery of Analog Circuit Topologies (ICLR 2025 · similarity 0.90)
- FALCON: An ML Framework for Fully Automated Layout-Constrained Analog Circuit Design (NeurIPS 2025 · similarity 0.89)
- AUTOCIRCUIT-RL: Reinforcement Learning-Driven LLM for Automated Circuit Topology Generation (ICML 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.