Generative AI for Automated Analog Circuit Design

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 288 papers96 conceptsfrozen 2026-10-04

For decades, the semiconductor industry has operated at two different speeds. Digital design is highly automated, moving from high-level code (Verilog) to physical silicon via standardized synthesis tools. Analog design, however, remains a "black art." It is a manual, iterative process where expert engineers spend weeks tweaking transistor dimensions and wiring to balance conflicting goals like gain, power, and noise.

As modern computing hits an "efficiency wall," analog computation is seeing a resurgence for its ability to perform low-power operations directly in the physics of silicon Yang Xue. To scale this resurgence, researchers are now using Large Language Models (LLMs) and Graph Neural Networks (GNNs) to bridge the automation gap.

The Abstraction Gap

The fundamental challenge in automating analog design is the lack of high-level abstraction. While digital logic can be simplified into discrete gates (AND, OR, NOT), analog circuits operate in a continuous, non-linear world where every physical connection introduces parasitic effects that can break the design.

To solve this, recent frameworks like AnalogCoder by Yao Lai and colleagues shift the focus from raw netlists to code generation. By prompting LLMs to generate Python scripts that interface with circuit simulators, the system creates a modular "circuit tool library." This allows the AI to work with functional subcircuits rather than individual transistors, mimicking the hierarchical way human experts think.

Topology Discovery as Sequence Generation

Finding the right "topology"—the arrangement of components—is a creative search problem. Traditional methods like Reinforcement Learning (RL) often require thousands of slow simulations to find a valid structure.

Newer approaches treat circuit topology as a language. LaMAGIC, developed by Chen-Chia Chang, fine-tunes LLMs to map performance specifications directly to circuit sequences. Similarly, AnalogGenie by Jian Gao and Weidong Cao represents circuits as Eulerian sequences. By solving the "Chinese Postman Problem" to minimize the length of these sequences, they reduce the complexity of the data the AI must learn, allowing it to discover novel, high-performance topologies that human experts might overlook.

Sizing and the Simulation Feedback Loop

Once a topology is chosen, the AI must "size" the components—determining the exact width and length of every transistor. This is a massive, non-linear search space.

The most successful systems today do not rely on "one-shot" generation. Instead, they use a closed-loop "diagnosis-and-repair" strategy. AnalogCoder-Pro and AnalogSAGE incorporate multimodal feedback, where the AI "sees" simulation waveforms and adjusts component values based on the error. By combining LLM reasoning with Bayesian Optimization, frameworks like LLANA achieve higher sample efficiency, finding optimal designs even when simulation data is scarce.

Toward Layout-Aware Synthesis

The final hurdle is the physical layout. In analog design, the geometry of the wires is just as important as the components themselves. A wire placed too close to another can introduce noise that ruins the circuit's performance.

Projects like OSIRIS by Giuseppe Chiari are building massive datasets to train models on "parasitic-aware" layout optimization. Meanwhile, LayoutCopilot by Yibo Lin and others allows designers to use natural language to guide these complex physical constraints. By translating intent ("reduce parasitics on this node") into executable EDA commands, these agents act as a bridge between human intuition and automated precision.

Through these combined efforts—shifting to code-based abstractions, treating topology as a sequence, and grounding AI reasoning in simulation feedback—the "black art" of analog design is finally becoming a scalable science.

Papers in this version

Showing 60 of 88 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 50 papers
  2. Version 1 · 2026-10-04 · 88 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_analog_design_automation_gap_d83815e4,
  title = {analog design automation gap},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/analog-design-automation-gap/rdr_d83815e438f5417f982089c66b840ccf},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_generative_ai_for_automated_analog_circu_d83815e4_v1,
  title = {Generative AI for Automated Analog Circuit Design},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/generative-ai-for-automated-analog-circuit-design/rdr_d83815e438f5417f982089c66b840ccf/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}