LLM-Assisted Analog Amplifier Sizing

An LLM-driven analog amplifier sizing system using gm/ID methodology, ngspice SPICE simulation, and self iteration on the SKY130 open-source PDK.

Why this project?

LLMs can reason about circuits, but they cannot size them reliably. They lack of domain knowledge and PDK information, and have no way to verify whether a design actually works. Meanwhile, traditional analog sizing tools require deep expert knowledge to set up and iterate.

This project bridges the gap. It gives the LLM agent everything a human designer uses — gm/ID look-up tables, analytical equations, SPICE simulation, and structured diagnosis — so the agent can size circuits the way an expert would, but faster and without manual intervention.

What makes it work:

1. gm/ID look-up tables map operating-point targets directly to transistor dimensions — no manual BSIM4 model evaluation, no guessing W and L
2. Procedural skill stack encodes the complete design methodology as markdown files that any LLM agent can read and execute step by step
3. SPICE-in-the-loop verifies every sizing iteration against ngspice — the agent never reports a result it hasn’t simulated
4. Root-cause fault trees link each spec failure to its root cause and recommended fix, with side-effect warnings and priority ordering
5. Numerical optimizer (CMA-ES / coordinate descent) fine-tunes the final sizing after the LLM converges

In a controlled experiment on a two-stage Miller OTA with 12 specs, the skill-equipped agent met all targets in 2 iterations. A bare LLM with the same model and tools but no skills failed to converge after 8 iterations, missing 2 specs.

Architecture

User (specs + netlist)
        |
        v
+-------------------+     HTTP (port 8001)     +--------------------+
|   AnalogAgent     |   |  CircuitCollector  |
|                   |                          |                    |
|  - LLM reasoning  |    /simulate/            | - Testbench gen    |
|  - gm/ID LUT      |    /register_circuit/    | - ngspice runner   |
|  - Skill stack    |    /health               | - Result parser    |
|  - Optimizer      |                          | - SKY130 PDK       |
+-------------------+                          +--------------------+

Supported topologies

Custom netlists are also supported. Drop a .sp SPICE subcircuit and the agent will parse the topology, register it, and size it.

Using the sizing flow

Fill in spec-form-template.md with your targets, then prompt the agent:

Use the skills to size the 5tota_single; use the specs in the template.

The agent reads the skill stack in skills/analog-amplifier/, identifies the topology, and runs the full flow:

1. Spec validation
2. Topology identification from the netlist
3. gm/ID-based initial sizing with LUT queries
4. Analytical pre-screening (poles, zeros, PM, gain, all specs)
5. SPICE simulation via CircuitCollector
6. Root-cause diagnosis and iterative fix (up to 10 iterations)
7. Design review with analytical vs SPICE comparison
8. Optional extreme PVT check (SS/85C + FF/-40C)
9. Optional numerical optimization (CMA-ES)

Agent compatibility

All entry files point to the same skill stack. skills/analog-amplifier/SKILL.md is the single source of truth.

Repository structure

AnalogAgent/
  skills/analog-amplifier/            Procedural skill stack (design flow, equations, fault trees)
  tools/                              Simulation bridge, optimizer, topology manager
  scripts/                            gm/ID LUT query API
  asset_new/                          Pre-computed LUT data (nfet/pfet, 5 corners, 3 temps)
  examples/                           Reference netlists (5T OTA + TSM, 3 variants each)
  agent/                              Programmatic agent entry point (Anthropic API)
  spec-form-template.md               User-facing spec form
  setup_and_run.ipynb                 Step-by-step setup and run notebook
  comparison_report_tsm_single.md     AnalogAgent vs pure-LLM controlled experiment
  README.md                           AnalogAgent documentation

CircuitCollector/
  CircuitCollector/                   Main package (API, runner, cache, templates, PDK)
  setup.py                            Package installation
  README.md                           CircuitCollector documentation

See AnalogAgent/README.md and CircuitCollector/README.md for detailed documentation of each component.

Tools and versions

Contributors

References

• P. G. A. Jespers and B. Murmann, Systematic Design of Analog CMOS Circuits: Using Pre-Computed Lookup Tables. Cambridge University Press, 2017.
• Razavi, B. 2000a. Design of Analog CMOS Integrated Circuits. McGraw-Hill, Inc.
• SkyWater SKY130 Open-Source PDK: https://github.com/google/skywater-pdk
• ngspice Open-Source SPICE Simulator: https://ngspice.sourceforge.io/

License

MIT