Skip to content

COBRA: An AI-Assisted Circuit-Level Optimizer for Open Source Based RFIC Design

COBRA logo

COBRA (Circuit-Level Open-Source Based RFIC AI-Assisted Optimizer) accelerates RFIC circuit optimization by combining surrogate model inference, SPICE simulation, and goal-driven optimization in one workflow.

Main Capabilities

  • Mixed component sources in one run: ONNX surrogates and fixed SNP files.
  • Goal-based optimization with Optuna samplers/pruners.
  • Design metrics for S-parameters and RF lumped quantities (e.g. L, R, Q, coupling).
  • Optional EM fine-tuning loop through Palace + ORCA geometry.
  • GUI workflow and script-first workflow.

Start Here

  • New users: go to Getting Started -> Installation.
  • First successful run: Getting Started -> Quickstart.
  • API details: API Reference.

Core Concepts

COBRA optimizes circuit behavior using a staged loop that combines surrogate prediction and SPICE simulation.

Optimization Workflow

flowchart LR
        A[Netlist .cir/.sp] --> B[Map components to ONNX or SNP]
        B --> C[Optimizer proposes parameters]
        C --> D[Surrogate stage predicts S-params]
        D --> E[Vector fit + Xyce circuit simulation]
        E --> F[Design-goal penalties]
        F --> C
        F --> G[Best result in results/timestamp_name]

Goals and Parameter Types

  • A design goal is a bound on one metric (for example S11, S21, L, Q, k), optionally in a frequency range such as 125-135ghz.
  • MODEL_INPUT parameters control ONNX surrogate inputs (example: X1:bottom_winding_diameter).
  • NETLIST_VARIABLE parameters patch parsed netlist values directly (example: Cshunt_p).
  • Linked parameters (linked_to) enforce symmetry/constraints by mirroring values.

Penalties are aggregated as:

\[L = \sum_i w_i \cdot p_i\]

where \(p_i\) is each goal penalty and \(w_i\) is the goal weight.

Inputs and Outputs

- Circuit netlist from Qucs-S/Xyce format.
- Component models (`.onnx` and/or `.sNp`).
- Design goals and optimization parameter definitions.

- Timestamped folder in `results/`.
- Predicted and surrogate S-parameter files.
- Vector-fitted SPICE subcircuits.
- Full optimization context JSON.
  • ORCA creates surrogate models that COBRA consumes.
  • Palace can be used for optional full-wave EM fine-tuning.

If you use COBRA in research, see citation information in the repository README.