Skip to content

ORCA: Open RF Integrated Circuit Automation

ORCA logo

ORCA (Open RF Integrated Circuit Automation) builds surrogate models of RF integrated circuit components by combining parametric GDS layout generation, full-wave electromagnetic simulation, and neural network training and export in one automated pipeline.

Main Capabilities

  • Parametric GDS layout generation via gdsfactory.
  • Full-wave electromagnetic simulation via Palace.
  • Automatic dataset construction from simulation results.
  • Neural network training (PyTorch) with feature engineering pipelines.
  • ONNX export for portable, simulator-ready surrogate models.
  • GUI workflow and script-first workflow.
  • Pre-trained models and geometry examples included.

Start Here

Core Concepts

ORCA automates the full loop from geometry to a trained, exported surrogate model.

Pipeline Workflow

flowchart LR
        A[Geometry Class] --> B[GDSGenerator]
        B --> X[DRCChecker]
        X --> C[GDSConverter]
        C --> D[PalaceSimulator]
        D --> E[S-Parameter Dataset]
        E --> F[ModelTrainer]
        F --> G[OnnxExporter]
        G --> H[ModelTester]
        H --> I[.onnx Surrogate]

Pipeline Stages

Stage Purpose
GDSGenerator Samples geometry parameters and writes GDS layout files
DRCChecker Snaps layouts to the manufacturing grid and drops those violating the SG13G2 design rules or with port markers off their metal
GDSConverter Converts GDS files to Palace-compatible mesh inputs
PalaceSimulator Runs full-wave EM simulations and stores Touchstone results
ModelTrainer Trains a PyTorch neural network on the simulation dataset
OnnxExporter Exports the trained model to portable ONNX format
ModelTester Evaluates prediction accuracy against held-out simulation data

Geometry Class

A geometry class defines everything ORCA needs to sample and simulate a component:

  • Parameter ranges via InputParameterIterator of RangeParameter / ChoiceParameter objects (min/max, step, data type and uniform or log sampling per parameter; space-filling Sobol' (default), Latin hypercube, random or grid picking, with samples on the boundary of the box).
  • GDS generation via create_gds_file(name, output_path, params).
  • Basis expansions of the model inputs via BasisExpansion (e.g. a Chebyshev expansion of frequency), chosen on ModelTrainer and shared by every architecture.
  • Dataset type (e.g. GeoToSParamDatasetSingleFrequency).
  • Model architecture via OrcaModel subclasses registered with register_model, chosen on ModelTrainer(model=...).
  • Hyperparameter search space via the model's hyperparameter_search_space() — Optuna distributions per tunable parameter, merged with the trainer's and the basis expansion's.

Inputs and Outputs

- A geometry class (`BaseGeometry` subclass) with parameter definitions.
- A stackup XML file describing the physical layer stack.
- A Palace simulation configuration file (`.simcfg`).
- A Palace executable (local or containerised via Apptainer).
- Touchstone (`.sNp`) S-parameter files per simulated geometry variant.
- Trained PyTorch model checkpoint.
- Exported ONNX surrogate model (`.onnx`).
- Full run context stored under `output/<geometry_name>/`.
  • COBRA consumes ORCA surrogate models for circuit-level RFIC optimization.
  • Palace is used as the full-wave EM solver backend.
  • gdsfactory generates the parametric GDS layouts.
  • gds2palace converts GDS layouts into Palace simulation setups.

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