ORCA: An AI-Assisted Surrogate Modelling Pipeline for RFIC Passives

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

Authors

Gianluca Simone, David Lurz, Martin Grund, Fabian Schneider, Michael Loose, Sascha Breun, Manuel Koch, Robert Weigel, Norman Franchi

Published

SBCCI 2026 (conference contribution)

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

• New users: go to Getting Started -> Installation.
• First successful run: Getting Started -> Quickstart.
• Bring your own geometry: User Guide -> Custom Classes.
• API details: see the source and docstrings in src/orca/.

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 --> 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
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 (sampling strategy, min/max per parameter).
• GDS generation via create_gds_file(name, output_path, params).
• Feature engineering via FeatureTransformPipeline (ratios, Chebyshev features, etc.).
• Dataset type (e.g. GeoToSParamDatasetSingleFrequency).
• Model factory via get_model(hyperparameters) — called by ModelTrainer during Optuna tuning.
• Hyperparameter search space via get_hyperparameter_search_space() — Optuna distributions per tunable parameter.

Inputs and Outputs

InputsOutputs

- 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>/`.

Related Projects

• COBRA consumes ORCA surrogate models for circuit-level RFIC optimization.
• Palace is used as the EM solver backend.

---

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