Skip to content

WORKFLOW.md

Latest commit

 

History

History
185 lines (117 loc) · 3.77 KB

WORKFLOW.md

File metadata and controls

185 lines (117 loc) · 3.77 KB
Raw

Workflow

This document describes the intended operational workflow of the project.

Stage 1: Create Layout Templates

Input:

  • process metadata
  • layer definitions
  • pixelized metal patterns
  • port locations

Tooling:

  • Data_process/JSON_layout_create/

Output:

  • one or more layout JSON files

Each layout JSON contains:

  • metadata
  • per-layer binary matrices
  • port definitions

Why this stage matters:

  • it is the upstream source of layout samples for the entire repository
  • the ADS automation stage consumes these JSON files directly
  • the dataset construction stage later reuses the same JSON files when aligning layouts with .sNp results

Important usage note:

  • for first-time quickstart validation, this stage can be skipped by using the sample JSON files already included under parallel_version/config_examples/json_layout/
  • for actual data generation campaigns, this stage is essential

Stage 2: Build ADS Designs

Input:

  • layout JSON files
  • layer mapping configuration
  • workspace configuration
  • PDK or reference technology configuration

Tooling:

  • parallel_version/subprocess_cli_parallel.py
  • parallel_version/subprocess_worker_parallel.py

Actions:

  • detect ADS Python runtime
  • create or open workspace and library
  • create cell and layout view
  • convert layout matrices to ADS geometry
  • place ports using border definitions
  • create rfpro_view

Output:

  • ADS workspace objects
  • generated layout and EM views

Stage 3: Run EM Simulations

Input:

  • created ADS design data
  • frequency configuration
  • export configuration

Actions:

  • configure simulation options
  • execute RFPro / ADS EM simulation
  • export simulation outputs

Output:

  • Touchstone .sNp
  • optional CSV exports
  • optional ADS dataset exports

Stage 4: Build Machine Learning Dataset

Input:

  • layout JSON files
  • matching .sNp files

Tooling:

  • Data_process/HDF5_create/create_hdf5.py

Actions:

  • convert layout matrices to padded tensors
  • interpolate S-parameters onto target frequencies
  • flatten complex responses into real-valued vectors
  • group compatible samples by port count and matrix size
  • write HDF5 datasets

Output:

  • HDF5 dataset files for training

Stage 5: Train CNN Surrogate Model

Input:

  • HDF5 dataset

Tooling:

  • Pytorch_Model/src/train.py
  • Pytorch_Model/src/dataset.py
  • Pytorch_Model/src/model.py

Actions:

  • load HDF5 tensors and metadata
  • split into training and validation sets
  • build CNN regressor
  • train on layout tensors and S-parameter vectors
  • save best-performing checkpoint

Output:

  • trained .pth weights

Stage 6: Verify Model Behavior

Input:

  • HDF5 dataset
  • trained model checkpoint

Tooling:

  • Pytorch_Model/src/tools/verify_model.py

Optional experimental helper:

  • Pytorch_Model/src/tools/experimental_gui/verify_model_gui.py

Actions:

  • run inference on selected samples
  • compare predicted and true S-parameter responses
  • inspect plots and error metrics
  • optionally use the experimental GUI wrapper for ad-hoc local inspection

Output:

  • qualitative and quantitative validation results

Recommended Public Execution Line

The recommended public execution line is:

  1. Data_process/JSON_layout_create/
  2. parallel_version/
  3. Data_process/HDF5_create/
  4. Pytorch_Model/

serial_version/ should remain reference material only.

Known Preconditions

End-to-end runtime validation requires:

  • Windows-compatible ADS installation
  • ADS Python runtime
  • RFPro / EM tooling
  • valid license access
  • available PDK or reference technology libraries

Known Repository Constraints

  • simulation outputs are runtime artifacts and should not be committed
  • training datasets and model weights should not be committed by default
  • some historical files reflect the original development machine and require cleanup before publication