CPU-Net (Cyclic Positional U-Net) is a transfer-learning framework that learns an ad-hoc translation between simulated and measured HPGe detector pulses. It is designed to emulate readout electronics effects in pulse-shape simulations using an unpaired, CycleGAN-style training strategy.

Paper: CycleGAN-driven transfer learning for electronics response emulation in high-purity germanium detectors (Machine Learning: Science and Technology, 2026).
DOI: 10.1088/2632-2153/ae3052

Data & pretrained weights (Zenodo):
DOI: 10.5281/zenodo.15311838

• CycleGAN backbone couples a REN (sim → data) and an IREN (data → sim), enforcing cycle- and identity-consistency.
• Positional U-Net generators use layer-wise positional encodings to better preserve pulse timing/structure (esp. tails).
• Attention-augmented RNN discriminators evaluate translations and guide adversarial training.
• Translated pulses reproduce ensemble distributions (e.g., current amplitude, drift time, tail slope) better than raw simulations.
• Architecture is adaptable to other scientific time-series domains where noise convolution / deconvolution is required.

conda env create -f environment.yml
conda activate cpu-net

pip install -e . --no-deps

pip install pytest
pytest -q

Note: We pin NumPy < 2 (see environment.yml / pyproject.toml) to avoid incompatibilities with compiled dependencies in common scientific Python stacks.

Data and pretrained weights are packaged on Zenodo: 10.5281/zenodo.15311838.

• fep_REN.pt, fep_IREN.pt — generators (sim → data, data → sim)
• fep_netD_A.pth, fep_netD_B.pth — discriminators
• _wf_ornl.pickle — real FEP / SEP / DEP detector pulses
• _wf_sim.pickle — matching Geant4 + siggen simulations

Each pickle entry:

{
    "wf":     np.ndarray,  # (800,) aligned & normalized waveform
    "tp0":    int,         # index of 0% rise
    "energy": float        # calibrated energy (keV)
}

Detector pulses were recorded with an inverted‑coax HPGe detector (serial V06643A) during a –228Th–flood calibration at Oak Ridge National Laboratory. Signals were digitised by FlashCam module; pygama handled conversion to HDF5.

Simulated pulses originate from a Geant4 model of the same setup. Energy deposits were fed to siggen to generate raw charge‑collection pulses.

The easiest way to reproduce figures/metrics is via the notebooks:

• notebooks/Training.ipynb — training loop (CycleGAN training of REN/IREN)
• notebooks/Validation.ipynb — validation on held-out datasets and distribution-level metrics

Core package code:

• src/cpunet/network.py — CPU-Net model definitions (Positional U-Net + discriminator(s))
• src/cpunet/dataset.py — dataset loading & preprocessing (pickled pulses)
• src/cpunet/tools.py — DSP helpers, metrics (IoU / ROC), plotting utilities

Notebooks:

• notebooks/Training.ipynb
• notebooks/Validation.ipynb

Tests:

• tests/test_imports.py — import/smoke test for packaging integrity

Typical training runtime: ~1 GPU-hour on an NVIDIA A100 (40 GB) for 7000 iterations.

• Maximum current amplitude (Imax): distinguishes single vs multi-site; translated pulses should match detector distribution.
• Drift time (Tdrift): time between 1% and 100% rise; checks realistic charge-collection timing after translation.
• (Optional) Tail decay constant τ: distribution-level check of RC decay behavior.

If you use this code or the associated data in academic work, please cite the paper:

• Kevin Bhimani, Julieta Gruszko, Morgan Clark, John Wilkerson, Aobo Li,
  CycleGAN-driven transfer learning for electronics response emulation in high-purity germanium detectors,
  Machine Learning: Science and Technology (2026). DOI: 10.1088/2632-2153/ae3052

Zenodo dataset DOI: 10.5281/zenodo.15311838

(Also see CITATION.cff.)

Released under the MIT License (see LICENSE).