CycleGAN-driven transfer learning for electronics response emulation in high-purity germanium detectors

Kevin Bhimani, Julieta Gruszko, Morgan Clark, John F. Wilkerson, Aobo LiMach. Learn. Sci. Technol 2026 · DOI

TL;DR

This paper introduces CPU-Net, a novel CycleGAN-based deep learning framework that translates simulated High-Purity Germanium (HPGe) detector pulses into realistic data-like pulses. This approach bypasses the need for complex, first-principles modeling of detector electronics, significantly improving the fidelity of simulations for rare-event searches.

What Problem Does This Paper Address?

High-Purity Germanium (HPGe) detectors are crucial for experiments searching for rare events like neutrinoless double-beta decay and dark matter. Accurately simulating the pulses from these detectors is vital for event classification and background rejection. Traditional pulse shape simulations (PSS) struggle to precisely model the complex effects of readout electronics, often requiring extensive parameter fitting and failing to capture subtle, detector-specific behaviors. This leads to discrepancies between simulated and real data, hindering the development of precise analysis cuts and reliable simulations.

Methods

The core of this work is the Cyclic Positional U-Net (CPU-Net), a deep learning architecture built upon the CycleGAN framework. CPU-Net employs a Positional U-Net (PU-Net) as its generator, which is a U-Net architecture enhanced with layer-wise positional encoding to better capture temporal information in the pulse shapes. This PU-Net is used for both the Response Emulation Network (REN), which translates simulated pulses to data-like pulses, and the Inverse Response Emulation Network (IREN), which performs the reverse translation.

The training is unsupervised, leveraging a CycleGAN approach to learn the mapping between simulated and real data pulses without requiring paired examples. This involves:

  • Adversarial Training: Two discriminators (one for simulated data, one for real data) are trained to distinguish between real and translated pulses, while the generators (REN and IREN) are trained to "fool" these discriminators.
  • Cycle Consistency Loss: This loss ensures that translating a pulse from one domain to another and back again results in a pulse that is very similar to the original. This is crucial for preserving the underlying physics information.
  • Identity Loss: This loss ensures that if a pulse from one domain is fed into the generator for that domain, the output should be very close to the input.

Datasets

The study utilizes characterization data from a LEGEND Inverted-Coaxial Point-Contact (ICPC) detector (V06643A).

  • Training Data: 110,000 Full Energy Peak (FEP) pulses from a 228Th source were used for training.
  • Validation Data: 12,000 Single Escape Peak (SEP) events and 3,000 Double Escape Peak (DEP) events were used for validation.
  • Simulated Data: Geant4 was used to simulate particle interactions, and siggen was used to generate pulse shapes from these interactions.

Key Insights and Ideas

  • Data-Driven Electronics Emulation: The most significant insight is that a deep learning model can effectively learn and replicate the complex, often unmodeled, effects of detector electronics without explicit physical modeling. This bypasses the limitations of traditional PSS.
  • Preservation of Physics: CPU-Net successfully translates pulse shapes while preserving critical topological information (e.g., single-site vs. multi-site events), which is essential for pulse-shape discrimination (PSD).
  • CycleGAN for Unpaired Data: The application of CycleGAN is key, as it allows for training on unpaired simulated and real data, a common challenge in experimental physics where perfectly matched datasets are difficult or impossible to obtain.
  • Positional U-Net for Temporal Features: The incorporation of positional encoding into the U-Net architecture is crucial for capturing the temporal dynamics of the pulse, particularly the rising and falling edges, which are vital for PSD.

Key Results

  • Improved Pulse Shape Parameter Agreement: CPU-Net achieves significant improvements in matching the distributions of key pulse shape parameters (drift time, maximum current amplitude, tail slope) between simulated and real data.
  • Drift Time (Tdrift): SEP dataset IoU increased from 39.5% to 62.3% (1.6x improvement). DEP dataset IoU increased from 5.4% to 22.5% (4x improvement).
  • Maximum Current Amplitude (Imax): SEP dataset IoU increased from 27.5% to 63.7% (2.3x improvement). DEP dataset IoU increased from 4.2% to 15.5% (3.7x improvement).
  • Tail Slope (τ): REN-translated pulses showed a mean τ of 53.75μs, deviating by only 1.6% from the measured data value of 54.60μs.
  • Realistic Pulse Features: The translated pulses capture subtle features like pulse overshoot and the impact of multiple RC decay constants, which are often missing in standard simulations.
  • Training Stability: The paper details strategies for achieving stable training in the complex CycleGAN framework, including careful weighting of losses and hyperparameter tuning.

Limitations and Future Work

  • Ensemble-Level Discrepancies: CPU-Net preserves ensemble-level discrepancies observed in the simulations (e.g., too-narrow peaks in distributions). The authors suggest that using more realistic charge cloud models in simulations could further improve agreement.
  • Uncertainty from Stochasticity: The multi-peaked distributions in IoU values indicate that adversarial training can converge to different modes, suggesting uncertainty arising from stochastic optimization and initialization. Using pre-trained weights is proposed as a strategy to mitigate this.
  • Future Work:
    • Denoising Detector Pulses: The bidirectional nature of CPU-Net could be used to denoise real detector pulses, potentially revealing subtle event topology features.
  • Quantitative Evaluation with Paired Data: Utilizing Compton scanning technique data to obtain position-tagged, paired waveforms would enable direct per-pulse evaluation.
  • Generalization: Studying how CPU-Net generalizes to other detectors, geometries, and changing experimental conditions.
  • Adaptation to Other Data Types: The framework could be adapted for time-series image data using CNN discriminators.

Summary updated

BibTeX
@article{bhimani2026cyclegan,
  author = {Kevin Bhimani and Julieta Gruszko and Morgan Clark and John F. Wilkerson and Aobo Li},
  title = {{CycleGAN-driven transfer learning for electronics response emulation in high-purity germanium detectors}},
  journal = {Mach. Learn. Sci. Technol},
  year = {2026},
  doi = {10.1088/2632-2153/ae3052},
}
Figure