CycleGAN-Driven Transfer Learning for Electronics Response Emulation in High-Purity Germanium Detectors

TL;DR

This work tackles the challenge of accurately modeling electronics response in pulse-shape simulations for HPGe detectors by learning a data-driven translation from simulated to data-like pulses. The authors introduce CPU-Net, a CycleGAN-based framework combining a Positional U-Net generator with an RNN-attention discriminator to perform unpaired, bidirectional translations between simulated and measured pulses, while preserving key topology information for PSD. Through rigorous data preparation, training with cycle-consistency and adversarial losses, and distribution-level validation on drift time, I_{max}, and tail-slope parameters, CPU-Net achieves up to a fourfold improvement in distribution-level agreement without relying on explicit electronic-transfer-function fitting. The method offers a flexible, detector-agnostic approach to electronics emulation that can enhance PSD performance and support broader applications in time-series data where first-principles models are incomplete.

Abstract

High-Purity Germanium (HPGe) detectors are a key technology for rare-event searches such as neutrinoless double-beta decay (\ensuremath{0νββ}) and dark matter experiments. Pulse shapes from these detectors vary with interaction topology and thus encode information critical for event classification. Pulse shape simulations (PSS) are essential for modeling analysis cuts that distinguish signal events from backgrounds and for generating reliable simulations of energy spectra. Traditional PSS methods rely on a series of first-principles corrections to replicate the effect of readout electronics, requiring challenging fits over large parameter spaces and often failing to accurately model the data. We present a neural network architecture, the Cyclic Positional U-Net (https://github.com/aobol/CPU-Net), that performs translations of simulated pulses so that they closely resemble measured detector signals. Using a Cycle Generative Adversarial Network (CycleGAN) framework, this {Response Emulation Network} (REN) learns a data-driven mapping between simulated and measured pulses with high fidelity, without requiring a predetermined response model. We use data from a High-Purity Germanium (HPGe) detector with an inverted-coaxial point contact (ICPC) geometry to show that \texttt{CPU-Net} effectively captures and reproduces critical pulse shape features, allowing more realistic simulations without detector-specific tuning. \texttt{CPU-Net} achieves up to a factor-of-four improvement in distribution-level agreement for pulse shape parameter reconstruction, while preserving the topology-dependent information required for pulse-shape discrimination.

Figures (10)

• Figure 1: Left: A cross section of the ICPC detector used, with the color bar depicting electric field magnitude and the white lines showing possible charge drift paths. The detector has $40.5$ mm radius and $96.0$ mm height. Image adapted with permission from the LEGEND Collaboration. Center: Four events in the detector, simulated using Geant4. Right: Pulse shape simulations, generated using siggen software, corresponding to the four events shown. Pulses are aligned by their start time and normalized to the total event energy.
• Figure 2: Left: Layer-wise breakdown of the Positional U-Net. The blue lines represent the skip connections between contracting and expanding paths of the U-Net. Positional encoding layers are also shown for all levels. Right: Flow of pulses in the RNN+Attention discriminator. The embedded waveforms pass through a bidirectional GRU and attention layer to form context-aware features. The output is concatenated with hidden states from the GRU. This is fed to a fully connected network which produces a classification decision using a sigmoid function.
• Figure 3: A flowchart representing the cycle-consistent adversarial training of the Cycle-GAN network. The solid lines indicate the flow of training instances through the PU-Net Response Emulation Network (REN) and Inverse Response Emulation Network (IREN) generators, and dotted lines represent how each stage is used to produce a subset of the losses used to train the network. The full list of the losses used is given in Table \ref{['tab:loss_summary']}. The red lines represent the forward-cycle, which begins with simulated pulses and transfers them to data-like pulses using the REN. The Data Discriminator $\delta_D$ differentiates data-like pulses from true data and is used for adversarial training of the REN. Completing the cycle through the IREN to produce recovered simulation pulses allows the use of an a weighted mean absolute error (L1) loss to ensure cycle consistency. The reverse cycle, shown by the blue lines, is trained simultaneously with the same strategy, beginning with data pulses.
• Figure 4: Left: Calibrated energy spectrum from a $^{228}$Th source at ORNL compared to the Monte Carlo simulation results. Key peaks from $^{228}$Th source are labeled. The simulation does not include detector surface effects, nor any electronic‑noise or energy resolution effects, thus resulting in a mismatch. Right: The simulation geometry of the ORNL characterization setup, showing the detector in green, the source in red, the aluminum cryostat and holder in dark grey, and the lead shielding in light grey.
• Figure 5: Training losses for CPU-Net. Curves are smoothed using a moving average of 10 iterations for clarity. The identity and cycle losses rapidly converge, indicating that the translators learn to preserve the cycle and identity within the first few hundred iterations. Following an initial training phase in which the discriminator dominates, the generator and discriminator losses oscillate. This adversarial dynamic helps both models to learn and improve each other.