Cyclic Positional U-Net (CPU-Net)

• The paper introduces CPU-Net as a learned post-simulation translator that corrects for missing detector-specific effects using adversarial and cycle-consistent training.
• CPU-Net integrates a 1D U-Net with layer-wise positional encoding to accurately capture temporal features, enhancing waveform tail fidelity.
• Cycle-consistent adversarial training with dual translators and discriminators ensures translated pulses match the statistical distribution of real HPGe detector signals.

Searching arXiv for the specified paper and closely related context. Cyclic Positional U-Net (CPU-Net) is a neural-network-based method for ad-hoc pulse shape simulation in high-purity germanium (HPGe) detectors. It was introduced to translate pulses generated by a conventional detector simulation into detector pulses that look like real readout waveforms, with high precision and low latency, without explicitly modeling all detector-specific nuisance effects from first principles. In the paper’s formulation, CPU-Net is not a replacement for the physical simulator; rather, it is a learned post-simulation translator that corrects missing detector and electronics effects through unpaired domain translation between simulated and detector pulse distributions (Li et al., 2022).

1. Definition and scientific setting

CPU-Net was proposed for HPGe detectors used in rare-event searches such as neutrinoless double-beta decay and dark matter experiments. In these detectors, a particle interaction produces electron-hole pairs, which drift to electrodes and induce a current or voltage waveform. The detailed pulse shape depends on deposited energy, interaction position, charge-cloud size, event topology such as single-site versus multi-site, charge trapping, surface effects, and front-end electronics response including preamplifier integration and RC discharge (Li et al., 2022).

Pulse-shape simulation is scientifically valuable because realistic simulated pulses with a known mapping to underlying particle interactions can be used to study why pulse-shape discrimination cuts accept or reject certain events, design improved cuts, test detector geometries, and generate training data for machine-learning analyses. The paper frames the central problem as a limitation of conventional simulation pipelines: electric-field simulation together with tools such as fieldgen, siggen, or SolidStateDetectors.jl, often coupled to Geant4, can simulate idealized induced signals, but reproducing the final detector pulse requires detector-specific corrections that are difficult to model a priori, may vary from detector to detector, and may change with operating conditions or over time (Li et al., 2022).

The authors describe this difficulty as a “curse of the first principle.” Even when the base physics simulation is correct, the detailed detector waveform still depends on many nuisance parameters and on expensive high-dimensional fitting. CPU-Net addresses that difficulty by learning a direct translation from simulated pulses to realistic detector pulses using transfer learning on unpaired datasets. Instead of modifying the simulator $P(X \mid Y)$ itself, the model learns an ad-hoc translation network that maps a simulated waveform $X$ into a corrected waveform $\hat{X}$ whose distribution matches real detector pulses $X'$ (Li et al., 2022).

2. Problem formulation and domain translation

The paper formalizes the transfer-learning problem through source and target domains and tasks. The source domain is

$\mathcal{D}_{Source}=\{\mathcal{Y},P(Y)\}$

where $Y$ contains reconstruction parameters such as energy, maximum current amplitude, tail slope, and related quantities. The source task is

$$\mathcal{T}_{Source}=\{\mathcal{X},P(X|Y)\} \qquad X\in \mathcal{X}.$$

The target domain is $\mathcal{D}_{Target}=\{\mathcal{Y}',P(Y')\}$ and the target task is

$\mathcal{T}_{Target}=\{\mathcal{X}',P(X'|Y')\},$

where $X'$ denotes real detector pulses (Li et al., 2022).

To connect these domains, the paper introduces the Ad-hoc Translation Network (ATN),

$X$0

while noting that the paper contains LaTeX typos and that this is the intended meaning. The translated collection $X$1 is required to become indistinguishable from $X$2, so that

$X$3

In this formulation, applying the learned translator $X$4 to source-task pulses approximates the target-task pulse distribution (Li et al., 2022).

The paper repeatedly uses the term transfer learning, but not in the sense of initializing from a pretrained external model and freezing layers. Instead, transfer learning denotes transfer from the simulation domain to the detector domain by learning domain translation. There is no stated pretrained backbone, no stated frozen layers, and no separate source-task pretraining stage distinct from adversarial training. A plausible implication is that the principal novelty lies in domain adaptation between waveform distributions rather than in parameter reuse from a prior model (Li et al., 2022).

3. Architecture: the “positional” component

The architectural backbone is a 1D U-Net adapted from image processing to time-series waveforms. The paper describes U-Net as having $X$5 convolutional layers for encoding, $X$6 upsample layers for decoding, and $X$7 skip or contracting paths connecting encoder and decoder features at corresponding scales. For pulse translation, the waveform is treated as a 1D sequence, and all convolutions are Conv1d operations. The design is intended to allow information at multiple scales to flow to the decoder and to improve reconstruction efficiency (Li et al., 2022).

The paper reports that a plain U-Net did not reproduce the tail of the waveform, especially the detector-electronics-induced tail region shown in the example figure. The authors attribute this failure to a lack of positional information in intermediate layer outputs. Because convolutional neural networks are translation-equivariant and do not inherently encode absolute sequence position, they can struggle when the desired transformation depends on where in the waveform a feature occurs. For HPGe pulses, this matters because the leading edge and the late-time tail correspond to physically different detector and electronics processes (Li et al., 2022).

To address this, the authors define a Positional U-Net (PU-Net) by adding a layer-wise positional encoding map $X$8 to each convolutional-layer output and each upsample-layer input. The paper states that the output of a 1D convolutional layer has shape

$X$9

where $\hat{X}$0 is the number of channels and $\hat{X}$1 is the sequence length. The positional encoding tensor has matching $\hat{X}$2 shape, is repeated across batch, and then added elementwise. Because each U-Net level has different channel count and sequence length, the positional encoding must be generated separately for each layer (Li et al., 2022).

The paper gives the following Transformer-inspired sinusoidal encoding matrix:

$\hat{X}$3

This is the explicit mathematical definition of the positional component in the paper (Li et al., 2022).

The appendix adds a further architectural detail: a reparameterization trick is used at the bottom of the PU-Net. The authors state that this is inspired by the VAE literature and is used to increase stochasticity in the latent representation, which empirically helped the ATN and IATN better learn the distribution of reconstruction parameters. They explicitly note that the latent vector is reparameterized, but they do not use a KL-divergence regularizer as in a standard VAE. This suggests that PU-Net incorporates stochastic latent sampling without becoming a VAE in the usual sense (Li et al., 2022).

The text provided does not state the number of encoder or decoder levels, the exact channel progression, the exact kernel sizes, the exact strides, the normalization type, the activation functions, whether residual blocks are used, whether attention is used inside the U-Net, or the exact waveform sequence length. Those details therefore remain unspecified in the available description.

4. Cycle-consistent adversarial training: the “cyclic” component

The term “cyclic” refers to CycleGAN-style cycle-consistent adversarial training on unpaired source and target waveforms. Because paired simulated and real waveforms with identical underlying physical event parameters are not available, supervised paired training is impossible, and the source and target pulse datasets may come from different parameter distributions:

$\hat{X}$4

The training procedure therefore uses a cycle-consistent framework rather than direct supervision (Li et al., 2022).

The CPU-Net setup contains four learned models:

• ATN $\hat{X}$5: simulated pulse $\hat{X}$6 detector-like pulse
• IATN $\hat{X}$7: detector pulse $\hat{X}$8 simulated-like pulse
• Source discriminator $\hat{X}$9
• Target discriminator $X'$0

Both $X'$1 and $X'$2 use the PU-Net structure. The paper describes two cycle paths:

$X'$3

for the source cycle, and

$X'$4

for the target cycle. Each path contributes an adversarial loss and a cycle-consistency loss, and the total training minimizes all losses from both directions simultaneously (Li et al., 2022).

The paper provides explicit losses in the appendix. For the simulated-pulse translation path, the cycle-consistency loss is

$X'$5

The same type of cycle loss is defined in the reverse direction for detector pulses, although the reverse-direction formula is not separately printed. The adversarial loss is given as

$X'$6

The identity loss is

$X'$7

According to the appendix, an additional three complementary losses are defined for the reverse path, giving a total of six losses optimized simultaneously (Li et al., 2022).

The identity term has a specific operational interpretation: since $X'$8 should output detector-pulse-like events, if a real detector pulse $X'$9 is already supplied as input to $\mathcal{D}_{Source}=\{\mathcal{Y},P(Y)\}$0, the network should preserve it rather than distort it. This design constrains the translator to behave as a correction operator rather than as an unconstrained waveform generator (Li et al., 2022).

5. Discriminators, data, and training procedure

The discriminators are not CNNs. The paper instead uses an attention-coupled recurrent neural network because position is naturally handled by recurrent processing. Before entering the RNN discriminator, a normalized waveform is multiplied by 500, rounded to the nearest integer, one-hot encoded, and passed through a PyTorch embedding layer with embedding dimension

$\mathcal{D}_{Source}=\{\mathcal{Y},P(Y)\}$1

The discriminator itself is a single-layer bidirectional RNN using GRU cells internally. The RNN produces a 64-dimensional intermediate hidden vector $\mathcal{D}_{Source}=\{\mathcal{Y},P(Y)\}$2 at each time step $\mathcal{D}_{Source}=\{\mathcal{Y},P(Y)\}$3 and a final output vector $\mathcal{D}_{Source}=\{\mathcal{Y},P(Y)\}$4 (Li et al., 2022).

An attention matrix $\mathcal{D}_{Source}=\{\mathcal{Y},P(Y)\}$5 is then used to compute attention scores between $\mathcal{D}_{Source}=\{\mathcal{Y},P(Y)\}$6 and each intermediate vector $\mathcal{D}_{Source}=\{\mathcal{Y},P(Y)\}$7:

$\mathcal{D}_{Source}=\{\mathcal{Y},P(Y)\}$8

A context vector is formed by weighted summation over $\mathcal{D}_{Source}=\{\mathcal{Y},P(Y)\}$9 using weights $Y$0. The context vector is concatenated with $Y$1 and then passed through fully connected layers to produce a single discriminator output (Li et al., 2022).

The source dataset $Y$2 is generated with siggen. It contains 65,536 single-site waveforms uniformly simulated at different positions in the detector active volume, and some are converted into multi-site waveforms by randomly stacking two single-site waveforms. The target dataset $Y$3 is collected from a local HPGe setup with a $Y$4Th radioactive source attached to a copper cryostat system containing a Broad Energy Ge detector. An energy cut of 2094–2104 keV selects Single Escape Peak events for the target dataset. The paper states that over 90% of SEP events are multi-site, and that including additional backgrounds in the window, the target dataset is about 79% multi-site (Li et al., 2022).

The datasets are split into 70% training and 30% validation. Each raw waveform is normalized to $Y$5 using

$Y$6

The paper states that inputs to ATN and IATN are normalized in this way, and outputs are also normalized using the same method. No further alignment, windowing, or segmentation detail is given in the provided text (Li et al., 2022).

The appendix states that ADAM optimizers are used for all losses, the total training cycle contains 3000 batches, the batch size is 16, and the learning rate decays linearly from 0.001 to 0 with decay step size 500. A single training trial takes about 0.5 GPU hours on an NVIDIA A100 GPU. This hardware statement is the paper’s main training-cost characterization (Li et al., 2022).

6. Physical interpretation, evaluation, and limitations

The paper argues that CPU-Net learns detector-specific corrections missing in siggen. Two effects are emphasized. First, simulated pulses have a sharp turning edge, and ATN smooths this region, which the authors interpret as learning non-zero preamplifier integration time. Second, siggen pulses have flat tails with $Y$7, and ATN transforms these into exponentially decaying tails, which the authors interpret as learning RC discharge in the readout electronics. The central physical interpretation is therefore that CPU-Net acts as a learned detector-response corrector (Li et al., 2022).

The experimental context uses two waveform classes important in HPGe analysis. Single-site events correspond to energy deposited in one location and produce a single sharply rising step. Multi-site events correspond to multiple depositions at different sites and produce multiple steps. A key pulse-shape parameter for distinguishing them is the maximum current amplitude $Y$8: at fixed energy, single-site waveforms tend to have larger or faster $Y$9 than multi-site waveforms. Another important parameter is the normalized tail slope $$\mathcal{T}_{Source}=\{\mathcal{X},P(X|Y)\} \qquad X\in \mathcal{X}.$$0, which characterizes the electronics tail (Li et al., 2022).

The main quantitative evaluation metric is Histogram Intersection over Union (HIoU) between a translated-pulse histogram and the detector-pulse histogram. The appendix defines it procedurally: for each histogram bin, take the minimum count as intersection, take the maximum count as union, sum over bins, and HIoU ranges from 0% to 100%. The compact equation

$$\mathcal{T}_{Source}=\{\mathcal{X},P(X|Y)\} \qquad X\in \mathcal{X}.$$1

is inferred from the appendix’s verbal definition rather than explicitly typeset in the paper (Li et al., 2022).

The reported results focus on $$\mathcal{T}_{Source}=\{\mathcal{X},P(X|Y)\} \qquad X\in \mathcal{X}.$$2 and $$\mathcal{T}_{Source}=\{\mathcal{X},P(X|Y)\} \qquad X\in \mathcal{X}.$$3.

Quantity	Raw simulated pulses	ATN outputs
$$\mathcal{T}_{Source}=\{\mathcal{X},P(X|Y)\} \qquad X\in \mathcal{X}.$$4 HIoU	0%	$$\mathcal{T}_{Source}=\{\mathcal{X},P(X|Y)\} \qquad X\in \mathcal{X}.$$5
$$\mathcal{T}_{Source}=\{\mathcal{X},P(X|Y)\} \qquad X\in \mathcal{X}.$$6 HIoU	12.4%	$$\mathcal{T}_{Source}=\{\mathcal{X},P(X|Y)\} \qquad X\in \mathcal{X}.$$7

For $$\mathcal{T}_{Source}=\{\mathcal{X},P(X|Y)\} \qquad X\in \mathcal{X}.$$8, the improvement is from 0% for raw simulated pulses to $$\mathcal{T}_{Source}=\{\mathcal{X},P(X|Y)\} \qquad X\in \mathcal{X}.$$9 for ATN outputs. This is consequential within the paper because the simulated pulses originally have zero tail slope, while the translated outputs recover a target-like tail distribution. For $\mathcal{D}_{Target}=\{\mathcal{Y}',P(Y')\}$0, the improvement is from 12.4% to $\mathcal{D}_{Target}=\{\mathcal{Y}',P(Y')\}$1. The paper notes that this improvement is substantial but less dramatic than for $\mathcal{D}_{Target}=\{\mathcal{Y}',P(Y')\}$2, reflecting that $\mathcal{D}_{Target}=\{\mathcal{Y}',P(Y')\}$3 is tied not only to electronics response but also to event-topology distribution mismatch between source and target (Li et al., 2022).

The uncertainty procedure is described explicitly. The appendix reports training 56 trials of CPU-Net with different random seeds, computing HIoU on the same validation dataset of 3000 pulses, fitting Fechner’s Distribution to the HIoU distribution, deriving asymmetric statistical uncertainties, and treating deviation of the reported result from the fitted mean as a conservative systematic component. The lower uncertainty is larger because some runs fail due to unstable GAN training (Li et al., 2022).

The paper’s claim that translated pulses become “indistinguishable” from detector pulses is supported by adversarial training against a discriminator, qualitative waveform agreement in example traces, and distributional agreement in reconstruction parameters including $\mathcal{D}_{Target}=\{\mathcal{Y}',P(Y')\}$4 and $\mathcal{D}_{Target}=\{\mathcal{Y}',P(Y')\}$5. The paper does not provide a standalone blinded human test or a classifier AUC demonstrating strict indistinguishability. A common misconception would therefore be to read the term as a formal impossibility result; in the paper, the evidence is adversarial and distributional rather than a proof of indistinguishability (Li et al., 2022).

Several limitations are stated explicitly. ATN can occasionally generate a small but sharp “spike” in output pulses, which the authors speculate may arise from activation functions, positional encoding addition, or general U-Net structure effects. CPU-Net does not fully reproduce the long tail of high-$\mathcal{D}_{Target}=\{\mathcal{Y}',P(Y')\}$6 events, specifically for $\mathcal{D}_{Target}=\{\mathcal{Y}',P(Y')\}$7; the authors suggest either that CPU-Net failed to learn this minority population or that such events are missing in siggen simulation. Some training runs fail outright and yield much worse HIoU, which the paper identifies as a typical GAN issue because there is no guarantee that the adversarial game reaches equilibrium (Li et al., 2022).

Within the scope of the paper, CPU-Net is significant because it offers a practical way to improve pulse-shape realism without explicit detector-by-detector nuisance modeling. The claimed advantages are data-driven correction of missing detector and electronics effects, operation on unpaired datasets, fast training, straightforward adaptation to multiple detectors and changing operating conditions, preservation of physically meaningful pulse properties through cycle consistency, and the availability of both a forward translator and an inverse translator. Compared with conventional approaches, the paper positions CPU-Net as avoiding hand-designed nuisance parameter models, expensive characterization campaigns, and high-dimensional fitting over degenerate parameter spaces. Compared with a plain neural translator, the paper identifies two defining contributions: positional encoding in a 1D U-Net for waveform-tail fidelity, and cycle-consistent adversarial training on unpaired simulation and data waveforms (Li et al., 2022).