TL;DR
This paper introduces a novel application of Generative Adversarial Networks (GANs), specifically Wasserstein GANs, to simulate scintillation signals in the EXO-200 experiment. By training on real calibration data, the GAN can generate realistic detector signals an order of magnitude faster than traditional Monte Carlo methods, while accurately capturing complex detector physics like position dependency and even identifying faulty channels.
What Problem Does This Paper Address?
Simulating particle detector responses is crucial for data analysis, detector design, and new physics searches. However, traditional Monte Carlo (MC) simulations, especially for complex detectors like the EXO-200 time projection chamber (TPC), are computationally expensive and time-consuming. This is particularly true for simulating scintillation light, which involves generating and tracking millions of photons. Furthermore, MC simulations rely on precise knowledge of detector geometry and material optical properties, which are often not perfectly known, leading to discrepancies between simulated and real data. The EXO-200 experiment, for instance, uses a simplified parametric simulation for light yield due to these challenges. This paper aims to address these limitations by developing a faster, more accurate, and data-driven simulation method using deep learning.
Methods
The core of this work is the application of a Wasserstein Generative Adversarial Network (WGAN). Unlike traditional GANs, WGANs use the Wasserstein distance as a loss function, which provides more stable training and better gradients, preventing issues like training collapse. The WGAN is trained on raw scintillation waveforms from the EXO-200 detector's avalanche photodiode (APD) channels.
Key methodological aspects include:
- Data Representation: APD waveforms are converted into image-like representations (74 channels x 350 time samples) to leverage GANs' strength in image generation.
- Label Conditioning: The generator is conditioned on physics labels (scintillation energy and spatial coordinates X, Y, Z) to produce waveforms with specific characteristics. These labels are derived from the EXO-200 reconstruction algorithm.
- Training Strategy: The WGAN is trained using the Adam optimizer with a specific learning rate schedule and a gradient penalty to enforce the Lipschitz constraint on the discriminator. The training involves alternating updates between the discriminator and the generator.
- Validation: The generated waveforms are evaluated at multiple levels: raw waveform comparison, signal feature comparison, and finally, by integrating them into the EXO-200 reconstruction framework to compare reconstructed energy spectra.
Datasets
The primary dataset used in this study is real calibration data from the EXO-200 experiment, specifically from 228Th calibration runs. This data is split into training and testing sets with a ratio of approximately 7:3. The training data consists of 215,789 events from phase II of the experiment. The labels (energy and position) for training are derived from the existing EXO-200 reconstruction algorithm applied to this calibration data.
Key Insights and Ideas
- Data-Driven Simulation: The GAN approach bypasses the need for detailed, often uncertain, physical models of light propagation and detector response, relying instead on learning directly from real data.
- Speed and Efficiency: The GAN can generate simulated waveforms an order of magnitude faster than traditional MC methods, significantly reducing computational costs.
- Capturing Complex Physics: The GAN successfully learns and reproduces subtle physics phenomena, such as the position-dependent light response across different APD planes and the anti-correlation between scintillation and charge signals.
- Identifying Detector Issues: The network implicitly learns to recognize and reproduce the behavior of "dead" APD channels caused by excessive leakage current, a testament to its ability to capture real-world detector characteristics.
- Reconstruction Compatibility: The generated waveforms, when fed into the standard EXO-200 reconstruction pipeline, produce energy spectra comparable to those from real data, demonstrating the practical utility of the simulated signals.
Key Results
- Speedup: The GAN achieves an average simulation rate of ~69 events/s on an Nvidia GeForce 1080Ti GPU, compared to ~4.2 events/s for the traditional EXO-200 simulation framework (which includes Geant4 and parametric optical simulation). This represents an approximate order of magnitude speedup.
- Energy Resolution: While the GAN-generated energy spectra are comparable to real data, they exhibit slightly worse energy resolution. For the 2.614 MeV 228Th peak, the resolution is 9.1% for GAN-generated data versus 7.6% for real data without position correction. After applying a light map correction, the resolutions improve to 2.0% for GAN and 1.3% for real data, respectively.
- Spatial Dependence: The GAN accurately reproduces the spatial dependence of APD amplitudes in both planes, showing good agreement with real data in terms of average amplitudes and standard deviations.
- Disconnected Channels: The GAN successfully identifies and reproduces the pattern of disconnected APD channels, matching the calibration data.
Limitations and Future Work
- Energy Label Uncertainty: The primary limitation acknowledged is the uncertainty in the energy labels derived from the EXO-200 reconstruction (~5% for the 228Th peak). This imperfect labeling is believed to contribute to the slightly worse energy resolution of the GAN-generated data.
- Training Data Source: The approach requires training on real data, which means it cannot be used to optimize the design of entirely new detectors as effectively as traditional MC methods that can explore hypothetical detector configurations.
- Single-Site Events: The current study focused on single-site events for simplicity. The authors suggest that extending the approach to multi-site events, which are common in TPCs, is a promising future direction.
- Improved Training Data: Future work could involve using a light source with known intensity and controlled deployment to generate training events with more accurate "truth" labels, potentially leading to better energy resolution.
- Joint Charge and Light Waveform Training: An interesting avenue for future research is to train the network on both charge and light waveforms simultaneously. If the network can learn the anti-correlation between these signals, it could potentially leverage the improved accuracy of rotated energy labels.
Summary updated