Abstract: Particle detection hardware often faces limitations when incident energy exceeds the dynamic range of the readout electronics, resulting in saturated signals and lost information. This study investigates the use of Artificial Neural Networks (ANNs) to restore these saturated $\gamma$ signal waveforms. By comparing Back Propagation (BP), Simple Recurrent (Elman), Radial Basis Function (RBF), and Generalized Radial Basis Function (GRBF) architectures against traditional mathematical fitting models, the researchers identified the GRBFNN as the most effective method for waveform recovery.
The Problem: Signal Saturation and Non-Linearity
In radiation detection, saturation occurs when a high-energy particle "maxes out" the voltage range of the readout electronics, producing a clipped, flat-topped waveform. This is not a simple matter of scaling; in liquid scintillation detectors, the pulse shape changes non-linearly with energy. As the energy increases, the pulse doesn't just grow taller; it also broadens significantly.
As shown in Figure 1, the normalized waveforms for different energy channels (from 40 to 240) exhibit distinct widths and decay characteristics. This non-linearity makes traditional geometric or linear reconstruction methods unreliable.
Methodology and Training Strategy
To address this, researchers from Sichuan University and the University of Edinburgh tested several ANN architectures. To train these networks, they utilized real-world background data from an EJ335 liquid scintillation detector. Since the "true" peak of a naturally saturated signal is unknown, the team employed a synthetic saturation strategy: they took healthy, unsaturated pulses and manually truncated them at random heights (4% to 99% of the peak) to create a ground-truth dataset.
The study compared four ANN types:
- Back Propagation (BP): Standard feed-forward network.
- Elman: A recurrent network suited for time-series data.
- Radical Basis Function (RBF): Known for simplicity and noise robustness.
- Generalized Radial Basis Function (GRBF): An advanced RBF offering higher accuracy and better avoidance of local minima.
Key Results: The Superiority of GRBFNN
The Generalized Radial Basis Function Neural Network (GRBFNN) emerged as the clear winner. While BP and RBF networks often produced physically unrealistic or discontinuous shapes, and Elman networks suffered from slow convergence, the GRBFNN reconstructed waveforms that were essentially identical to the original signals.
As illustrated in Figure 6, the GRBFNN (data 3) accurately predicts the missing peak of the clipped signal (data 2), closely matching the original ground-truth waveform (data 1). In contrast, traditional mathematical models like the Marrone model—a complex double-exponential fit—showed significant discrepancies, proving inadequate for high-energy saturation recovery.
Conclusion
The application of GRBFNNs provides a robust computational solution to hardware limitations in particle physics. By accurately "hallucinating" the missing peaks of saturated signals, this method effectively extends the dynamic range of existing detectors, allowing researchers to recover valuable high-energy data that would otherwise be discarded.
Summary updated