Neural Networks for 3D Characterisation of AGATA Crystals

TL;DR

This work tackles the problem of precise 3D gamma-ray interaction localisation in the AGATA detector by replacing traditional PSCS-derived bases with LSTM-based neural networks trained on experimental Strasbourg data. A masked loss function enables training from 2D ground-truth scans to predict full 3D positions, yielding experimental signal bases that improve PSA performance beyond both simulated bases and PSCS-derived bases. The NN-based bases achieve mean position errors around 2 mm (on two known axes) and exhibit near-isotropic, lower-variance signal bases, enhancing both accuracy and consistency. The study also demonstrates the feasibility of emulated 1-D scans for faster crystal characterisation, and discusses practical implications, including the computational cost of training and the potential for broader application to other AGATA crystals.

Abstract

Precise localisation of gamma-ray interactions is crucial for the performance of the Advanced GAmma Tracking Array (AGATA). The Pulse Shape Analysis (PSA) method used for the position estimation of gamma-ray interactions relies on a simulated signal database. The Pulse Shape Comparison Scanning (PSCS) method was used to scan AGATA crystals in order to produce an experimental database of signals. This paper presents a novel approach using Long Short-Term Memory (LSTM) neural networks to determine the 3D interaction position of gamma rays within AGATA crystals, trained on data from IPHC Strasbourg, allowing for the construction of an experimental database. A custom masked loss function is introduced to enable training with incomplete position information. The database generated by this new method outperforms the existing simulated database, and the experimental database obtained from the conventional PSCS algorithm.

Figures (12)

• Figure 1: An example of ADL simulated AGATA detector signals (traces) for interactions at different radial positions (R) within segment A3. The AGATA nomenclature (top right) and crystal structure (bottom right) are shown for reference. The varying shapes and polarities of the transient signals in neighbouring segments (B3, F3, A2, A4) highlight the position sensitivity used for PSA and NN-based positioning.
• Figure 2: Schematic diagram illustrating the NN architecture.
• Figure 3: The error (mm) on the validation dataset at photopeak energy versus segment number for different training methods: using a uniformly sampled subset (blue), using the full dataset with a single model trained with balanced batches per segment (orange), and training one model per segment (green). Error is calculated on the two known axes. The segments index is explained in the text.
• Figure 4: Training and validation loss curves for segment F6, without (blue/orange) and with (green/red) the Gaussian noise layer, demonstrating reduced overfitting.
• Figure 5: The mean absolute error on the validation dataset versus segment number. Comparison between a model trained on the standard 2D dataset (blue) and a model trained on the simulated 1D dataset (orange). The error shown is calculated only for the axis that was unknown during the 1D training but known in the original 2D dataset. The segment indexing follows the same nomenclature as defined for Fig. \ref{['fig:Training_process_error_methodo']}.