A Domain Adaptive Position Reconstruction Method for Time Projection Chamber based on Deep Neural Network

TL;DR

The paper addresses domain shift between Monte Carlo simulations and real data in transverse position reconstruction for dual‑phase Time Projection Chambers. It introduces a CycleGAN‑based domain adaptation framework to translate simulated signals into data‑like distributions, followed by a Deep Residual Network for XY regression. The approach yields substantial improvements: a $60.6\%$ improvement in boundary reconstruction on a RELICS prototype and at least $27\%$ improvement in a simulated 50‑kg TPC when evaluating resolution versus event energy. This method reduces biases from model‑data mismatch and shows potential for broader applicability to pattern‑based detector reconstructions.

Abstract

Transverse position reconstruction in a Time Projection Chamber (TPC) is crucial for accurate particle tracking and classification, and is typically accomplished using machine learning techniques. However, these methods often exhibit biases and limited resolution due to incompatibility between real experimental data and simulated training samples. To mitigate this issue, we present a domain-adaptive reconstruction approach based on a cycle-consistent generative adversarial network. In the prototype detector, the application of this method led to a 60.6% increase in the reconstructed radial boundary. Scaling this method to a simulated 50-kg TPC, by evaluating the resolution of simulated events, an additional improvement of at least 27% is achieved.

Figures (16)

• Figure 1: Schematic illustration of the training framework. The upper part (Domain Adaption stage) establishes mappings between the simulation domain and the data domain through two Generators F and G. Their outputs are evaluated by discriminators A and B, which attempt to distinguish between the generator-produced samples and the real samples belonging to the corresponding domain.The blue and red arrows represent two training cycles. The black arrows represent other training processes. The lower part (Regression stage) uses paired coordinates together with the transformed data as input to a regressor for position reconstruction. The green arrows represent the inference process.
• Figure 2: The left panel illustrates the architecture of the generators $G$ and $F$, which are responsible for translating data between two domains in the CycleGAN framework. The right panel shows the structure of the discriminators $D_X$ and $D_Y$, which are used to distinguish between real and generated samples. A discriminator outputs a score between 0 and 1; outputs greater than 0.5 are classified as real, while outputs less than 0.5 are considered fake.
• Figure 3: The right panel shows the architecture of the Deep Residual Network, while the left illustrates the detailed structure of a single residual block. The input and output channels $X$ and $Y$ are adjusted according to the channel labels indicated in the network diagram on the right.A shortcut connection is added in each residual block to mitigate gradient vanishing and improve overall network performance.
• Figure 4: (a) XY plane density; (b) ZX plane density. Relative scintillation intensity distribution in a single 5 mm $\times$ 5 mm electrode grid (maximum value = 1). The color represents the normalized scintillation intensity in each plane. This distribution is calculated for a single cell using Garfield++ based on the electric field simulated by COMSOL. The strongest scintillation occurs at the center of the cell where the electric field is maximal. Near the grid edges, scintillation can also appear due to the influence of the anode wires.
• Figure 5: Scintillation simulated in gaseous Xenon. Each point of a certain color represents the location where an individual electron emits a photon in gaseous xenon, approximately tracing its path under the influence of the electric field. The blue horizontal line indicates the liquid surface level. The red region illustrates the shape of the anode wires, and each red rectangle spaced by 5 mm represents the position of an anode wire perpendicular to the X-axis, extending outward along the vertical coordinate system. The initial positions of the electrons in liquid xenon are uniformly distributed.