Speaker
Description
The CEPC tracker relies on a high-granularity Time Projection Chamber (TPC) to provide precise track reconstruction and particle identification. Traditional reconstruction relies on signals digitized using pad-like methods, which fails to utilize the full signal features during track reconstruction. To more accurately explore the true capabilities of pixel-readout TPCs, we incorporated realistic ionization and drift-diffusion effects into the simulation and digitization processes to obtain comprehensive signals analogous to real data as inputs for track reconstruction. We adapt a multi-stage tracking method originally developed for MDC tracking and apply it to CEPC TPC track finding. The pipeline first uses a graph neural network (GNN) to suppress noise hits. DBSCAN is then applied to the GNN-filtered hits to cluster primary track hits. Finally, RANSAC fitting is used to complement the DBSCAN results, which specially retrieves missed, broken, and poorly clustered tracks that cannot be identified by density-based clustering. In parallel, we adopt the AIDE intelligent agent framework to autonomously search, design, iterate and optimize track finding models for CEPC TPC. The search space covers supervised, unsupervised, and deep-learning strategies, and every candidate is quantitatively evaluated with customized tracking metrics, including hit clustering efficiency, hit clustering purity, and track reconstruction efficiency. Preliminary simulation results on CEPC qqH Monte Carlo samples show that both the GNN-clustering pipeline and the agent-optimized models provide promising TPC track reconstruction performance.
Finally, we outline future optimization plans for CEPC TPC tracking, including improving track finding efficiency and suppressing fake track rates, optimizing the reconstruction of low-momentum tracks, and joint reconstruction with silicon detectors. These efforts aim to further enhance tracking performance for future CEPC physics analyses.
