Speaker
Description
The Deep Underground Neutrino Experiment (DUNE) employs Liquid Argon Time Projection Chambers (LArTPCs) as its primary detection technology, taking advantage of their high spatial and calorimetric resolution. To extract ionization charge information from raw waveforms and to reconstruct events of interest, the Wire-Cell Toolkit (WCT) is under active development. However, when particle tracks are oriented nearly perpendicular to the readout electronics, signal extraction becomes challenging due to limited charge sharing and insufficient capture of the Region of Interest (ROI).
In this study, we investigate advanced machine learning models to enhance signal reconstruction in these difficult event topologies within the WCT framework. This approach shows improved performance compared to conventional ROI-based methods. However, inference using deep neural networks leads to a significant increase in memory consumption. To mitigate this, we employ profiling techniques to identify memory bottlenecks, and we are currently optimizing key components through methods such as parameter pruning and restructuring of the data processing flow.
