Algorithms for the machine-learning era

Dorota Stefan (CERN) and Robert Sulej (Fermilab and NCBJ Warsaw) are actively pushing algorithm development for reconstruction towards the machine-learning era. LArTPC detectors produce image-like projections of particle trajectories, but due to very large sizes of these images, current reconstruction approaches have only been able to process strongly reduced representations of the data.

Size is not the only challenge facing LArTPC reconstruction. It turns out that even when a complete LArTPC image is converted into a set of points identified on particle traces, important features easily visible to the human eye are very hard for algorithms to recognize. Exploration of the full potential of LArTPC detectors, therefore, requires cutting-edge computing technologies, such as deep neural networks.

“The standard reconstruction tools perform much better when complemented with the modern computing vision techniques,” said Sulej. “CERN and Fermilab are continuing to develop algorithms that work with these newer techniques.”

Stefan and Sulej collaborate with OpenLab and the TechLab project at CERN to help build the large, specialized computing resources required for DUNE’s research.

A 2D and 3D view of interactions of two cosmic rays and a beam particle. Courtesy: D. Stefan and R.Sulej


“We are building up the participation of European scientists in ProtoDUNE software and analysis activities, overseeing and supporting those who are fresh to the challenges of LArTPCs,” said Stefan. “We aim to join together efforts from both the single- and dual-phase LArTPC projects into ProtoDUNE data analysis.”

Leigh Whitehead, Photo: Warwick University press

Leigh Whitehead, who joined CERN last October, is using reconstruction algorithms developed by Stefan and Sulej to study ProtoDUNE’s capability to distinguish beam events from the large number of cosmic-ray muon interactions, and to see how the reconstruction of the beam interactions is affected by the presence of the cosmic rays.

In parallel, he is studying how well ProtoDUNE can expect to know the particle content of the beam entering into the detector. He has written an interface between beam simulations (provided by the beam group at CERN) and LArSoft in order to characterize the detector response to known particles. The output files provide fodder for the algorithms he runs for his reconstruction studies.

DUNE’s pattern-recognition efforts and proton decay searches are also going to use these modern algorithms.

If progress continues apace, collaborators should expect to reconstruct detector events using machine-learning techniques combined with conventional approaches once ProtoDUNE starts taking data.