A modular neutrino detector years in the making

Researchers at the U.S. Department of Energy's SLAC National Accelerator Laboratory have joined collaborators from around the world to build a prototype neutrino detector that has now captured its first neutrino interactions at Fermi National Accelerator Laboratory (Fermilab). 

The prototype detector will help fine-tune a full-size version of the DUNE Near Detector Liquid Argon (ND-LAr) detector in the coming years for the international Deep Underground Neutrino Experiment (DUNE), led by Fermilab, and in the meantime help illuminate some specific neutrino properties.

Researchers will also use the detector to test advanced machine learning techniques, developed at SLAC, that are expected to play a key role in processing the vast amount of data generated by DUNE.

“The prototype is going to be very important because it’s the only source of neutrino beam data at energies comparable to the DUNE beam that will be available before DUNE is running,” said James Sinclair, a SLAC scientist working on the project. “We are excited to be completing this critical step in the experiment and are now ready to study the data that’s coming in.”

A modular design for an unusual problem

Neutrinos are fundamental particles unlike any other. They can pass through almost all matter largely unseen and can change forms along the way – a phenomenon called neutrino oscillation. Scientists think a better understanding of their unusual properties could help answer some of the most challenging questions about the origin of matter in the universe and the pattern of neutrino masses. 

To detect neutrinos, physicists use what's called a time projection chamber (TPC) – a vast tank of liquified noble gases such as argon. When a particle enters the chamber from outside, two things happen. First, interactions between the particle and argon atoms create flashes of light called scintillation. Second, the particle can knock electrons free from argon atoms, ionizing them. TPCs typically include photosensitive equipment to detect scintillation and an electric field that guides free electrons to one end of the detector, where – traditionally – a wire mesh picks them up as an electrical current. By comparing details of the flash with the time it takes electrons to arrive at the mesh, researchers can identify key details including what kinds of particles they're picking up and how fast those particles are moving.

The idea is to capture as many neutrino interactions as possible with a large volume of argon and a relatively small amount of detector equipment, almost all of which stays on the periphery of that volume. 

But something more is needed for DUNE, said SLAC scientist Hiro Tanaka, the technical director for the DUNE near detector and head of SLAC's efforts on the DUNE project. Unlike many other neutrino experiments, DUNE will produce a very large number of neutrinos and beam them in bunches toward DUNE's near detector outside Chicago. Over the course of just a few microseconds, scientists expect to see multiple neutrino interactions in the near detector. The trouble is, all those interactions make it hard to tell which flash of light belongs to which neutrino, in part because large tanks of liquid argon scatter and diffuse each individual flash. It also makes it hard to tell which electron comes from which ionization event, since any one electron takes milliseconds to reach the edge of a TPC, during which time many interactions may have occurred. 

It was out of these concerns that the newly minted prototype, called the 2x2 detector, was born. On one level, the idea is simple: Rather than use one giant TPC, break the device into a set of four TPC modules arranged in a two-by-two grid – hence the name. Each module actually contains two separate volumes of argon with an opaque wall in the center. That wall effectively creates eight optically separate TPC tanks, so that it's less likely to mistake one neutrino flash for another. It also serves as the source of the electric field that draws ionization electrons to the sides of the detector module.  

In addition, each module contains a new system for detecting ionization electrons developed at DOE's Lawrence Berkeley National Laboratory that picks up not just when the electrons arrive, but also precisely where, in contrast to the traditional wire-based designs, where the information provided by each plane of wires can be difficult to reconcile in the high interaction-rate environment of the DUNE Near Detector. Combined with the light flashes, this will help researchers determine where neutrino interactions occurred for the first time without ambiguity in three dimensions.

Scientists will also use data from the prototype to study electron neutrinos, which are one of three known neutrino types. Nearly all of the neutrinos that come out of the neutrino beam at Fermilab will be muon neutrinos, but 1 in about 1,000 will be electron neutrinos.

"DUNE needs to measure the oscillation of muon neutrinos to electron neutrinos by counting both interactions," Sinclair said. "We know that the interaction probability of electron neutrinos is different from muon neutrinos. The 2x2 will allow us to study and verify the new detector's capability to identify and study electron neutrino interactions."

Four sturdy boxes make a novel detector 

Although the module system might seem simple, it faces a practical challenge, Tanaka said: It puts a lot more stuff in the way of detecting neutrinos. It fell to longtime SLAC mechanical engineer Knut Skarpaas VIII and his colleagues to design a system that was light, sturdy, and could withstand the very cold temperatures of liquid argon. 

Skarpaas worked on many of the components for the TPC modules with collaborators at the University of Bern and Berkeley Lab. When Skarpaas first heard about the prototype detector, he walked up to a chalkboard and sketched a possible design of it. Many years later, the detector looked nearly identical to those initial drawings.

After completing the design, Skarpaas and the team focused on building the prototype’s electrostatic field cages, the boxes that contained all of the detector’s electronic components and the liquid argon. This cage defines the volume of the prototype, and everything has to fit inside of that volume.

Additionally, the team had to squeeze a high-voltage cathode, which guides those ionization electrons toward an anode, into the cage without touching any other metal parts. If metal touched the cathode, this could create an electrical arc, jeopardizing the detector equipment.

Perhaps the most difficult part of the building process was selecting the right power cable. The cable feeds electricity to the high-voltage cathode and makes the whole detector work, and it needs to be straight, cannot touch any other parts and must be able to shrink up to two inches due to the cold temperature inside of the detector. If the cable bends under these cold temperatures, it could shatter.

After many long days inside of a machine shop at SLAC, Skarpaas and the team finished assembly, and shipped the modules to the University of Bern for testing.

“Putting all of a detector’s pieces together is like being the conductor of an orchestra,” Skarpaas said. “You have to understand what everyone needs for their science goals and then meld these needs together to build the detector.”

Advanced machine learning techniques

DUNE's primary goal is to explore some of the deepest questions about the composition of the universe by studying neutrino properties. To do this, researchers need to not only capture neutrino interactions, but also make sense of the data generated by these interactions.

In the case of the prototype detector, the data generated by up to thousands of neutrino interactions per day would be impossible for scientists to study manually image by image. Researchers therefore invented new machine learning techniques for this amount of data. Machine learning is a type of artificial intelligence that detects patterns in large datasets, then uses those patterns to make predictions and improve future rounds of analysis.

“By eye, it might be easy to find the information you need in a single image generated by the detector,” SLAC researcher Francois Drielsma said. “But it is difficult to teach a machine to perform this task. Sometimes there is the thought that if something is simple for a human being, it should be simple for a machine. But that is not necessarily true.”

Still, humans aren't up to scanning millions of images at a time. They've also struggled to use traditional programming techniques to help identify objects in detector data, so Drielsma’s group started working on a machine learning technique called neural networks, a type of algorithm loosely modeled after the human brain. Once a neural network is trained on a large set of data – whether from particle interactions or astronomical images – it can automatically analyze other complex datasets, almost instantaneously and with great precision.

The program is working better each day, and researchers will continue to fine-tune its performance over the coming years while the prototype detector is collecting data.

“It’s going to be a difficult task to train the program to do everything we want accurately,” Drielsma said. “But when things are difficult, they can be really entertaining.”

This research was supported in part by the DOE Office of Science.

For questions or comments, contact SLAC Strategic Communications & External Affairs at communications@slac.stanford.edu.