Machine Learning Tools Are Helping Scientists Analyze Neutrino Problems More Efficiently

Scientists are using a new machine learning tool called SCNNs to analyze data faster and more efficiently than ever before. The breakthrough marks a paradigm shift in the industry as AI becomes more useful in real-world applications.

Physicists spend a lot of time looking at nothing when trying to solve problems about the smallest particles in the world. This is thanks to “sparse” data, or the areas between matter. In a rectangular photo of the Eiffel Tower, for instance, just 0.02% would feature the landmark’s iron. The rest would be taken up by empty space.

For algorithms tasked with analyzing sparse images or data sets, looking at blank space is a waste of time and chews into valuable computing resources. Crunching meaningless numbers for sparse areas takes up just as much computing power as calculating the data that yields a result. That’s why researchers at the forefront of physics are turning to artificial intelligence (AI) and a new type of machine learning tool called sparse convolutional neural networks (SCNNs).

This marks a monumental shift in the industry as computer science advancements begin leading the way in data analysis in physics. Researchers hope this will make analyzing complex data sets more efficient to pave the way for new discoveries in the field.

A Long Road to SCNNs

The idea for SCNNs first emerged in 2012 when Benjamin Graham, a University of Warwick student, developed a neural network to decipher Chinese handwriting. His development altered the time’s leading convolutional neural network (CNN) technology to focus on areas where at least one pixel had a value—thus ignoring much of the blank space within an image. Graham’s solution achieved a 2.61% error rate before he moved on to 3D-object recognition. He went on to Facebook’s AI Research department and in 2017 released early details of the world’s first SCNN.

Now, Kazuhiro Terao, a physicist from the SLAC National Accelerator Laboratory in California, is bringing the concept to his field and his work at the Fermi National Accelerator Laboratory. Teraro’s projects are vast, but his experiments involve neutrinos. The latter are some of the most elusive, albeit abundant, mass-containing particles in the universe.

Even so, researchers have trouble finding neutrinos inside particle detectors. That’s because neutrino interactions inside detectors generate sparse data.

Using traditional methods to find them means sorting through massive amounts of dead space. Terao first applied SCNNs to the task in 2019, using the tech to analyze data from the Deep Underground Neutrino Experiment (DUNE). DUNE is scheduled to come online in 2026 and will instantly become the world’s largest neutrino physics experiment.

Neutrinos will be shot from Chicago’s Fermilab to an underground laboratory some 800 miles away in South Dakota. Oscillations of the particles along their journey could give scientists a better understanding of some of their key properties. SCNNs are expected to play a critical role in analyzing the data gathered from the experiment.

“With a sparse CNN, we analyze the entire image at once—and do it much faster,” Terano says.

This lets the researchers know which data is worth investigating and which data should be discarded. All without taking the lengthy amount of time it used to.

New Paradigm

Physics experiments demand an immense amount of computational resources. Finding ways to analyze data more efficiently is a top priority as both grant funding and time are limited. To optimize experiments, physicists have historically designed their own algorithms for analyzing data. This gives researchers precise control and efficacy for their specific research. Unsurprisingly, though, it takes time away from research to develop these algorithms.

With SCNNs becoming more powerful and more effective, the need for physicists to write their own algorithms for sparse data environments is lessened.

Harvard University physicist Carlos Arguelles-Delgado said in an interview, “In physics, we are used to developing our own algorithms and computational approaches. We have always been on the forefront of development, but now, on the computational end of things, computer science is often leading the way.”

Indeed, this sentiment is true not just of physics, but every industry as artificial intelligence becomes more capable. With generative AI applications like ChatGPT able to produce high-quality work and expertly analyze data, leaders in every sector are taking notice. While SCNNs are likely relegated to the realm of physics for now, it will be interesting to see how AI innovations continue to shape the way work is done and discoveries are made.

Author