A lot can happen in the blink of an eye. In a laboratory setting, it takes the average person to see the light and press the button in about a fifth of a second, and the same interval allows a hummingbird to flap its wings about 12 times. Meanwhile, in that same split second, specialized computer hardware that analyzes particle collider data can use artificial intelligence to make more than 10 million decisions about whether to keep or discard the information created by the collision event.
“Neural network algorithms allow us to gain deeper, more efficient insights into our data and make discoveries much more quickly than with traditional, simpler techniques.”
Nhan Tran, Director of AI Coordination Office, Fermilab
At the U.S. Department of Energy’s Fermi National Accelerator Laboratory, researchers are leading an open-source collaboration that pushes the limits of machine capabilities and embeds neural networks directly into physical hardware in the form of efficient, customized digital circuits. At the heart of this effort is hls4ml, a software framework developed with the expertise of Fermilab researchers. Hls4ml allows you to create ultra-fast decision-making hardware for applications ranging from particle physics to fusion science and beyond.
Humanity’s most ambitious scientific projects, many led by or supported by Fermilab researchers, generate staggering amounts of data. Particle collider detectors such as the CMS at CERN’s Large Hadron Collider investigate the universe at its most fundamental level. Fermilab is the US host laboratory that facilitates the participation of hundreds of US physicists from more than 50 institutions in the CMS experiment at CERN.
“The high-luminosity LHC’s CMS upgrade will generate nearly six times as much data once it becomes operational in the 2030s,” said Anadi Canepa, senior scientist at Fermilab and spokesperson for the international CMS collaboration. “Our updated trigger system gives us access to more detailed information, expanded coverage, and expanded timing information. The challenge is that if we want to analyze all of this additional data, we need to do it quickly.”
Neural networks (algorithms inspired by the way the human brain processes information) learn by passing data through interconnected layers and adjusting connections to recognize patterns and make predictions. But learning alone is not enough. To deliver real-world value, these networks must also be efficiently deployed.
“Neural network algorithms can help us gain deeper insights into data more efficiently and make discoveries much faster than traditional, simpler techniques,” said Nian Tran, director of Fermilab’s AI Coordination Office.
After a network is modeled and trained, researchers need a clear path to accelerating it in hardware. That’s where hls4ml comes in.
“Hls4ml takes the code for neural networks that can be written in open source machine learning libraries like PyTorch or TensorFlow and essentially turns them into a set of logic gates,” Tran explained.
Traditionally, machine learning algorithms have been run using the central processing unit (commonly referred to as the CPU) and graphics processing unit (GPU) found in laptops and desktop computers.
“As these techniques become more widely used, it’s natural to wonder if there are more efficient approaches,” said Giuseppe Di Guglielmo, principal engineer at Fermilab.
By moving neural networks to specialized hardware, such as field-programmable gate arrays and application-specific integrated circuits, researchers can perform many calculations at once and make decisions faster while using less power.
“Although the program is more complex, it allows us to run advanced algorithms in real time when latency and power are important,” Di Guglielmo added.
Programming these devices has traditionally required deep expertise. However, hls4ml makes it possible for a wider range of researchers to prepare decision-making hardware for particle detector triggers.
“The hls4ml team is making triggers easier to use,” says Canepa. “Anyone with a new idea can now create and run triggering algorithms. Hls4ml is absolutely essential for a successful CMS upgrade, as we will be collecting an unprecedented, extremely large and complex data set. Without a capable triggering system to select events, we will not be able to save the most interesting collisions.”
“Many fields of cutting-edge science face the challenge of big data and explore the nature of the universe on very short timescales, so research communities from fusion energy to neuroscience to materials science are very interested in what we are doing to enable new capabilities through the power of AI,” Tran added.
Fermi National Accelerator Laboratory is America’s premier national laboratory for particle physics and accelerator research. Fermi Forward Discovery Group manages Fermilab for the U.S. Department of Energy’s Office of Science. Visit Fermilab’s website. www.fnal.gov Follow us on social media.
