ÌÇÐÄÊÓƵ

It's elementary: Problem-solving AI approach tackles inverse problems used in nuclear physics and beyond

Solving life's great mysteries often requires detective work, using observed outcomes to determine their cause. For instance, nuclear physicists at the U.S. Department of Energy's Thomas Jefferson National Accelerator Facility analyze the aftermath of particle interactions to understand the structure of the atomic nucleus.

This type of subatomic sleuthing is known as the inverse problem. It is the opposite of a forward problem, where causes are used to calculate the effects. Inverse problems arise in many descriptions of physical phenomena, and often their solution is limited by the experimental data available.

That's why scientists at Jefferson Lab and DOE's Argonne National Laboratory, as part of the QuantOm Collaboration, have led the development of an artificial intelligence (AI) technique that can reliably solve these types of puzzles on supercomputers at large scales.

"We set out to prove we could use generative AI to better understand the structure of the proton," said Jefferson Lab Data Scientist Daniel Lersch, a lead investigator on the study. "But this framework isn't bound to nuclear physics. Inverse problems can be anything."

The system is called SAGIPS (Scalable Asynchronous Generative Inverse-Problem Solver). It relies on high-performance computing and generative AI models, which can produce new text, images or videos based on data the algorithms are trained on.

SAGIPS was built for QuantOm. Its goal is to better understand fundamental nuclear physics by using advanced computational methods, and the SAGIPS system was recently in the journal Machine Learning: Science and Technology.

The problem

Inverse problems can be found in most areas of science, from astrophysics to chemistry to medical imaging. The process can be likened to reverse engineering, said Nobuo Sato, a Jefferson Lab theoretical physicist and author on the paper.

"Imagine throwing a ball into a dark hole," Sato said. "If the ball bounces back in a particular pattern, you can play around with different directions and in principle infer what kind of surface is inside."

In the SAGIPS study, the ball is an electron. It's part of a "toy" nuclear physics problem based on inclusive deep inelastic scattering, in which an electron is measured after interacting with another particle.

But the math behind inverse problems can be a little fuzzy. Solutions are represented as probabilities instead of concrete answers. Using a problem-solver like SAGIPS can add clarity and definition to those probabilities, reducing uncertainties and bringing scientists closer to an answer.

SAGIPS ran a machine learning (ML) algorithm on the Polaris supercomputer cluster at the Argonne Leadership Computing Facility, a DOE Office of Science user facility in the Advanced Scientific Computing Research (ASCR) portfolio. Using 400 processing cores, SAGIPS solved the toy problem and showed promise for solving larger problems on an even more powerful supercomputer.

"This technique scales linearly with the available computing resources, which means we could process much bigger problems on a much bigger cluster," said Malachi Schram, Jefferson Lab's head of data science and a co-author on the paper. "That's the heart of it."

How it works

SAGIPS uses generative adversarial networks (GANs), which are dueling neural networks that interact to produce meaningful data. One is constantly trying to trick the other, and the other is trying to spot fakes.

"Let's say you want an image of a cat," Lersch said. "The generator doesn't know what a cat looks like, so it takes a guess. The discriminator has access to real cat pictures and can communicate to the generator that it produced a fake. The generator will then attempt to draw a more realistic picture of a cat."

This back and forth continues until a realistic image is produced.

The GANs are hosted on graphics processing units (GPUs). SAGIPS groups GPUs in a ring-like fashion with remote memory access, allowing the GPUs to directly access each other's data and share their own. This reduces communication bottlenecks and speeds up processing. SAGIPS exploits Gustafson's law, running faster as more computer resources are available.

"Imagine that the processing of data on one machine is slower because of the statistical variations in the workflow," Schram said. "They show up maybe seconds behind. If this happens at scale, these delays become horrendous. With an asynchronous ring-all-reduce approach, the GPUs don't have to wait."

Finding solutions

SAGIPS was supported by SciDAC through the QuantOm project (Quantum Chromodynamics Nuclear Tomography). QuantOm is a DOE Office of Science partnership between ASCR and the Office of Nuclear ÌÇÐÄÊÓƵics, and it is interested in the 3D imaging of protons, neutrons and nuclei by running particle accelerator data on supercomputers. The goal is to support current and future DOE Office Science user facilities, such as the Continuous Electron Beam Accelerator Facility (CEBAF) at Jefferson Lab and the future Electron-Ion Collider at Brookhaven National Laboratory.

"It's very easy to use some of these algorithms for different projects, and this minimizes the overhead costs of redesigning and rebuilding software stacks," Schram said. "Jefferson Lab's Data Science Department is focused on building capabilities, not one-off solutions."

Now that the concept is proven on Argonne's Polaris cluster, the SAGIPS team wants to leverage it on the exascale (1 quintillion flops, or 1 exaflop). Another computer at Argonne, named Aurora, is the first in the world to reach that speed.

"It's fascinating that bridging the gap between experimentalists and theorists includes another experiment in and of itself," Sato said. "And that experiment is called high-performance computing."