A breakthrough in nuclear physics at Florida Polytechnic University has created an advanced machine learning model that predicts nuclear binding energies with unprecedented accuracy, helping scientists better understand the building blocks of matter.
Dr. Ian Bentley, professor and chair of the University’s Department of Physics, developed the technique, called the Four Model Tree Ensemble, which combines several machine learning models.
He presented this innovation recently at an international astrophysics conference in Germany, where the model’s prediction precision was discussed. He also published two papers on the subject this year in the journal Physical Review C.
“When this discovery happened, I was in disbelief. I was doubtful and began looking around to see if there were better-performing existing models,” Bentley said. “There has been a lot of work applying machine learning to the problem, but our new approach seems to be the best at predicting recent nuclear mass measurements.”
Nuclear masses reflect the amount of mass contained in an atom’s nucleus. They are essential for understanding how elements form in the universe.
Bentley said that while scientists understand how the sun creates lighter elements like helium, carbon, and oxygen through thermonuclear fusion, the processes that produce heavier elements such as gold, lead and uranium are an active area of research.
“The idea is that extreme astrophysical environments like supernovas and neutron star-neutron star mergers form these, but we need accurate input data to be able to simulate the events,” he said. “There are also groups conducting experiments to create new nuclei and they want to know what mass values to expect.”
Bentley’s work bridges the theoretical and experimental efforts.
Existing methods have produced strong results by using neural networks or kernel-based machine learning methods in their modeling. In contrast, his approach utilizes an advanced algorithm that combines many decision trees to model the data, resulting in much greater accuracy.
“Four years ago, no one was approaching it this way,” he said.
In addition to his participation in the Germany conference, Bentley has been invited to speak at two national laboratories.
Plans for additional research in the area are already underway with a research team that includes undergraduate students James Tedder and Anthony Fiorito.
“Anthony is going to teach me how to apply physics-informed machine learning more effectively,” Bentley said. “It’s exciting – they’ll learn the physics from me, and I’ll learn machine learning from them.”
Contact:
Lydia Guzmán
Director of Communications
863-874-8557
