The origin of heavy elements in the universe is theoretically the result of neutron star collisions. The high temperature and high density conditions generated by neutron star collisions are enough to allow free neutrons to merge with atomic nuclei and form new elements in an instant. Testing this theory and answering other astrophysical questions requires predictions over the vast range of masses of atomic nuclei.Scientists at Los Alamos National Laboratory were the first to use machine learning algorithms, an application of artificial intelligence, to successfully build an atomic mass model of the entire nuclide map -- all possible combinations of protons and neutrons that define an element and its isotopes.
Theoretical physicist Matthew Mumpower said: "There may be thousands of nuclei in nature that have not yet been measured. Machine learning algorithms are very powerful because they can find complex correlations in the data that theoretical nuclear physics models struggle to produce efficiently. These correlations can provide scientists with information about 'missing physics', which in turn can be used to strengthen modern nuclear models of atomic masses."
Simulate the rapid neutron capture process
Recently, Mumpower and colleagues (including former Los Alamos summer student Mengke Li and postdoc Trevor Sprouse) published a paper in Physics Letters B describing the use of a physics-based machine learning mass model to simulate an important astrophysical process. The r process, the rapid neutron capture process, is an astrophysical process that occurs in extreme environments, such as the environment created by neutron star collisions.
Heavy elements may come from this "nucleosynthesis". In fact, half of the heavy isotopes in the universe down to bismuth and all of the thorium and uranium were probably produced by this "nucleosynthesis" process.
However, modeling this process requires theoretical predictions of atomic masses, which current experiments cannot achieve. The research team used physical information machine learning methods to randomly select from the large mass database of Atomic Mass Evaluation (AtomicMassEvaluation) and train a model. Next, the researchers used the masses of these predictions to simulate the r process. The model allowed the research team to simulate R-process nucleosynthesis for the first time with the quality of machine learning predictions - a major first since machine learning predictions often collapse on extrapolation.
"We have shown that machine learning of atomic masses can open the door to predictions beyond experimental data," Mumpower said. "The key point is that we tell the model to obey the laws of physics. By doing this, we are able to make physics-based inferences. Our results are on par with or better than contemporary theoretical models, and can be updated immediately as new data becomes available."
Study nuclear structure
The r-process simulation complements the research team's application of machine learning to studies related to nuclear structure. In a recent article published in Physical Review C that was selected as an "Editor's Suggestion," the team used machine learning algorithms to reproduce nuclear binding energies with quantitative uncertainty; that is, they were able to determine the energy required to separate an atomic nucleus into protons and neutrons, as well as the associated error bars for each prediction. Therefore, this algorithm provides information that would require significant computational time and resources to obtain from current nuclear modeling.
In related work, the research team used their machine learning model to combine precise experimental data with theoretical knowledge. These results inspired some of the first experimental activities at the new Rare Isotope Beam Facility, which aims to expand the known area of the nuclear map and reveal the origins of heavy elements.
Compiled from:ScitechDaily