Machine Learning the ground state masses of atomic nuclei

Triangle Nuclear Theory seminar

Invited presentation on 03/2023

Machine Learning and Artificial Intelligence methods offer wide applicability to a class of problems known collectively as "inverse problems." The solution to such problems involves the calculation of the causal factors that are responsible for a set of observations. Inverse problems are frequently encountered in nuclear physics, where measurements exist and these observations must be reconciled with theoretical models and interpretation. I will discuss a probabilistic machine learning technique applied to the binding energy of atomic nuclei. The set of observations comes from the Atomic Mass Evaluation (AME), which totals over 2000 data points. I show that inclusion of physics-based inputs as well as physics-informed training yields neural networks that are capable of describing these observations with a high degree of precision. Because the method is stochastic, it also provides an estimate of uncertainty for each prediction. I discuss the capacity of such modeling to extrapolate into regions of unmeasured nuclei that is needed for applications such as astrophysics.

Contact Me


Matthew Mumpower
Los Alamos National Lab
MS B283
TA-3 Bldg 123
Los Alamos, NM 87545

Office Phone

(505) 667-5671