Using Binding Energies to Gain Insights on Machine Learning

Ian Bentley, Florida Polytechnic University

Thu, Mar 19, 2026, 14:00

Auditorium / Zoom

Abstract

Nuclear binding energies provide an excellent testing ground for developing machine learning regression models. The moderately sized data set of about 2500 accurately measured values and the well studied behavior allow for training and test sets to be set up and appropriate physical features to be utilized. Neural networks, kernel-based approaches, such as support vector machines, and gaussian process regression, and even ensembled tree-based approaches have seemingly matched and potentially surpassed the accuracy of prior models. This talk will discuss observed behavior among these approaches when used to model binding energy residuals. A composite model which predicts the Atomic Mass Evaluation 2020 for N>7, and Z>7 with a standard deviation of less 100 keV will also be discussed.

Zoom: https://uvic.zoom.us/j/83875570623?pwd=kQKlwDCfScrIu0J5TunWqnho0jNRg8.1

Meeting ID: 838 7557 0623 Password: TRColloq