Utilizing Machine Learning Techniques to Predict Binding Energies

The best theoretical nuclear binding energies can typically reproduce Atomic Mass Evaluation (AME) experimental values for N>7 and Z>7, with standard deviations ranging from 300-500 keV. Recently, the use of machine learning techniques informed by a variety of physical features as model inputs, have been able to achieve fits near this range. We have included four different machine learning approaches (Neural Networks, Gaussian Process Regression,Support Vector Machines, and Ensemble of Trees). The last approach has produced some interesting results comparable to the other more common techniques. Our methodology involves training these machine learning models on the difference between semi-empirical fits and the experimental data. This has reduced the need for a dozen or so features to be used, and fits on the order of 200 keV have been achieved for both a training set using AME 2012 and for an independent testing set using AME 2020 values for different nuclei.