High Precision Binding Energies from Physics Informed Machine Learning

TL;DR

This paper develops physics-informed machine learning models to accurately predict nuclear binding energy residuals, achieving high precision and demonstrating strong extrapolation and prediction capabilities.

Contribution

The paper introduces a machine learning approach that combines physical features and ensemble methods to improve binding energy predictions beyond existing mass models.

Findings

Achieved a standard deviation of 17 keV on training data.

Attained a standard deviation of 92 keV on AME 2020 data.

Demonstrated effective extrapolation and prediction of new mass measurements.

Abstract

Twelve physics-informed machine learning models have been trained to model binding energy residuals. Our approach begins with determining the difference between measured experimental binding energies and three different mass models. Then four machine learning approaches are used to train on each energy difference. The most successful ML technique, both in interpolation and extrapolation, is the least squares boosted ensemble of trees. The best model resulting from that technique utilizes eight physical features to model the difference between experimental atomic binding energy values in AME 2012 and the Duflo Zuker mass model. This resulted in a model that fit the training data with a standard deviation of 17 keV and that has a standard deviation of 92 keV when compared all of the values in the AME 2020. The extrapolation capability of each model is discussed, and the accuracy of predicting new mass measurements has also been tested.