Controlling extrapolations of nuclear properties with feature selection

TL;DR

This paper introduces a feature selection method to improve the robustness of machine learning models predicting nuclear properties, specifically nuclear binding energies, by identifying input variables that reduce model bias and enhance predictive accuracy.

Contribution

The paper presents a novel feature selection technique based on probability distribution functions to control and reduce model bias in nuclear property predictions, especially for extrapolations.

Findings

Feature selection improves the accuracy of nuclear binding energy predictions.

The method systematically reduces model bias without increasing uncertainty.

Application to DFT demonstrates enhanced predictive power across the nuclear chart.

Abstract

Predictions of nuclear properties far from measured data are inherently imprecise because of uncertainties in our knowledge of nuclear forces and in our treatment of quantum many-body effects in strongly-interacting systems. While the model bias can be directly calculated when experimental data is available, only an estimate can be made in the absence of such measurements. Current approaches to compute the estimated bias quickly lose predictive power when input variables such as proton or neutron number are extrapolated, resulting in uncontrolled uncertainties in applications such as nucleosynthesis simulations. In this letter, we present a novel technique to identify the input variables of machine learning algorithms that can provide robust estimates of model bias. Our process is based on selecting input variables, or features, based on their probability distribution functions across the entire nuclear chart. We illustrate our approach on the problem of quantifying the model bias in nuclear binding energies calculated with Density Functional Theory (DFT). We show that feature selection can systematically improve theoretical predictions without increasing uncertainties.