Comparative Study of Feedforward Neural Network and NGBoost for Proton Separation Energy Predictions

Abstract

Accurate prediction of nuclear proton separation energies is very important to understand the fundamental nuclear structure and rapid proton capture (rp-process) astrophysical simulations. While standard global mass models are built on powerful theoretical bases, the complex interaction between the macroscopic liquid-drop behavior and discrete quantum shell effects is still a major computational challenge. Moreover, the ability of such highly localized phenomena at the microscopic level by continuous or discrete machine learning (ML) structures requires a strict comparative analysis. In this research, we quantitatively evaluate two different predictive models of a continuous-mapping Feedforward Neural Network (FNN) and a tree based piecewise-discrete Natural Gradient Boosting (NGBoost) algorithm. Both architectures were trained using comprehensive NUBASE2020 data of 3378 empirically assessed nuclides, using a physics-informed feature space, which explicitly captures macroscopic, microscopic, and non-linear correlational effects. Both algorithms were able to map the overall nuclear dynamics with a high score of over 0.98 on R 2 . The FNN showed that it can generalize well and effectively learn smooth macroscopic trends. However, NGBoost achieved higher predictive precision, resulting in a significant reduction of test error when compared to the continuous model. Our analysis provides evidence that continuous FNN structures tend to be better at capturing high-level structural trends, while discrete additive tree NGBoost structures are better suited to capturing extreme, non-linear microscopic discontinuities. As a result, NGBoost develops a calculation system that is widely accurate and strong to calibrate the predictions of nuclear masses and analyze the characteristics of carried out exotic nuclei.