Jilin University Researchers Launch Physics-Informed Deep Learning Model For Global High-Precision Nuclear Charge Density Prediction

Jilin University’s new deep learning model predicts nuclear charge density with 50% more accuracy than traditional theories, aiding fundamental physics research.

By: AXL Media

Published: Mar 21, 2026, 7:09 AM EDT

Source: Information for this report was sourced from Nuclear Science and Techniques

A Paradigm Shift In Nuclear Structure Research

The study of nuclear charge density, a fundamental property for understanding the internal structure of an atom, has long been hampered by the limitations of traditional theoretical models. While frameworks like density functional theory have provided the basis for nuclear physics for decades, they often struggle with predictive precision. Researchers at Jilin University have addressed this challenge by pioneering a "physics-informed" deep learning paradigm. This model shifts the focus from purely theoretical calculations to a data-driven approach that successfully integrates experimental charge radius data, marking a decisive evolution in how scientists map the distribution of protons within a nucleus.

The Mechanics Of Physics Informed Training

The team’s innovative approach utilizes a two-stage optimization process to ensure the neural network remains grounded in physical reality. Initially, the DNN was trained to predict Fourier-Bessel coefficients based on the Relativistic Continuum Hartree-Bogoliubov (RCHB) theory. In the second stage, the model was fine-tuned using experimental data from 1,014 different nuclei. By using inputs such as proton and neutron numbers, proximity to "magic numbers," and pairing parameters, the model outputs 17 specific coefficients. This dual-layered strategy allows the AI to provide a unified and highly precise description of both charge density and charge radius simultaneously.

Validating Accuracy Against Established Theories

To test the model's performance, the researchers applied the DNN to isotopes of nickel, palladium, mercury, and bismuth. The results showed a root-mean-square error of only 0.0149 fm, a significant improvement over previous deep learning attempts and the traditional RCHB theory. The model proved particularly adept at predicting the central density and the "tail" structures of specific nuclei like chromium and zinc. When compared to experimental values, the DNN’s deviations were heavily concentrated within the 0–0.01 fm range, whereas traditional theoretical calculations remained far more dispersed and less reliable.