In a groundbreaking advance at the crossroads of nuclear physics and artificial intelligence, researchers from Jilin University have unveiled a revolutionary deep learning-based model that dramatically enhances the precision of nuclear charge density predictions. Capturing the intricate distribution of charge within atomic nuclei has long posed a formidable challenge due to experimental limitations and the inherent complexity of nuclear forces. However, this latest study leverages a sophisticated deep neural network (DNN), intricately informed by physics, to transcend traditional theoretical models and deliver a quantum leap in predictive accuracy.
The atomic nucleus’s charge density profile is paramount to comprehending nuclear structure and the subtleties of nuclear forces. Conventional approaches predominantly hinge upon density functional theory and mean-field models such as the relativistic continuum Hartree–Bogoliubov (RCHB) framework. While these methodologies have delivered valuable insights, their predictive fidelity is often constrained, particularly when attempting a global and systematic description spanning thousands of nuclei. The Jilin University team’s innovative approach channels the power of data-driven machine learning while embedment of physical laws, forging an unprecedented synergy between empirical evidence and theoretical sophistication.
Central to their methodology is a two-stage “physics-informed” training paradigm. Initially, the DNN is trained on theoretical data generated by RCHB calculations, specifically targeting the prediction of Fourier–Bessel (FB) coefficients, which serve as comprehensive descriptors of nuclear charge density distributions. In the refining phase, the model undergoes fine-tuning against an extensive dataset of experimental charge radii encompassing over one thousand nuclides. This dual-step optimization not only anchors the DNN in rigorous physical theory but also calibrates it robustly against real-world measurements.
The model’s architecture accepts fundamental nuclear parameters as inputs: proton number (Z), neutron number (N), proximity to nuclear magic numbers, and pairing effects among nucleons. From these, it outputs 17 Fourier–Bessel coefficients that intricately reconstruct the radial charge density profile with exceptional precision. This design elegantly encapsulates both the global structural features and subtle shell effects within charge distributions, enabling a unified model that simultaneously predicts charge radii and detailed density profiles across a diverse nuclear landscape.
Validation results underscore the transformative impact of this approach. When tested on representative isotopic chains such as nickel, palladium, mercury, and bismuth, the DNN predictions align remarkably well with experimental data, boasting a root-mean-square deviation (RMSD) as low as 0.0149 femtometers. By contrast, traditional RCHB theory manifests a considerably broader scatter in deviation, evidencing the model’s superior accuracy and consistency. Furthermore, in regions of nuclear matter where experimental data is sparse or uncertain, the DNN demonstrates robust extrapolative capabilities, significantly enhancing confidence in its predictive power.
Beyond mere radii, the model effectively reconstructs detailed density profiles, capturing features such as central density depressions and the tail behavior at large radii with heightened fidelity compared to earlier theoretical models. These improvements are critical for understanding nuclear surface properties, neutron skins, and the interplay of nuclear forces at extreme isospin values. The capacity to predict such subtle density variations holds promise for deepening insights into the nuclear equation of state, a fundamental descriptor with ramifications spanning astrophysics and particle physics.
The implications of this research ripple far beyond nuclear theory. High-precision nuclear charge distributions serve as foundational inputs for atomic physics, refining theoretical calculations of atomic spectra and transitions. Moreover, they play an instrumental role in calibrating fundamental constants and testing the limits of quantum electrodynamics under strong-field regimes. In astrophysics, accurate nuclear densities inform reaction networks essential for simulating nucleosynthesis processes occurring in stellar explosions and neutron star mergers, thereby enriching our comprehension of elemental abundances in the cosmos.
This pioneering study exemplifies the paradigm shift from purely analytical modeling toward a hybrid framework where machine learning is seamlessly integrated with domain-specific physics. Such “physics-informed” artificial intelligence frameworks hold the promise to revolutionize not only nuclear physics but also other disciplines confronted with complex, multiscale problems that defy traditional treatment. The Jilin University team’s success signals an era where empirical data, theory, and AI converge to push the frontiers of scientific understanding.
Looking forward, the research group envisions broadening the model’s applicability across the entire nuclear chart, including lighter and highly exotic nuclei that challenge both experiment and theory. Incorporation of additional experimental observables, such as charge form factors and transition probabilities, is also anticipated to refine model robustness. Parallel advancements in neural network architectures and training algorithms will further enhance performance, potentially catalyzing new discoveries and applications in fundamental and applied nuclear science.
Professor Jian Li, leading the research, reflects on the broader significance: “Our work demonstrates that deeply embedding physical principles within machine learning architectures yields predictive models that surpass traditional methods in accuracy and reliability. The integration of experimental data into this framework not only advances nuclear structure modeling but also establishes a robust data foundation for related fields including atomic physics and fundamental physics. This is a transformative step toward intelligent scientific prediction.”
As artificial intelligence methodology continues to mature, this research heralds a new era where the intricate microcosm of the atomic nucleus can be elucidated with unparalleled clarity and confidence. The proficiency exhibited by the physics-informed DNN approach promises to unlock profound insights into nuclear phenomena, inform experimental design, and stimulate innovation across scientific domains reliant on precise nuclear data.
This novel fusion of deep learning with nuclear theory could reshape the landscape of nuclear physics, offering an intelligent predictive tool that bridges the gap between complex theoretical models and the wealth of experimental knowledge. The ongoing refinement and expansion of such models will likely spearhead interdisciplinary advances, illuminating the structure of matter and underpinning transformative technologies grounded in nuclear science.
For the detailed study and further technical exposition, refer to the original publication available via DOI: 10.1007/s41365-026-01905-6.
Subject of Research: Not explicitly specified in the text beyond nuclear structure physics.
Article Title: Predictions of charge density distributions for nuclei with Z ≥ 8
News Publication Date: 19-Mar-2026
Web References: http://dx.doi.org/10.1007/s41365-026-01905-6
Image Credits: Jian Li
Keywords
Nuclear charge density, deep neural network, physics-informed machine learning, nuclear structure, Fourier–Bessel coefficients, Relativistic Continuum Hartree–Bogoliubov theory, charge radius prediction, data-driven nuclear modeling, computational nuclear physics, nuclear equation of state, atomic physics applications, fundamental constants, quantum electrodynamics
