Hierarchical Neural Filtering of Nuclear Mass Residuals and Spectral Signatures of Quantum Chaos

Abstract

In complex quantum many-body systems such as atomic nuclei, the interplay between regular collective motion and irregular intrinsic dynamics gives rise to fluctuations that cannot be fully captured by existing global theoretical models. Nuclear mass, which exhibits smooth trends across the nuclear chart together with localized deviations, provides a sensitive observable for investigating such irregular dynamics. In this work, we employ a variety of neural network architectures, which serve as controlled nonlinear filters within a Hierarchical Residual Decomposition framework to progressively extract and suppress the chaotic many-body signature (characterized by $1/f$ spectral correlations) in nuclear mass residuals. The resulting Physics-Informed Neural Ensemble (PINE) model combines multiple mass models and neural network architectures, enabling a systematic suppression of coherent and chaotic components, after which the remaining fluctuations are analyzed using Fourier-based spectral diagnostics across different mass regions. Our results show that hierarchical neural residual learning efficiently removes the dominant low-frequency correlations and suppresses the quantum-chaotic spectral rigidity, driving the residuals toward the uncorrelated white-noise limit. This systematic suppression provides a quantitative diagnostic of the underlying scale-dependent complexity and many-body correlation structure of nuclear mass deviations.