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

This paper introduces a Physics-Informed Neural Ensemble (PINE) model utilizing a Hierarchical Residual Decomposition framework to systematically filter chaotic many-body signatures from nuclear mass residuals, demonstrating that hierarchical neural learning can suppress quantum-chaotic spectral rigidity and drive deviations toward an uncorrelated white-noise limit.

The Gist

Imagine you are trying to predict the weight of every single apple in a massive orchard. You have a very good rule of thumb (a "global model") that says, "Big apples weigh more, small apples weigh less." This rule works well for most apples, but if you look closely, there are always tiny differences between your prediction and the actual weight. Maybe a specific apple is slightly heavier because of a unique pattern of seeds inside, or slightly lighter because of a tiny bruise.

In the world of physics, scientists do the same thing with atomic nuclei (the tiny cores of atoms). They have complex mathematical formulas to predict the mass of every nucleus. But just like with the apples, there are always small "residuals"—tiny differences between the predicted mass and the real, measured mass.

For a long time, scientists wondered: Are these tiny differences just random noise (like static on a radio), or do they hide a secret, complex pattern?

This paper introduces a new way to answer that question using Artificial Intelligence (AI), but not in the usual way. Here is how they did it, explained simply:

1. The Problem: The "Messy" Leftovers

The scientists started with three different, highly respected formulas (models) for predicting nuclear mass. Even with these advanced formulas, there were still leftover errors.

• Some errors were smooth and predictable (like a gentle slope).
• Some errors were chaotic and jagged (like a rocky path).

The goal was to separate the smooth parts from the chaotic parts to see what was really going on inside the nucleus.

2. The Solution: The "Hierarchical Filter"

Instead of using AI to just guess the final weight of the apple (which is what most people do), the authors used AI as a specialized filter. They built a "sieve" with different levels of mesh size.

• The First Layer (The Coarse Sieve): They used a simple AI to catch the big, smooth errors. Think of this as a net that catches the large rocks but lets the sand through.
• The Second Layer (The Medium Sieve): They took what was left over and ran it through a slightly more complex AI to catch the medium-sized bumps.
• The Final Layers (The Fine Sieve): They kept going, layer by layer, using increasingly complex AI networks. Each layer was trained only on the mistakes the previous layers missed.

They called this the Hierarchical Residual Decomposition (HRD). It's like peeling an onion, where each layer reveals a slightly more detailed texture of the remaining errors.

3. The "PINE" Ensemble

To make sure they weren't just seeing patterns that belonged to one specific formula, they combined the results from all their different AI layers and all three original physics formulas. They mixed them together like a smoothie to create a final, super-accurate prediction tool they called PINE (Physics-Informed Neural Ensemble).

4. The Discovery: Turning Chaos into Silence

The most exciting part of the paper is what happened when they analyzed the "leftovers" after all this filtering.

• Before Filtering: The leftover errors looked like a chaotic, noisy song with a lot of structure. In physics terms, they had "1/f correlations" (a specific type of complex, rhythmic chaos) and "spectral rigidity" (meaning the errors were stiff and connected over long distances). It was like a drumbeat that kept a steady, complex rhythm.
• After Filtering: Once the AI layers stripped away all the smooth trends and the organized chaos, the remaining errors looked like white noise.

The Analogy: Imagine a crowded room where everyone is talking in a complex, rhythmic chant (the chaotic nuclear dynamics). The AI filters are like a series of sound engineers who mute the bass, then the mids, then the highs. By the end, all that is left is the sound of people shuffling their feet and breathing—completely random, unconnected, and flat.

5. What This Means

The paper claims that by using this "peeling" method, they successfully removed almost all the long-range, organized patterns from the nuclear mass errors.

• The Result: The remaining tiny errors are now mostly random and local. They don't stretch across the whole chart of elements; they are just small, isolated quirks.
• The Conclusion: This proves that the "chaos" in atomic nuclei isn't just random noise. It has a structure that can be systematically removed. Once you remove the big, smooth physics and the complex, organized chaos, what's left is just the fundamental, uncorrelated "fuzz" of the quantum world.

In short: The authors built a multi-stage AI machine that acts like a high-tech filter. It stripped away all the predictable trends and complex patterns from nuclear mass errors, leaving behind a "flat" signal that proves the remaining mysteries are truly random and local, rather than part of a giant, hidden global pattern.

Technical Summary

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

Problem Statement
In complex quantum many-body systems like atomic nuclei, regular collective motion coexists with irregular intrinsic dynamics. While global theoretical mass models (e.g., macroscopic-microscopic or self-consistent mean-field approaches) successfully describe smooth trends in nuclear binding energies, they inevitably leave behind residuals. A central question in the field is whether these residuals represent mere model deficiencies or if they encode intrinsic, irreducible quantum-chaotic many-body effects. Specifically, the authors investigate whether the remaining fluctuations in nuclear mass residuals exhibit structured complexity (characterized by $1/f$ spectral correlations and spectral rigidity) or if they are purely stochastic. Existing machine learning approaches often focus on maximizing predictive accuracy by learning the full mass surface; this work posits that treating neural networks as controlled nonlinear filters applied specifically to residuals offers a more rigorous diagnostic tool for isolating the underlying physics of these fluctuations.

Methodology
The authors propose a Hierarchical Residual Decomposition (HRD) framework, culminating in a Physics-Informed Neural Ensemble (PINE) model. The methodology proceeds as follows:

1. Residual Definition: The input is the mass residual $\Delta M(Z, N) = M_{exp}(Z, N) - M_{model}(Z, N)$, calculated against three distinct global mass models: Duflo–Zuker (DZ10), Finite-Range Droplet Model (FRDM), and Hartree–Fock–Bogoliubov (HFB24).
2. Hierarchical Filtering: Instead of a single network, the authors employ a sequence of neural networks trained recursively on the residuals.
   • Architecture: The hierarchy utilizes Fully Connected Feedforward Neural Networks (FFNNs) for global smooth approximation and Mixture-of-Experts (MoE) architectures to adaptively specialize in distinct residual regimes (e.g., shell-dominated vs. deformation-dominated regions).
   • Process: At each stage $i$, a network $f_i$ is trained on the residual field $\Delta M_{i-1}$ from the previous stage. The new residual is $\Delta M_i = \Delta M_{i-1} - f_i$.
   • Regularization Strategy: Training begins with strong $L_2$ regularization to capture broad, long-range correlations. Regularization is progressively relaxed to increase model capacity, allowing the network to resolve and suppress increasingly localized, short-range chaotic fluctuations.
3. Ensemble Construction (PINE): The final PINE model is a weighted linear combination of six independently filtered residual fields (FFNN and MoE variants for each of the three base models). The weights are optimized to minimize the overall prediction error while preserving structures robust across different physical descriptions.
4. Spectral Diagnostics: To quantify the suppression of correlations, the authors apply:
   • One-dimensional Fourier Analysis: Using two complementary "boustrophedon" orderings (mass-based and shell-distance-based) to map the 2D nuclear chart to 1D sequences.
   • Two-dimensional Fourier Analysis: Performed directly on the $(Z, N)$ lattice to analyze spatial correlations without sequential ordering artifacts.
   • Spectral Rigidity ($\Delta_3$): The Dyson–Mehta statistic is used to measure the persistence of long-range cumulative correlations.

Key Contributions

• Controlled Nonlinear Filtering: The paper reframes neural networks not merely as predictive tools but as spectral filters designed to systematically extract and suppress coherent structures from residuals.
• Hierarchical Decomposition: The HRD framework successfully separates residual dynamics into global (low-frequency), intermediate, and localized (high-frequency) components by varying the network capacity and regularization.
• Quantitative Diagnostic of Chaos: The study provides a quantitative measure of the "resolution threshold" required to drive nuclear mass residuals toward an uncorrelated white-noise limit, thereby diagnosing the scale-dependent complexity of many-body correlations.

Results

• Suppression of Low-Frequency Correlations: Before filtering, the residuals from all three base models exhibit strong low-frequency dominance with steep negative spectral slopes (indicating coherent, long-range structures). After HRD filtering, the spectral slopes collapse toward zero (e.g., from $\approx -0.95$ to $\approx -0.06$ for DZ-MoE in mass-based ordering).
• Transition to White Noise: The Fourier power spectra of the PINE residuals become significantly flatter across the full frequency range. The radial power spectrum slope improves from $\approx -1.726$ (averaged model) to $\approx -0.211$ (PINE), approaching the white-noise limit.
• Elimination of Spectral Rigidity: The Dyson–Mehta $\Delta_3$ statistic, a measure of long-range rigidity, drops dramatically from $\approx 270$ in the averaged model to $\approx 1.19$ in the PINE model. This indicates a near-complete elimination of the rigid, collective correlations characteristic of quantum-chaotic systems.
• Localization of Remaining Fluctuations: The remaining fluctuations in the PINE model are predominantly localized, fragmented, and concentrated at higher frequencies. They are less coherent across the nuclear chart and show weaker dependence on global shell structures, particularly in the HFB and DZ models.
• Model Independence: Despite the different systematic deficiencies of the DZ, FRDM, and HFB models, the post-filtered residuals from all three converge toward a similar weakly correlated regime, suggesting that the hierarchical learning procedure successfully isolates the learnable global physics.

Significance and Claims
The authors claim that the hierarchical neural filtering approach effectively isolates the "irreducible" complexity of the nuclear many-body problem. By driving the residuals toward a white-noise limit, the framework demonstrates that the dominant low-frequency correlations in nuclear mass deviations are largely systematic and learnable (associated with mean-field effects, shell evolution, and collective trends). The persistence of only weak, localized high-frequency fluctuations suggests that the remaining deviations are governed by short-range many-body effects, local shell fluctuations, and stochastic contributions that do not organize coherently over large scales.

The paper asserts that this methodology serves as a general spectral filtering tool capable of decomposing residual dynamics into global, intermediate, and localized components. It provides a quantitative diagnostic for the underlying scale-dependent complexity of nuclear mass deviations, distinguishing between model deficiencies and intrinsic many-body dynamics. The authors note that this approach can be extended to other nuclear observables (radii, decay properties, excitation spectra) and broader machine learning problems where the goal is the systematic isolation of hidden correlation structures across multiple scales.