⚛️ nuclear theory

Bridging Ab Initio Symmetries and Global Nuclear Masses with Interpretable Neural Networks

This paper demonstrates that interpretable neural networks incorporating Wigner's SU(4) and Elliott's SU(3) symmetries not only achieve competitive accuracy in predicting global nuclear masses but also reveal that these fundamental symmetries govern binding energies across the entire nuclear chart, offering crucial physical insights into phenomena like symmetry restoration near the neutron dripline.

Original authors: Phong Dang, Evander Espinoza, Xiaoliang Wan, Michela Negro, Jerry P. Draayer, Feng Pan, Tomas Dytrych, Daniel Langr, David Kekejian

Published 2026-06-29✓ Author reviewed ⓘ

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Phong Dang, Evander Espinoza, Xiaoliang Wan, Michela Negro, Jerry P. Draayer, Feng Pan, Tomas Dytrych, Daniel Langr, David Kekejian

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the atomic nucleus as a tiny, chaotic dance floor filled with protons and neutrons. For decades, physicists have tried to write a single "rulebook" to predict how tightly these dancers hold onto each other (their binding energy, which determines the atom's mass). Traditionally, they've used a "liquid drop" model, treating the nucleus like a blob of water where the main factors are just how big the blob is and how much the dancers repel each other.

This paper asks a different question: Are there hidden, elegant dance patterns (symmetries) that govern the entire nuclear chart, from the lightest atoms to the heaviest, super-heavy ones?

The authors, led by Phong Dang, decided to test this using a new kind of "smart" computer program called an Interpretable Neural Network. Here is the breakdown of their journey:

1. The Old Way vs. The New Clues

Think of the traditional "liquid drop" model as a weather forecast that only looks at temperature and humidity. It's okay, but it misses the wind patterns.

The authors brought in two specific "dance patterns" discovered in physics:

Wigner's SU(4): A symmetry that treats protons and neutrons as interchangeable partners in a specific way.
Elliott's SU(3): A symmetry related to the shape of the nucleus (whether it's round like a ball or stretched like a football).

Instead of just feeding these patterns into a "black box" computer that spits out a number, they built a system designed to explain itself. They wanted to see how these patterns influence the mass, not just what the mass is.

2. The Three "Smart" Models

The team built three different versions of their AI to test these ideas:

FINN (The Point Predictor): This model just looks at the dance patterns and says, "I predict the mass is X." It's like a student who memorizes the answer key.
GINN (The Uncertainty Calculator): This model is more cautious. It says, "I predict the mass is X, but I'm only 80% sure." It's like a weather forecaster who gives a percentage chance of rain.
WINN (The Rulebook Writer): This is the star of the show. Instead of a black box, WINN is a bilinear formula. Imagine it as a recipe with eight specific ingredients (the dance patterns). The AI doesn't just guess the mass; it learns how much of each "ingredient" to mix in for every specific atom. It's like a chef who learns the exact ratio of salt, pepper, and sugar needed for a dish, rather than just guessing the taste.

3. The Results: Symmetry Wins

When they tested these models against real-world data (using a database of known atomic masses called AME2016 and testing them on newly discovered atoms from AME2020), the results were surprising:

The Liquid Drop was weak: It made big mistakes, especially for exotic atoms.
Adding Wigner's Symmetry (SU(4)) was a game-changer: Just adding this one symmetry cut the prediction errors in half. It proved that Wigner's old idea from 1937 still holds power across the entire nuclear chart, not just in small atoms.
WINN was the champion: The "Rulebook" model (WINN) achieved the lowest error rate (0.430 MeV). This is competitive with the most complex, high-tech models in the world, but WINN did it with only eight simple terms.

The Big Takeaway: The authors argue that the reason WINN worked so well isn't just because it's a good calculator. It's because the "eight ingredients" (the symmetries) actually capture the true physics of how nuclei stick together. The universe seems to follow these simple, elegant rules.

4. What Did They Discover?

Because their model was "interpretable" (we could see the recipe), they found two fascinating things:

The "Neutron Drip Line" Rescue: Near the edge of the nuclear chart where atoms are packed with too many neutrons, the Wigner symmetry (SU(4)) suddenly became very strong. It's as if the "dance partners" (protons and neutrons) started holding hands much tighter in these extreme conditions, restoring a symmetry that was previously thought to be broken.
The Superheavy Surprise: In the heaviest, super-heavy elements, a specific "four-body" interaction (a quartic term) suddenly became important. This suggests that in these massive nuclei, groups of four particles might be forming special clusters, a hint of new physics in the heaviest regions of the universe.

Summary

In simple terms, this paper is a victory for simplicity and elegance. The authors showed that you don't need a massive, complicated computer to predict atomic masses if you understand the underlying "dance moves" (symmetries) of the particles. By building a model that respects these ancient rules, they created a tool that is not only accurate but also tells us why the universe is built the way it is. They didn't just predict numbers; they uncovered the hidden order of the atomic world.

Technical Summary: Bridging Ab Initio Symmetries and Global Nuclear Masses with Interpretable Neural Networks

Problem Statement
While ab initio nuclear theory has established Wigner's SU(4) and Elliott's SU(3) symmetries as dominant features in light and intermediate-mass nuclei, it remains unclear whether these symmetries govern nuclear binding across the entire nuclear chart, including exotic isotopes near the driplines. Traditional machine learning (ML) approaches to nuclear mass prediction often function as "black boxes," capturing patterns without providing physical insight. Conversely, traditional mass formulas (e.g., Liquid Drop Model) lack the microscopic detail to predict masses with high precision. The authors aim to bridge this gap by developing interpretable neural network (NN) models that utilize symmetry-based features to test if SU(3) and SU(4) symmetries can serve as a governing principle for global nuclear systematics, prioritizing physical insight over mere high-precision benchmarking.

Methodology
The study employs a three-stage methodology involving data curation, feature engineering, and model development:

Data Collection and Splitting:
- Nuclear masses were collected from the AME2016 dataset (excluding theoretical extrapolations) and split into a training set (1908 nuclei, 80%) and a test set (478 nuclei, 20%).
- The AME2020 dataset provided a validation set of 71 nuclei first reported in that edition, serving as a benchmark for extrapolation to unseen, often exotic, nuclei.
- Light nuclei ( $Z < 8, N < 8$ ) were excluded.
Symmetry-Inspired Feature Engineering:
The models utilize nine physics-informed features grouped into three categories:
- Liquid Drop Model (LDM) Features: Proton number ( $Z$ ), neutron number ( $N$ ), surface term proxy ( $A^{2/3}$ ), and isospin projection ( $T_z = (N-Z)/2$ ).
- SU(4) Features: The second ( $C_2$ ), third ( $C_3$ ), and fourth ( $C_4$ ) order Casimir operators. These describe spin-isospin polarization forces. The authors hypothesize that the dominant irreducible representation (irrep) with the lowest $C_2$ value, identified in ab initio calculations for light nuclei, extends across the chart.
- SU(3) Features: The total harmonic oscillator quanta ( $N\hbar\omega$ ) and the second ( $C_2$ ) and third ( $C_3$ ) order Casimir operators of the SU(3) group, which encode quadrupole collectivity and deformation.
- Feature Selection: Correlation analysis using the Maximal Information Coefficient (MIC) revealed that the cubic SU(4) Casimir ( $C_3$ ) was weakly correlated with other features and exhibited sign fluctuations between odd-mass nuclei. Consequently, $C_3$ [SU(4)] was excluded from the final models as an irrelevant feature.
Neural Network Architectures:
Three distinct models were constructed:
- FINN (Feature-Informed NN): A standard single-output NN mapping features directly to binding energy (BE). It uses a Mean Squared Error (MSE) loss and a physics-informed Garvey-Kelson (GK) regularization term.
- GINN (Gaussian-Informed NN): A two-output NN predicting both the mean BE ( $\mu_{BE}$ ) and the uncertainty ( $\sigma_{BE}$ ). It employs a negative log-likelihood (NLL) loss to quantify uncertainty.
- WINN (Wigner-Informed NN): An interpretable, bilinear mass formula where the coupling strengths are learned by a NN. The formula is defined as:
  $BE = a + bZ^2 + cT_z^2 + dN\hbar\omega + eC_2[\text{SU(3)}] + fC_3[\text{SU(3)}] + gC_2[\text{SU(4)}] + hC_4[\text{SU(4)}]$
  Here, the coefficients $\{a, \dots, h\}$ are functions of $N$ and $Z$ learned by the network, allowing the model to adapt the symmetry contributions across the nuclear landscape.

Key Results

Performance Metrics:
- The LDM baseline achieved a validation Root-Mean-Square Error (RMSE) of 1.270 MeV (FINN).
- Adding SU(4) features alone reduced the validation RMSE to 1.024 MeV, demonstrating that Wigner's symmetry carries predictive information beyond bulk properties.
- The combination of LDM, SU(4), and SU(3) features yielded the best performance for the black-box models: 0.624 MeV (FINN) and 0.793 MeV (GINN).
- WINN achieved the lowest validation RMSE of 0.430 MeV, outperforming both FINN and GINN. This result is competitive with state-of-the-art mass models despite using only eight symmetry-motivated terms.
Extrapolation Capability:
- When compared to the WS3 and HFB-26 models on nuclei beyond the AME2020 boundary, WINN showed good agreement (within 5 MeV) for nuclei within 5 neutrons of the known region.
- Errors increased systematically with distance from the known chart, reflecting the inherent limits of extrapolating a compact eight-term model.
Physical Insights:
- SU(4) Restoration: The quadratic SU(4) Casimir ( $C_2$ ) showed a pronounced enhancement (contributing 10–40% of the binding energy) near the neutron dripline ( $Z \approx 8-60$ ), signaling the restoration of Wigner's symmetry due to the suppression of spin-orbit splitting.
- Superheavy Region: The quartic SU(4) Casimir ( $C_4$ ) gained significant strength in the superheavy region ( $N \gtrsim 126$ ), becoming competitive with $C_2$ , potentially linked to the dominance of $\alpha$ -systematics.
- Pairing: The model successfully captured pairing effects without an explicit pairing term, as evidenced by the absence of systematic sign inversions or odd-even staggering in the residuals.
- Activation Functions: The study found that piecewise-linear activation functions (ReLU, PReLU) outperformed smooth functions (Tanh, SiLU) in reproducing the sharp drops in two-neutron separation energy ( $S_{2n}$ ) at shell closures.

Significance and Claims
The paper claims that embedding physical symmetries directly into the neural network architecture (as in WINN) is more effective than treating them merely as input features. The primary significance lies not in the absolute precision of the mass predictions, but in the physical insight gained:

Validation of Symmetries: The results provide evidence that Wigner's SU(4) and Elliott's SU(3) symmetries are not limited to light nuclei but act as governing principles for the entire nuclear chart, including exotic isotopes.
Interpretability: Unlike standard ML models, WINN reveals how symmetries influence binding energy in different regions (e.g., the enhancement of SU(4) near the dripline).
Emergent Principles: The work elevates emergent symmetries from being descriptors of individual nuclei to a framework for understanding global nuclear systematics.
Future Directions: The authors suggest that while the current symmetry basis is compact and effective, future work could extend the framework to other observables (e.g., $\beta$ -decay) or address symmetry breaking through improved irrep selection or by leveraging the NN's capacity to absorb non-linearities associated with broken symmetries.