Architecture as physical prior: cooperative neural network for nuclear masses

This paper introduces the Cooperative Neural Network (CoNN), a baseline-free architecture that embeds physical priors directly into its structure to predict nuclear binding energies with high accuracy using only proton and neutron numbers, effectively replacing the need for hand-crafted physics features or theoretical baselines.

The Gist

Imagine you are trying to predict the weight of every single type of Lego castle in the universe. You have a list of every castle that has ever been built (the data), but you also need to guess the weight of castles that haven't been built yet, or ones that are so weird they don't exist in your list.

In the world of physics, these "castles" are atomic nuclei (the cores of atoms), and their "weight" is called binding energy. Knowing this weight is crucial because it tells us how stars explode, how new elements are formed, and whether a nucleus will stay stable or fall apart.

For decades, scientists have used two main ways to guess these weights:

1. The "Physics Formula" Way: Using complex math based on how we think atoms work. It's good, but sometimes it misses the details.
2. The "AI Correction" Way: Using a computer to learn the mistakes of the Physics formulas. It's very accurate, but it's like a student who can only do math if they have a textbook right next to them. They can't do it from scratch.

The New Idea: "Architecture as Physical Prior"

This paper introduces a new AI model called CoNN (Cooperative Neural Network). The authors ask a bold question: What if we don't need a textbook (a physics baseline) or a list of special features? What if we just build the AI's "brain" in a way that naturally understands physics?

Think of it like this:

• Old AI: A general-purpose student who is given a raw list of numbers (Proton count, Neutron count) and told, "Figure out the weight." The student gets confused because the data is too messy and complex.
• CoNN: Instead of a general student, the authors built a team of four specialized experts who work together. They are all inside one computer program, but each has a specific job that matches how nature actually works.

The Four Experts (The Architecture)

The CoNN breaks the problem down into four parts, just like a physicist would:

1. The Smooth Trend Expert (The Macroscopic Branch):

   • The Job: This expert handles the "big picture." It knows that as you add more bricks to a Lego castle, the weight generally goes up in a smooth, predictable curve.
   • The Analogy: Imagine a painter who only knows how to paint a smooth, flat blue sky. They get the background right, but they can't paint the trees or clouds.

2. The "Magic Number" Expert (Shell Embeddings):

   • The Job: In the atomic world, certain numbers of protons or neutrons (like 2, 8, 20, 50) make the nucleus extra stable, like a perfect square of Legos. These are called "magic numbers."
   • The Analogy: This expert is like a librarian who knows that books on specific shelves (the magic numbers) are heavier or lighter than expected. They add a little "jolt" of energy whenever the count hits these special numbers.

3. The "Regional Map" Expert (Correlation Grid):

   • The Job: Sometimes, protons and neutrons interact in complex ways depending on where they are on the map of all possible atoms.
   • The Analogy: This is like a weather forecaster looking at a 2D map. They know that in the "Rare Earth" region of the map, the weather (energy) behaves differently than in the "Actinide" region. They draw a flexible grid over the map to catch these local patterns.

4. The "Odd/Even" Expert (Pairing Network):

   • The Job: Nuclei with an even number of particles are usually more stable than those with an odd number. It's a "sawtooth" pattern.
   • The Analogy: This expert is like a bouncer at a club who only lets in pairs. If you have an odd number of people, the energy is different. This expert specifically looks at whether the numbers are even or odd and adjusts the prediction accordingly.

How They Work Together

The magic of CoNN isn't just that it has these experts; it's how they are trained.

Usually, if you put four experts in a room, they might argue. The "Smooth Trend" guy might try to explain the "Magic Number" jumps, or the "Odd/Even" guy might try to fix the "Big Picture."

The authors used a special training method called Alternating Training:

1. First, they let the "Smooth Trend" expert learn the big picture alone.
2. Then, they freeze that expert and let the other three experts learn the leftover details (the jumps, the maps, the odd/even stuff).
3. They switch back and forth.

This ensures that each expert stays in their lane, just like a well-organized orchestra where the bassists don't try to play the violin solos.

The Results: Why It Matters

The team tested this model on the latest data (AME2020), which includes 3,558 different atomic nuclei.

• Accuracy: CoNN predicted the weights with an error of only 0.269 MeV. This is incredibly precise.
• The "From Scratch" Win: Most AI models need a physics textbook (a baseline) to be this accurate. CoNN did it using only the proton and neutron counts. It learned the physics rules just by looking at the structure of its own brain.
• Discovery: Even though the AI wasn't told what "magic numbers" were, it figured them out on its own! The "Magic Number" expert developed strong signals exactly at the numbers physicists know are special. It also correctly predicted the "sawtooth" pattern of odd/even stability.

The Bottom Line

This paper is a breakthrough because it changes how we build AI for science. Instead of feeding the AI a list of "physics facts" (features) and hoping it learns, we build the physics facts directly into the AI's skeleton.

It's the difference between giving a student a cheat sheet and teaching them to think like a physicist. The CoNN proves that if you design a neural network with the right "inductive biases" (structural assumptions that match reality), it can learn the laws of the universe directly from the data, without needing a textbook to hold its hand.

Technical Summary

Here is a detailed technical summary of the paper "Architecture as physical prior: cooperative neural network for nuclear masses" by Peiwen Zai, Wei Cheng, and Feng-Shou Zhang.

1. Problem Statement

Predicting nuclear binding energies is critical for understanding nuclear stability, decay energetics, and astrophysical processes like the rapid neutron-capture process (r-process). While theoretical models (e.g., Density Functional Theory, macroscopic-microscopic models) exist, they often struggle with accuracy (typically 0.5–0.8 MeV RMSD) or rely on complex parameterizations.

Machine Learning (ML) has emerged as a powerful tool, but current approaches face two main limitations:

1. Residual Correction: Most ML models train on the difference between experimental data and an existing theoretical baseline. While accurate, these are not standalone predictors and are tethered to the errors of the baseline model.
2. Direct Prediction with Feature Engineering: Models attempting to predict binding energy directly from proton ($Z$) and neutron ($N$) numbers often fail to achieve high accuracy unless they rely on hand-crafted physics features (e.g., parity indicators, distances to shell closures). This requires significant domain expertise and limits the model's ability to generalize if feature choices are suboptimal.

The Core Challenge: Can a neural network predict nuclear masses directly from raw $(Z, N)$ inputs with high accuracy, without relying on external theoretical baselines or manually engineered input features, by instead embedding physical knowledge directly into the network's structure?

2. Methodology: The Cooperative Neural Network (CoNN)

The authors propose the Cooperative Neural Network (CoNN), a modular architecture that decomposes the binding energy prediction into four structurally constrained components based on the macroscopic-microscopic decomposition of nuclear physics.

Input: Only the proton number ($Z$) and neutron number ($N$).
Output: Predicted Binding Energy ($B_{pred}$).

The model is defined as:
$$B_{pred} = E_{Macro} + E_{Shell} + E_{Cor} + E_{Pair}$$

A. Architectural Modules

1. Macroscopic Branch ($E_{Macro}$):

   • Function: Captures the smooth, bulk liquid-drop trends.
   • Structure: A fully connected neural network (encoder-decoder) with a narrow bottleneck (16 dimensions).
   • Physical Bias: The narrow bottleneck and spectral bias of gradient descent force the network to learn low-frequency, smooth variations, preventing it from fitting high-frequency noise (shell effects/pairing).

2. Shell Embeddings ($E_{Shell}$):

   • Function: Captures discrete shell effects at magic numbers.
   • Structure: Learnable scalar embeddings indexed by $Z$ and $N$ ($e_Z[Z] + e_N[N]$).
   • Physical Bias: Mimics the independent-particle model where shell closures cause discrete energy jumps. This module cannot capture proton-neutron correlations.

3. Regional Correlation Grid ($E_{Cor}$):

   • Function: Captures non-separable, collective effects (e.g., deformation) that depend on both $Z$ and $N$ jointly.
   • Structure: A learnable 2D grid ($50 \times 60$) spanning the nuclear chart, using bilinear interpolation.
   • Physical Bias: Enforces spatial continuity to model smooth regional variations (like quadrupole deformation) while remaining distinct from the discrete shell jumps.

4. Pairing Network ($E_{Pair}$):

   • Function: Captures the high-frequency odd-even staggering.
   • Structure: A small Multi-Layer Perceptron (MLP) taking inputs: parity bits ($Z \mod 2, N \mod 2$) and normalized $Z/100, N/100$.
   • Physical Bias: The modulo operation is a fixed, non-learnable transformation that explicitly extracts parity, allowing the network to learn the specific amplitude of the staggering without needing to infer parity from raw numbers.

B. Training Protocol

To prevent the modules from "stealing" each other's contributions (e.g., the macroscopic branch absorbing shell effects), the authors use an alternating training protocol:

1. Warmup: Train only the macroscopic branch to establish a smooth baseline.
2. Cooperative Training: Alternate between updating the macroscopic branch (frozen microscopic modules) and the microscopic modules (frozen macroscopic branch).
   • The macroscopic branch learns the residual after subtracting the current microscopic predictions.
   • The microscopic branch learns the residual after subtracting the current macroscopic prediction.
   • Asymmetry: The microscopic modules use a learning rate 10x higher than the macroscopic branch, allowing them to adapt quickly to structured residuals while the macroscopic branch remains anchored to smooth trends.

3. Key Contributions

• Architecture as Physical Prior: Demonstrates that embedding physical constraints (smoothness, discreteness, parity, spatial continuity) directly into the network architecture can replace the need for hand-crafted input features.
• Baseline-Free Direct Prediction: Achieves state-of-the-art accuracy for direct mass prediction using only $(Z, N)$ inputs, without relying on a theoretical baseline model (unlike residual correction methods).
• Interpretability: The modular design ensures that the learned components correspond to distinct physical phenomena (bulk, shell, collective, pairing), making the "black box" transparent.
• Cooperative Training Strategy: Introduces a specific optimization schedule that successfully enforces the separation of physical scales within a single model.

4. Results

The model was evaluated on the AME2020 dataset (3,558 nuclei).

• Accuracy:
  • Overall RMSD: 0.269 MeV on all 3,558 nuclei.
  • Interpolation: 0.419 MeV on a held-out 20% subset.
  • Extrapolation: 0.728 MeV on 122 nuclei measured after AME2016 (a strict test of generalization).
• Comparison:
  • Outperforms standard direct-prediction ML models (e.g., plain MLP, KAN-2) which struggle to go below 0.8 MeV without feature engineering.
  • Matches or exceeds models that use hand-crafted features (e.g., ANN7, KAN-11) while using only 2 inputs instead of 7–11.
  • Significantly outperforms traditional macroscopic-microscopic models (FRDM2012, WS4) on extrapolation tasks, particularly in the superheavy region.
• Ablation Studies:
  • A parameter-matched plain MLP achieved 0.836 MeV, proving the CoNN's success is due to inductive bias, not model size.
  • Removing the pairing module caused RMSD to jump to 1.257 MeV, confirming odd-even staggering is the largest single source of error in microscopic corrections.
• Physical Emergence:
  • Shell Effects: The learned embeddings naturally developed extrema at canonical magic numbers ($Z, N = 20, 28, 50, 82, 126$) without explicit supervision.
  • Pairing: The pairing module reproduced the expected sawtooth pattern and the $1/\sqrt{A}$ scaling of pairing gaps.
  • Correlations: The 2D grid successfully identified regions of collective deformation (rare-earth and actinide sectors) and non-separable proton-neutron correlations near doubly-magic nuclei.

5. Significance and Outlook

• Paradigm Shift: The paper establishes "Architecture as Physical Prior" as a viable alternative to "Feature Engineering." It suggests that the network structure itself should encode the physics, allowing the model to learn the specific parameters from data.
• Interpretability: Unlike standard deep learning models, CoNN provides a decomposed view of nuclear binding energy, recovering known physical signatures (magic numbers, pairing gaps) autonomously.
• Limitations & Future Work:
  • Extrapolation: The model struggles in regions far from training data (e.g., superheavy elements) because the discrete embeddings and finite grid cannot extrapolate. Future work suggests replacing discrete embeddings with continuous parametrizations.
  • Uncertainty: The current ensemble spread is qualitative; future iterations aim to incorporate calibrated Bayesian uncertainty quantification.
  • Extension: The framework could be extended to predict other observables like charge radii and $\beta$-decay properties.

In conclusion, the CoNN demonstrates that by respecting the multi-scale nature of nuclear physics through architectural constraints, neural networks can achieve high-precision, baseline-free predictions and offer deep physical insights directly from raw data.