Systematic Study on the $\alpha$-particle preformation factor in the theory of $\alpha$-decay based on the Tabular Prior-data Fitted Network (TabPFN)

This paper develops a hybrid approach combining the Tabular Prior-data Fitted Network (TabPFN) with the Coulomb and Proximity Potential Model to accurately predict $\alpha$-particle preformation factors, thereby significantly improving $\alpha$-decay half-life calculations and suggesting $N=184$ as a neutron magic number for superheavy nuclei.

The Gist

Here is an explanation of the paper, translated from complex nuclear physics into everyday language using analogies.

The Big Picture: Predicting the Unpredictable

Imagine you are trying to predict when a specific, unstable balloon will pop. In the world of atoms, this "pop" is called Alpha Decay. Sometimes, an atom spits out a tiny bundle of particles (an alpha particle) to become more stable.

Physicists have known for a long time how to calculate when this happens, but there's a missing piece of the puzzle. They know the "pressure" inside the balloon (the energy) and the "thickness" of the rubber (the barrier), but they don't know exactly how likely the balloon is to form a weak spot right before it pops. This weak spot is called the Preformation Factor.

For decades, scientists have had to guess this number with simple rules of thumb. This paper says, "Let's stop guessing and let a super-smart computer learn the pattern instead."

The New Tool: The "TabPFN" Chef

The authors used a new type of Artificial Intelligence called TabPFN (Tabular Prior-data Fitted Network).

Think of traditional AI training like teaching a student by showing them a textbook and then a test. You have to study hard, memorize, and then take the exam.

TabPFN is different. Imagine a chef who has already tasted millions of different soups (synthetic data) and learned the fundamental rules of flavor. When you give this chef a new recipe with just a few ingredients (real nuclear data), they don't need to study the recipe again. They instantly recognize the pattern based on their massive prior experience. They can predict the taste of a soup they've never seen before just by looking at the list of ingredients.

In this paper, the "ingredients" are the properties of the atom (how many protons, how many neutrons, how squashed the atom is, etc.), and the "taste" is the probability of the alpha particle forming.

How They Did It: The Three-Step Recipe

1. The Setup (The Physics): They used a standard physics model (called CPPM) to calculate what the alpha decay should be if the "weak spot" probability was perfect.
2. The Reality Check: They compared their calculation to real-world experimental data. The difference between the two told them what the "real" preformation factor actually was for 498 different atoms.
3. The AI Lesson: They fed these 498 examples into the TabPFN chef. The AI looked at the atomic "ingredients" and learned the complex, hidden rules that determine how likely an alpha particle is to form.

What They Discovered: The "Odd-Even" Dance

The AI didn't just spit out numbers; it learned the physics behind them. It found patterns that humans already suspected but struggled to calculate perfectly:

• The Odd-Even Staggering: Imagine a dance floor. If everyone is paired up (even numbers of protons and neutrons), the dance is smooth and easy. But if one person is left alone (an odd number), the dance gets awkward and harder. The AI found that atoms with "unpaired" particles are much less likely to form an alpha particle. It's harder to get a group of four to form if one of your dancers is missing a partner.
• The Shell Closures: Atoms have "magic numbers" of particles that make them extra stable, like a full shelf in a bookcase. The AI learned that when an atom is near these magic numbers, the "weak spot" is much harder to form, making the atom last longer.

The Results: Sharper Predictions

When the authors used the AI's predictions to calculate how long these atoms would last (their half-lives), the results were amazing:

• Old Way: Without the AI, their predictions were off by a huge margin (like guessing a 10-year lifespan and being off by 5 years).
• New Way: With the AI, the predictions became incredibly accurate, reducing the error by nearly 90%.

The Crystal Ball: Predicting Superheavy Elements

The coolest part? The AI was asked to predict the behavior of elements that are so heavy and unstable they barely exist in nature (Elements 117 to 120).

The AI predicted that for these super-heavy atoms, there is a specific number of neutrons (N = 184) that acts like a "super-magic number." It's like finding a hidden shelf in the bookcase that holds the books perfectly still. This suggests that if scientists can build an atom with exactly 184 neutrons, it might be surprisingly stable and last longer than its neighbors.

The Takeaway

This paper is a victory for Machine Learning in Physics. It shows that instead of just building complex mathematical formulas to describe the universe, we can teach computers to "feel" the patterns in the data.

By using a "chef" who has tasted millions of soups, the researchers were able to perfectly predict the flavor of a new, exotic soup (superheavy atoms) and even guess where the next "magic number" of stability lies. It's a new way of seeing the invisible building blocks of our universe.

Technical Summary

Here is a detailed technical summary of the paper "Systematic Study on the $\alpha$-particle preformation factor in the theory of $\alpha$-decay based on the Tabular Prior-data Fitted Network (TabPFN)".

1. Problem Statement

The $\alpha$-decay half-life is a critical observable for understanding the structure of unstable atomic nuclei, particularly in the superheavy region. The theoretical calculation of the half-life ($T_{1/2}$) relies on the Gamow theory of quantum tunneling, expressed as:
$$T_{1/2} = \frac{\ln 2}{\lambda} = \frac{\ln 2}{P_\alpha \nu P}$$
where $\nu$ is the assault frequency, $P$ is the barrier penetration probability, and $P_\alpha$ is the $\alpha$-particle preformation factor.

Key Challenges:

• Uncertainty in $P_\alpha$: Traditionally, $P_\alpha$ is treated as an empirical constant or derived from semi-empirical formulas. This assumption fails to capture complex nuclear structure effects, particularly near closed shells (e.g., $Z=82, N=126$) and in odd-A or odd-odd nuclei, leading to significant discrepancies between theoretical and experimental half-lives.
• Limitations of Existing ML: While machine learning (ML) has been applied to $\alpha$-decay, many previous studies used direct learning (predicting half-lives directly) or required extensive retraining on specific datasets. There is a need for a method that can effectively learn the underlying physical mapping between nuclear structure properties and the preformation factor without overfitting, especially given the limited size of experimental nuclear data.

2. Methodology

The authors propose a hybrid approach combining the Coulomb and Proximity Potential Model (CPPM) with the Tabular Prior-data Fitted Network (TabPFN).

A. Extraction of Experimental Preformation Factors

To train the ML model, the authors first extracted "experimental" $P_\alpha$ values ($P_\alpha^{exp}$) for 498 nuclei (172 even-even, 261 odd-A, 65 odd-odd) using the CPPM framework:

1. Theoretical Baseline: They calculated the theoretical half-life ($T_{1/2}^{cal}$) assuming $P_\alpha = 1$ using the CPPM, which includes nuclear, Coulomb, and centrifugal potentials.
2. Extraction: Using the experimental half-life ($T_{1/2}^{exp}$), they derived $P_\alpha^{exp}$ via the relation:
   $$P_\alpha^{exp} = \frac{T_{1/2}^{cal}}{T_{1/2}^{exp}}$$
3. Deformation: Calculations were performed for both spherical and deformed potentials (incorporating quadrupole deformation $\beta_2$) to ensure robustness.

B. The TabPFN Model

Instead of training a standard neural network from scratch, the authors utilized TabPFN, a Transformer-based foundation model pre-trained on millions of synthetic datasets generated via Structural Causal Models (SCMs).

• In-Context Learning (ICL): TabPFN does not require gradient-based retraining. It learns the mapping from input features to the target label ($P_\alpha$) directly within the forward pass using the provided dataset as context.
• Input Features: The model was fed 9 physical descriptors:
  1. Mass number ($A$)
  2. Proton number ($Z$)
  3. Neutron number ($N$)
  4. $\alpha$-decay energy ($Q_\alpha$)
  5. Minimum orbital angular momentum ($l$)
  6. Quadrupole deformation ($\beta_2$)
  7. Proton parity ($Z_p$)
  8. Neutron parity ($N_p$)
  9. Pairing effect ($\delta$)

C. Validation Strategy

• Training/Test Split: 399 nuclei for training, 99 for testing.
• Extrapolation: The model was tested on 41 newly evaluated nuclei (Z > 104) from NUBASE2020 to assess generalization to unexplored regions.
• Superheavy Prediction: The model predicted $P_\alpha$ for superheavy nuclei with $Z = 117–120$.

3. Key Contributions

1. Novel Application of TabPFN: This is the first application of the TabPFN foundation model to nuclear physics preformation factors, demonstrating its efficacy on small-to-medium tabular datasets without the need for task-specific retraining.
2. Systematic Feature Analysis: The study systematically quantified the impact of specific nuclear descriptors (pairing, parity, deformation, angular momentum) on prediction accuracy, identifying the optimal feature set (TabPFN12).
3. Physical Insight Extraction: The model successfully captured complex physical phenomena, including odd-even staggering (OES) and shell closure effects, directly from the data without explicit shell-correction terms in the input.
4. Improved Half-Life Predictions: By integrating ML-derived $P_\alpha$ into the CPPM, the authors achieved a drastic reduction in the root-mean-square deviation (RMSD) of predicted half-lives compared to experimental data.

4. Key Results

A. Model Performance

• Accuracy: The optimal model (TabPFN12, using all 9 features) achieved a root-mean-square deviation of $\sigma_{rms} = 0.211$ for $\log_{10} P_\alpha$ across the entire dataset.
• Comparison: This represents a 63.8% improvement over the empirical formula from Ref. [63] and a 66.8% improvement over Ref. [44].
• Feature Importance: Including pairing effects ($\delta$), parities ($Z_p, N_p$), and deformation ($\beta_2$) significantly reduced errors. For instance, adding $\beta_2$ reduced $\sigma_{rms}$ from 0.247 to 0.216.

B. Physical Correlations Discovered

• Linear Correlations: The model confirmed a linear relationship between $\log_{10} P_\alpha$ and $Q_\alpha^{-1/2}$ (extending the Geiger-Nuttall law to preformation factors) and a negative linear correlation with the fragmentation potential ($V_{frag}$).
• Odd-Even Staggering (OES): The model successfully reproduced the OES pattern, showing that unpaired nucleons (odd $N$ or odd $Z$) suppress $P_\alpha$ compared to even-even neighbors.
• Shell Effects: Clear variations in $P_\alpha$ were observed at shell closures ($Z=82, N=126$), confirming the model's ability to learn shell structure implicitly.

C. Half-Life Predictions

• Training Set: Incorporating TabPFN-predicted $P_\alpha$ reduced the RMSD of half-life predictions from 2.028 (without $P_\alpha$) to 0.211 (an 89.5% improvement).
• Extrapolation: For 41 new nuclei (Z > 104), the RMSD improved from 2.920 to 0.771, demonstrating strong generalization capabilities.
• Deformed Potentials: The approach remained robust when applied to deformed Coulomb + Woods-Saxon potentials, yielding $\sigma_{rms} = 0.232$.

D. Superheavy Nuclei ($Z=117-120$)

• The model predicted $P_\alpha$ for superheavy isotopes, revealing a sharp decrease in half-lives between $N=184$ and $N=186$.
• Magic Number Suggestion: This rapid change suggests $N=184$ is a potential new neutron magic number in the superheavy region.

5. Significance

• Refining Nuclear Models: The study demonstrates that machine learning can extract high-fidelity nuclear structure information (like preformation factors) directly from decay data, offering a more accurate alternative to empirical formulas.
• Guiding Experiments: The reliable predictions for superheavy nuclei ($Z=117-120$) provide crucial guidance for future synthesis experiments, specifically regarding stability islands and the search for the $N=184$ shell closure.
• Methodological Advancement: It establishes TabPFN as a powerful tool for nuclear physics, capable of handling small datasets and capturing complex non-linear relationships with minimal computational cost, paving the way for its application to other nuclear properties (masses, radii, reaction rates).