Further exploration of binding energy residuals using machine learning and the development of a composite ensemble model

This paper introduces the Four Model Tree Ensemble (FMTE), a composite machine learning model that combines three new residual-based models with a prior model to predict nuclear binding energies with high accuracy, identifying the least-squares boosted ensemble of trees as the superior approach for interpolating and extrapolating binding energy residuals.

The Gist

The Big Picture: Predicting the Weight of Atoms

Imagine the universe is a giant Lego set. The "bricks" are atoms (specifically, their nuclei). To understand how the universe works—how stars explode, how heavy elements are forged, or how to build better nuclear batteries—we need to know the exact weight (binding energy) of every single Lego brick.

However, there are thousands of different types of atomic bricks. Scientists have measured the weight of many, but there are still huge gaps, especially for the weird, unstable, and rare bricks found at the edges of the periodic table.

The Problem:
Scientists have mathematical formulas (called "Mass Models") to guess the weight of these unmeasured bricks. Think of these formulas like weather forecasts. Sometimes they are pretty good, but often they are off by a significant amount. If you are trying to predict a solar storm (an astrophysical event), being off by a little bit can lead to a completely wrong prediction.

The Goal:
The authors of this paper wanted to build a "Super-Weatherman" for atoms. They wanted a system that could predict atomic weights with extreme precision, even for bricks that have never been weighed before.

---

The Strategy: The "Correction Team"

Instead of trying to build a new formula from scratch, the team decided to use a team of existing "forecasters" (the four mass models: FRDM, HFB, WS, and DZ) and ask a new question: "Where do you usually get it wrong?"

1. The Old Models: These are like four different weather apps. They all predict the temperature, but they all have specific blind spots.
2. The Residuals (The Mistakes): The team looked at the "residuals." In plain English, this is just the difference between what the old models predicted and what the actual experiment measured.
   • Analogy: If a weather app says it will be 70°F, but it's actually 75°F, the "residual" is +5°F.
3. The Machine Learning (ML) Crew: The team hired four different types of AI detectives (SVM, GPR, Neural Networks, and Tree Ensembles) to study these mistakes. Their job wasn't to predict the weather; their job was to learn the pattern of the mistakes so they could fix them.

The Breakthrough: The "Tree" Detective Wins

The team tested four different AI styles to see which one was best at learning these patterns.

• The Neural Network: Like a brain trying to memorize every single detail. It was good, but sometimes it got confused.
• The Support Vector Machine: Like a strict rule-follower. It was okay, but a bit rigid.
• The Gaussian Process: Like a smooth artist. It drew nice curves but sometimes missed the sharp edges.
• The LSBET (Least-Squares Boosted Ensemble of Trees): This was the winner.
  • The Analogy: Imagine a group of 3,000 junior detectives. Each one looks at a tiny piece of the data and makes a guess. Then, they pass the baton to the next detective, who looks at the mistakes the previous group made and tries to fix them. They keep doing this, layer by layer, until the errors are tiny.
  • Why it won: It was the best at both interpolation (filling in the blanks between known data) and extrapolation (guessing what happens in totally unknown territory, like the edge of the universe).

The Masterpiece: The "Four Model Tree Ensemble" (FMTE)

The team didn't just pick one winner; they built a "Super-Team."

They took the best-performing AI models (mostly the "Tree" detectives) and combined them into one Composite Model called the FMTE.

• The Analogy: Imagine you are trying to guess the price of a rare coin. You ask four experts.
  • Expert A is great at old coins but bad at new ones.
  • Expert B is great at silver but bad at gold.
  • The FMTE is like a committee that listens to all of them, but gives more weight to the expert who is currently right.
  • The Result: This committee (FMTE) is incredibly accurate.

The Results: How Good is It?

• The Old Way: The original formulas were off by about 200 to 700 keV (a unit of energy). That's like guessing the weight of a car and being off by 500 pounds.
• The FMTE Way: The new model is off by only 34 keV on average. That's like guessing the weight of a car and being off by a single apple.
• The Catch: While it's amazing, it's not perfect yet. For the most extreme, unstable atoms (near the "neutron drip line," where atoms fall apart), the model still struggles a bit. It's like the weatherman is great at predicting rain in your town but still gets confused when predicting a hurricane in the middle of the ocean.

Why Does This Matter?

1. For Astronomers: To understand how stars create heavy elements (like gold and uranium) in supernovas, we need to know the exact weights of unstable atoms. The FMTE gives them a much better map.
2. For Experimenters: It tells scientists exactly which atoms to go measure next. If the model says "This atom is weird," scientists can go to a lab (like the one at Michigan State University) and test it.
3. For Physics: It proves that combining old-school physics formulas with modern AI is a winning strategy.

Summary in One Sentence

The authors took four imperfect physics formulas, used AI to learn exactly how they were wrong, and combined them into a "Super-Model" that predicts the weight of atoms with the precision of a master jeweler, helping us understand the building blocks of the universe.

Technical Summary

1. Problem Statement

Traditional microscopic and micro-macro nuclear binding energy models (e.g., FRDM, HFB, WS) typically reproduce experimental values with standard deviations ranging from 200 to 700 keV. However, modern astrophysical calculations, particularly those involving the rapid neutron-capture process (r-process), require significantly higher precision, with a suggested maximum standard deviation of 50 keV. Furthermore, existing models diverge significantly far from the valley of stability (near the neutron drip line), leading to uncertainties in predicting the limits of nuclear matter. While Machine Learning (ML) has been used to model binding energies directly or correct residuals, there is a need to identify the most robust ML approach for modeling binding energy residuals (the difference between experimental and theoretical values) and to determine if a composite ensemble can outperform individual models.

2. Methodology

Data Sources and Preprocessing

• Training Data: The Atomic Mass Evaluation (AME) 2012 dataset, with 400 isotopes removed to create an independent test set.
• Test Data: A combination of AME 2020 data, including:
  • 121 new isotopes not present in AME 2012.
  • 57 isotopes with significant value changes (>100 keV) between AME 2012 and 2020.
  • 326 "seed" values (1 in 7 of the remaining isotopes) to test interpolation and extrapolation.
• Target: The models predict the binding energy residuals ($\Delta B = B_{expt} - B_{model}$) rather than the binding energy itself.

Base Mass Models

Four theoretical mass models were selected as baselines for residual calculation:

1. FRDM 2012: Finite-Range Droplet Model (liquid droplet with microscopic corrections).
2. HFB 31: Hartree-Fock-Bogoliubov (microscopic interactions and pairing).
3. WS 4: Weizsäcker-Skyrme (micro-macro model).
4. WSRBF: WS 4 with Radial Basis Function correction (used as a benchmark).

Machine Learning Approaches

Four distinct ML algorithms were trained to model the residuals for each of the four mass models (resulting in 16 initial models):

1. Support Vector Machine (SVM): Used with a Gaussian kernel.
2. Gaussian Process Regression (GPR): Tested with Exponential, Squared Exponential, Rational Quadratic, and Matérn 5/2 kernels.
3. Fully Connected Neural Network (FCNN): Optimized with two hidden layers and Tanh activation functions.
4. Least-Squares Boosted Ensemble of Trees (LSBET): A gradient boosting approach using decision trees.

Feature Engineering

Models utilized physical features including:

• Nucleon counts: Proton number ($Z$), Neutron number ($N$), Mass number ($A$), Isospin projection ($T_Z$).
• Shell parameters: Shell scaling parameters ($\nu, \zeta$) based on magic numbers, and subshell numbers ($N_S, Z_S$).
• Parity: Boolean operators for even/odd $N$ and $Z$ ($N_E, Z_E$).
• Shape parameters: Deformation parameters ($\beta_2, \beta_3, \beta_4, \beta_6$) and charge radius ($R_C$) specific to each mass model.
• Feature Selection: Shapley value analysis was used to determine the hierarchy of feature importance, revealing that the ML algorithm type dictated feature influence more than the specific mass model.

Composite Model Construction

The authors developed the Four Model Tree Ensemble (FMTE). This is a weighted ensemble of four specific LSBET models:

• WSLSBET (Weizsäcker-Skyrme based)
• DZLSBET (Duflo-Zuker based, from prior work)
• FRDMLSBET
• HFBLSBET
  Weights were optimized using amplitude cycling to minimize the Mean Average Deviation (AE) on the test set.

3. Key Contributions

• Identification of Optimal ML Algorithm: The study determined that Least-Squares Boosted Ensemble of Trees (LSBET) is the superior approach for modeling binding energy residuals. It demonstrated better interpolation and extrapolation capabilities compared to SVM, GPR, and FCNN.
• Feature Efficiency: LSBET models achieved high accuracy using fewer physical features (often just $N, Z, T_Z, A, \nu, \zeta, N_E, Z_E$) compared to other methods, suggesting robustness against overfitting.
• Development of FMTE: The creation of a composite ensemble model that leverages the strengths of multiple mass models and ML corrections.
• Analysis of Overfitting: The paper provides a critical analysis of overfitting, noting that while some models (like WSRBFGPR) showed excellent training performance (eV scale), they failed significantly on test sets (keV scale), highlighting the necessity of independent test sets like AME 2020.
• Extrapolation Behavior: The study analyzed model behavior near the neutron drip line, showing that LSBET models generally correct theoretical models inward toward a midpoint, avoiding the extreme deviations seen in raw mass models.

4. Results

• Performance Metrics (AME 2020 Test Set):
  • The FMTE model achieved a Standard Deviation ($\sigma$) of 76 keV and a Mean Average Deviation (AE) of 34 keV for nuclei with $N>7$ and $Z>7$.
  • This represents a significant improvement over the original mass models (e.g., WS had $\sigma \approx 295$ keV) and other ML approaches.
  • The FMTE's AE of 34 keV is comparable to the average experimental uncertainty of the AME 2020 data (23 keV).
• Comparison with Recent Measurements:
  • When tested against 207 recent mass measurements (post-AME 2020), the FMTE reduced the average AE from ~454 keV (original models) to 206 keV.
  • The constituent LSBET models reduced the average AE to 307 keV.
• Garvey-Kelson Relations: The FMTE successfully reproduces the "Wigner cusp" at $N=Z$ and maintains smooth behavior elsewhere, aligning well with experimental trends.
• Limitations:
  • For a set of 33 newly measured isotopes (previously unmeasured), the standard deviation of FMTE was 376 keV.
  • This indicates that while FMTE excels at interpolation and near-extrapolation, it still falls short of the 50 keV target required for high-precision astrophysical r-process calculations.
  • Evidence suggests potential overfitting in the training data, likely due to evolving shell structures far from stability that current physics inputs do not fully capture.

5. Significance

This work establishes LSBET as the preferred machine learning methodology for refining nuclear mass models via residual correction. The FMTE model represents the current state-of-the-art in data-driven nuclear mass prediction, offering a significant reduction in error compared to traditional theoretical models.

While it does not yet meet the strict 50 keV threshold for all astrophysical applications, it provides a highly accurate tool for:

1. Guiding new experimental measurements at facilities like FRIB.
2. Providing improved weighted averages for regions far from stability where individual models diverge.
3. Serving as a baseline for future ML calculations involving other nuclear properties (e.g., charge radii, half-lives).

The study underscores that while ML can significantly correct systematic errors in mass models, the ultimate precision is limited by the underlying physics assumptions of the base models and the availability of experimental data in exotic regions.