Neural-Network Correlation Functions for Light Nuclei with Chiral Two- and Three-Body Interactions

This paper demonstrates that neural networks can generate highly expressive trial wave functions for light nuclei using chiral two- and three-body interactions, achieving ground-state energies within 0.45% of Green's Function Monte Carlo results and representing a 91% improvement over standard variational Monte Carlo methods.

The Gist

The Big Picture: Predicting the "Heartbeat" of Tiny Nuclei

Imagine you are trying to predict exactly how a tiny, complex machine (like a light atomic nucleus) behaves. In the world of physics, this machine is made of protons and neutrons dancing around each other. To understand them, scientists need to write a "recipe" called a wave function. This recipe tells us the probability of finding these particles in specific spots and how they influence one another.

The problem is that these particles don't just dance in pairs; they have complex group dynamics. Sometimes, three particles interact in ways that two particles alone cannot explain. Finding the perfect recipe for this dance is incredibly hard. If the recipe is too simple, the prediction is wrong. If it's too complicated, it takes a supercomputer forever to calculate.

The Old Way: Guessing and Checking

Traditionally, scientists used a method called Variational Monte Carlo (VMC) to find these recipes. Think of this like trying to tune a radio to a clear station. You have a dial with about 30 knobs (parameters). You turn them manually or with a basic algorithm to get the clearest signal (the lowest energy state).

However, this method has limits:

1. It's slow: Turning 30 knobs to find the perfect setting takes a lot of computing power.
2. It's rigid: The "knobs" are fixed formulas. If the real physics requires a weird, complex shape that the formula doesn't allow, the radio stays fuzzy.
3. It misses the group: When three particles interact, the old recipes often struggle to capture that extra layer of complexity.

Another method, called Green's Function Monte Carlo (GFMC), is like a "gold standard" reference. It is incredibly accurate but requires starting with a very good recipe to work efficiently. If the starting recipe is bad, the calculation gets stuck or takes too long.

The New Solution: The "Smart Chef" (Neural Networks)

The authors of this paper introduced a new tool: Neural Networks (NNs).

Think of a neural network not as a set of fixed knobs, but as a super-smart chef who can learn to cook any dish. Instead of giving the chef a fixed recipe with 30 knobs, you give them a blank slate and say, "Make the best-tasting dish possible." The chef tastes the dish, realizes it needs more salt or a different spice, and adjusts the ingredients automatically.

In this paper, the "dish" is the wave function, and the "taste" is the energy of the nucleus. The lower the energy, the better the dish.

How the "Chef" works in this study:

1. Learning the Pairs: The neural network looks at two particles (a pair) and learns how they interact based on their distance.
2. Learning the Crowd: Crucially, the network also looks at the other particles surrounding that pair. It learns that "When Particle A and B are close, but Particle C is also right next to them, the interaction changes." This allows the model to handle the tricky three-body interactions that old methods missed.
3. The Training: The team used a computer simulation (VMC) to let the neural network "practice" millions of times. Every time the network guessed a wave function, it calculated the energy. If the energy was high, the network tweaked its internal connections to do better next time.

The Results: A Near-Perfect Match

The team tested this "Smart Chef" on the lightest nuclei: Tritium ($^3$H) and Helium-3 ($^3$He). These are nuclei made of three particles (two neutrons and one proton, or vice versa).

They compared their neural network results against the "gold standard" (GFMC):

• The Old Way (Standard VMC): The energy prediction was off by a noticeable margin.
• The New Way (Neural Network VMC): The prediction was incredibly close to the gold standard.
  • For the softest version of the nuclear force they tested, the neural network was 91% better than the standard method.
  • The final energy result was within 0.45% of the gold standard.

To put that in perspective: If the gold standard says a ball weighs 100 grams, the old method might guess 95 grams, but the neural network guessed 99.55 grams.

Why This Matters

The paper shows that neural networks can act as a powerful "translator" for quantum physics. They can take the messy, complex rules of how protons and neutrons interact (including the tricky forces that only happen when three particles are together) and turn them into a highly accurate wave function.

This is a big deal because it means scientists might not need to rely on the incredibly expensive and time-consuming "gold standard" calculations for every problem. Instead, they can use these neural networks to generate a near-perfect starting point, making the study of atomic nuclei faster and more efficient.

Summary

• The Problem: Predicting how tiny atomic nuclei behave is hard because the particles interact in complex groups, and old math tools are too rigid or slow.
• The Solution: The authors used Neural Networks (AI) to "learn" the perfect mathematical recipe for these interactions.
• The Innovation: The AI learned not just how pairs of particles interact, but how a third particle changes the game.
• The Outcome: The AI-generated recipe was almost as accurate as the most expensive, time-consuming method in physics, but it was found much faster. It proved that AI can be a powerful tool for solving fundamental problems in nuclear physics.

Technical Summary

Technical Summary: Neural-Network Correlation Functions for Light Nuclei with Chiral Two- and Three-Body Interactions

Problem Statement
Accurately determining the ground-state energies of light nuclei via ab initio quantum many-body techniques requires high-quality trial wave functions. While Green's Function Monte Carlo (GFMC) serves as a benchmark for such calculations, it relies heavily on the initial trial wave function to control the fermion-sign problem during imaginary-time evolution. Traditionally, these trial functions are constructed using Variational Monte Carlo (VMC) with parameterized correlation functions. However, optimizing these parameters is computationally demanding, particularly for systems involving three-body interactions where many-body effects beyond simple pairwise correlations must be accounted for. Furthermore, standard VMC frameworks often require solving complex second-order differential equations or neglecting intricate components of the nuclear Hamiltonian, such as full tensor, spin-orbit, and three-body terms derived from chiral effective field theory (EFT). There is a need for more efficient, expressive tools to generate trial wave functions that can capture the full complexity of microscopic nuclear forces.

Methodology
The authors propose a neural-network (NN) based approach to construct variational trial wave functions for light nuclei ($A \le 3$). The methodology integrates the following components:

• Wave Function Ansatz: The trial wave function $|\psi\rangle$ is constructed using a Jastrow-Slater form (Eq. 1) multiplied by additional three-body operators (Eq. 3) to account for three-body forces. The ansatz includes operators acting on spin-isospin states coupled with coordinate-dependent functions.
• Neural Network Architecture: The coordinate-dependent correlation functions are generated by two distinct neural networks, $NN_a$ and $NN_b$:
  • $NN_a$ processes the inter-particle distance $r_{\alpha\beta}$ to produce an intermediate feature $\tilde{R}_{\alpha\beta}$.
  • $NN_b$ generates all spatial functions (e.g., $f^{(c)}_{ij}$, $u^{(x)}_{ij}$, and three-body spatial functions) by taking as input the pairwise distance $r_{ij}$ and a summed feature representing the influence of all other particles ($\sum_{k \neq i,j} (\tilde{R}_{ik} + \tilde{R}_{jk})$). This structure explicitly allows the correlation between a pair to depend on the presence of a third particle, addressing three-body effects.
  • The networks utilize residual structures with ReLU activation functions.
• Training and Sampling: The networks are trained using VMC to minimize the energy expectation value, adhering to the Rayleigh-Ritz principle. To overcome the correlation length issues typical of classical Markov Chain Monte Carlo (Metropolis-Hastings), the authors employ Normalizing Flow (nflow) models. These models generate independent samples distributed according to $|\psi|^2$ via a series of invertible transformations, significantly improving sampling efficiency.
• Interactions: The study utilizes local chiral interactions at Next-to-Leading Order (N2LO) within chiral EFT. The interactions include full two-body and three-body forces, incorporating tensor, spin-orbit, charge-independence breaking (CIB), and charge-symmetry breaking (CSB) terms without simplifying pion-exchange contributions.

Key Results
The study focuses on the triton ($^3$H) and helium-3 ($^3$He) nuclei, as well as the deuteron ($^2$H) for two-body-only benchmarks.

• Performance on $^3$H: Using the "softest" N2LO chiral interaction ($(R_0, \Lambda) = (1.2 \text{ fm}, 1 \text{ GeV})$), the neural-network VMC achieved a ground-state energy of $E = -8.30 \pm 0.09$ MeV. This result is within 0.45% of the benchmark GFMC value ($E = -8.34$ MeV) and falls within one-half of a standard deviation of the GFMC result.
• Improvement over Standard VMC: Compared to standard parameterized VMC approaches, the neural-network method demonstrated a 91.2% improvement in capturing the ground-state energy (defined as the reduction in the gap between the parameterized VMC result and the GFMC result).
• Robustness: The method remained effective even with finer-resolution potentials ($R_0 = 1.0$ fm) and when including Coulomb interactions. For $^3$H and $^3$He, the neural network predictions fell within the one-sigma error bands of GFMC calculations (which include chiral EFT truncation uncertainties).
• Two-Body Systems: For $^2$H and $^3$H calculated without three-body interactions (using only $NN_b$), the results were nearly indistinguishable from GFMC within one-sigma uncertainty, demonstrating the network's capability to model complex two-body correlations.

Significance and Claims
The paper claims that neural networks offer a powerful framework for constructing effective trial wave functions for quantum Monte Carlo calculations. Specifically, the authors demonstrate that:

1. Neural networks can efficiently incorporate the full complexity of chiral two- and three-body interactions, including tensor and spin-orbit forces, without ignoring pion-exchange contributions.
2. The proposed architecture, which explicitly models the influence of surrounding particles on pairwise correlations, successfully captures the majority of the ground-state energy estimated by GFMC at the variational level alone.
3. The combination of neural-network wave functions with normalizing flow sampling provides a computationally efficient alternative to traditional parameterized VMC, yielding highly accurate results for light nuclei ($A \le 3$).

The authors note that while the current antisymmetrization limits the application to $A \le 4$ nuclei, the framework is extensible to larger systems by utilizing neural networks to construct Slater determinants. The work establishes the potential of machine learning to solve the variational optimization problem in nuclear physics with high expressiveness and accuracy.