Addressing the ground state of the deuteron by physics-informed neural networks

This paper demonstrates that Physics-Informed Neural Networks (PINNs) can accurately solve the many-body Schrödinger equation for the deuteron ground state using realistic nucleon-nucleon interactions, achieving a relative binding energy error of approximately $10^{-6}$ compared to established numerical benchmarks.

The Gist

Imagine you are trying to find the perfect recipe for a cake, but you don't have the recipe card. You only have a few rules: "It must taste sweet," "It must not burn," and "It must hold its shape."

In the world of physics, scientists are trying to find the "recipe" for the most basic building blocks of matter: the Deuteron. A deuteron is a tiny, stable pair made of one proton and one neutron stuck together. It's the simplest "nucleus" in the universe.

For decades, physicists have used complex math and supercomputers to figure out exactly how these particles behave. But this new paper introduces a clever new tool: Physics-Informed Neural Networks (PINNs). Think of this as teaching a computer to be a "smart guesser" that learns the laws of physics while it tries to solve the puzzle.

Here is the story of what they did, explained simply:

1. The Old Way vs. The New Way

• The Old Way (Traditional Math): Imagine trying to solve a maze by drawing every single possible path on a giant piece of paper. It takes forever, and if the maze gets too big (like a complex atom), the paper runs out.
• The New Way (PINNs): Imagine giving a robot a map of the maze's rules (the laws of physics) and saying, "Find the exit." The robot doesn't need to see the whole maze at once. It just needs to know the rules: "Don't hit the walls," "Keep moving forward," and "Stay inside the lines." The robot learns the path by trial and error, guided by the rules.

2. The "Smart" Robot (The Neural Network)

The authors built a digital brain (a Neural Network) to act as this robot. But instead of just guessing, they "informed" it with physics.

• The Rules of the Game: They told the robot: "You must follow the Schrödinger equation" (the master rulebook for how tiny particles move).
• The Goal: The robot's job is to find the state where the deuteron is most stable (the "ground state"). In our cake analogy, this is finding the exact mix of ingredients that makes the cake perfect without burning.

3. The Two Different Kitchens (Position vs. Momentum)

The team tested their robot in two different "kitchens" (mathematical spaces):

• Kitchen A (Position Space): This is like looking at the deuteron as if it were a physical object sitting on a table. They used a simplified model (the "Minnesota potential") to see if the robot could learn the basics.

  • Result: The robot learned quickly and got the recipe right, but with a tiny bit of error (about 1% off). It was a good practice run.

• Kitchen B (Momentum Space): This is a much harder kitchen. Here, they looked at the deuteron not by where it is, but by how fast its parts are moving. They used two very advanced, realistic models (N4LO and CD-Bonn) that include "high-speed" particles.

  • The Challenge: One of these models (CD-Bonn) is like a recipe that requires ingredients at incredibly high speeds. It's very "spiky" and hard to predict.
  • The Result: The robot was amazing. For the most difficult model, it got the answer right to six decimal places. It was so accurate that the difference between the robot's answer and the "perfect" math answer was almost zero.

4. How Did They Teach the Robot?

The robot learns by minimizing a "Loss Function." Think of this as a scorecard.

• If the robot's guess violates the laws of physics (like the particles flying apart when they should stick), the score gets a big penalty.
• If the robot's guess doesn't match the known boundaries (like the particles disappearing at the edge of the universe), it gets another penalty.
• The robot adjusts its "brain" millions of times to lower the penalty score until it finds the perfect solution.

5. Why Does This Matter?

This paper is a big deal because:

1. It Works: It proves that AI can solve the hardest problems in nuclear physics, not just simple ones.
2. It's Efficient: The robot found the answer with incredible precision, even for models that are notoriously difficult to solve.
3. The Future: Right now, the robot only solved the simplest nucleus (the deuteron). But if this method works, scientists can use it to solve the "recipes" for heavier, more complex atoms and even molecules. This could help us understand how stars burn or how to create new materials.

The Bottom Line

The authors took a "smart guesser" (AI), taught it the strict laws of nuclear physics, and asked it to figure out how a proton and neutron hold hands. The AI didn't just guess; it learned the physics so well that it solved the problem with near-perfect accuracy. It's like teaching a student the laws of gravity and then asking them to calculate the orbit of a satellite, and they get it right on the first try.

This opens the door to using AI as a powerful new microscope for the universe, helping us understand the very fabric of reality.

Technical Summary

1. Problem Statement

The paper addresses the challenge of solving the quantum many-body Schrödinger equation for atomic nuclei, specifically focusing on the deuteron (a bound state of one proton and one neutron). Traditional methods like Variational Monte Carlo (VMC) and Diffusion Monte Carlo (DMC) are computationally expensive. While Neural Network Quantum States (NQS) have shown promise, they typically rely on reinforcement learning without explicit physical constraints beyond the variational principle.

The authors aim to demonstrate the feasibility of using Physics-Informed Neural Networks (PINNs) to solve the deuteron ground state problem. The specific challenges addressed include:

• Handling realistic nucleon-nucleon interactions in momentum space, which involve strong high-momentum correlations.
• Solving integro-differential equations where the Hamiltonian includes non-local potential terms.
• Achieving high accuracy (benchmarked against exact numerical diagonalization) without labeled training data.

2. Methodology

The authors implement a PINN framework tailored for eigenvalue problems. The core approach involves minimizing a composite loss function that encodes the physical laws of the system directly into the neural network training.

A. Loss Function Construction

The total loss function ($L_{PINN}$) is a weighted sum of five components:

1. $L_{Schrod}$ (PDE Loss): Enforces the Schrödinger equation $H|\psi\rangle = E|\psi\rangle$. It calculates the mean squared error between the Hamiltonian acting on the network output and the energy eigenvalue times the output.
2. $L_{BCs}$ (Boundary Conditions): Ensures the wavefunction vanishes at the boundaries ($x=0$ and $x=M$) and satisfies specific conditions for partial waves (e.g., $\phi(x) = x\psi(x)$ for $L=0$).
3. $L_{norm}$ (Normalization): Enforces the condition $\int |\psi|^2 dV = 1$.
   • Coordinate Space: Uses an auxiliary output method where a secondary network output integrates the probability density (mesh-free).
   • Momentum Space: Uses finite difference methods for direct integration (more computationally efficient per epoch).
4. $L_{var}$ (Variational Principle): A custom term designed to guide the network toward the lowest energy state. It includes a decay parameter that diminishes over training epochs to prevent the network from getting stuck in local minima once the correct energy is approached.
5. $L_{int}$ (Integration Loss): Specific to the auxiliary output method, ensuring the auxiliary output correctly represents the integral of the probability density.

B. Network Architecture

• Type: Feed-forward neural networks.
• Inputs: Radial coordinate ($r$) in position space or relative momentum ($k$) in momentum space.
• Outputs:
  • Position Space: Wavefunction $\psi(r)$ and auxiliary integral $\nu(r)$.
  • Momentum Space: Components $\phi_S(k)$ and $\phi_D(k)$ corresponding to the $^3S_1$ and $^3D_1$ partial waves.
• Differentiation: Automatic differentiation is used to compute derivatives required for the kinetic energy term in the Hamiltonian.

C. Models Tested

The study evaluates three interaction models:

1. Minnesota Potential: A simplified coordinate-space model (no angular momentum mixing).
2. N4LO ($\chi$EFT): A high-precision chiral effective field theory interaction in momentum space (5th order), with a cutoff of 550 MeV/c.
3. CD-Bonn: A high-precision potential in momentum space designed to induce strong short-range (high-momentum) correlations, requiring a much larger momentum range (up to ~8.5 GeV/c).

3. Key Contributions

• First Application to Ab Initio Nuclear Structure: This is the first demonstration of using PINNs to solve an ab initio nuclear structure problem (the deuteron) with realistic interactions.
• Momentum Space Implementation: The authors successfully tackled momentum-space interactions, which are integral-differential in nature and computationally more demanding than coordinate-space differential equations.
• Novel Variational Loss Strategy: They introduced a specific loss term ($L_{var}$) that dynamically balances the search for the ground state, allowing the network to converge to the correct eigenvalue without prior knowledge of the energy.
• Dual Normalization Strategies: The paper compares and implements two distinct normalization techniques (auxiliary output vs. finite difference) to optimize performance in different domains.

4. Results

The PINN results were benchmarked against exact numerical diagonalization and experimental values.

| Interaction Model | Domain | Relative Error (vs. Numerical Benchmark) | Fidelity ($F_\psi$) | Observations |
| :--- | :--- | :--- | :--- :--- |
| Minnesota | Position | $1.1\%$ | High | Converged to ~ -2 MeV. Demonstrated feasibility but lower precision due to model simplicity. |
| N4LO | Momentum | $-5.21 \times 10^{-5}$ | $0.9999933$ | Achieved high accuracy. Training required $\sim 4 \times 10^5$ epochs but was faster per epoch due to finite difference normalization. |
| CD-Bonn | Momentum | $-2.76 \times 10^{-7}$ | $1 - 10^{-9}$ | Highest Accuracy. Successfully handled strong high-momentum correlations. The error relative to experiment was $-6.01 \times 10^{-4}$, indicating the error source is the potential model, not the PINN method. |

• Training Dynamics: The training process showed distinct phases: initial rapid energy drop, a plateau where boundary conditions were refined (scattering states removed), and final asymptotic convergence.
• Efficiency: While momentum space required more epochs than coordinate space, the computational cost per epoch was significantly lower (approx. 1/10th) due to the efficiency of the finite difference integration method.

5. Significance and Future Directions

• Validation of PINNs in Nuclear Physics: The study proves that PINNs can achieve relative errors on the order of $10^{-6}$ to $10^{-7}$ for nuclear eigenstates, rivaling or exceeding traditional numerical methods in terms of precision for specific benchmarks.
• Scalability: The framework is not limited to ground states; the authors note its potential for excited states and scattering solutions by adapting boundary condition losses.
• Future Work:
  • Extending the implementation from 1D radial equations to full 3D space.
  • Incorporating three-nucleon interactions.
  • Applying the method to heavier nuclei and molecules.
  • Integrating standard variational ansätze (like Slater determinants) to handle the Pauli exclusion principle for multi-nucleon systems.

In conclusion, this work establishes a robust foundation for using Physics-Informed Neural Networks as a powerful tool for ab initio nuclear theory, offering a pathway to solve complex many-body problems that are currently computationally prohibitive.