Constructing Inverse Potentials from Scattering Phase Shifts using Physics-Informed Neural Networks: Application to Neutron-Alpha Scattering

This paper presents a physics-informed neural network framework that successfully reconstructs the neutron-alpha scattering potential by embedding hard structural constraints and differentiable numerical solvers, thereby accurately recovering the $P_{3/2}$ resonance parameters and demonstrating the reliability of machine learning for nuclear inverse scattering problems.

The Gist

Imagine you are a detective trying to figure out what a hidden object looks like, but you can't see the object itself. All you have are the ripples it creates when you throw pebbles at it. In the world of nuclear physics, scientists do this all the time: they shoot neutrons at tiny atomic cores (like the alpha particle, which is the nucleus of a helium atom) and watch how the neutrons bounce off. The way they bounce—specifically, the angle and timing—tells them about the invisible "force field" or potential that exists between the neutron and the core.

The challenge is the Inverse Problem: It's easy to predict how a pebble bounces if you know the shape of the rock it hits. But figuring out the exact shape of the rock just by looking at the ripples? That's incredibly difficult. Many different shapes could create the same ripples, making the answer unstable and confusing.

This paper introduces a new, clever detective tool called Physics-Informed Neural Networks (PINNs) to solve this puzzle for the first time in this specific context. Here is how they did it, explained simply:

1. The "Smart" Detective (The Neural Network)

Usually, scientists guess a shape for the force field (like a specific mathematical curve) and tweak the numbers until the ripples match the experiment. This paper used a Neural Network, which is like a super-flexible, digital clay model. Instead of guessing a fixed shape, the network can mold itself into any shape it wants to fit the data.

2. The Crucial Rule: The "Finite-Range" Envelope

Here is the paper's biggest breakthrough. In nuclear physics, there is a hard rule: the force between a neutron and an alpha particle must disappear completely once you get far enough away. It's like a magnet; if you pull it far enough away, the pull becomes zero. It doesn't just get weak; it stops.

• The Mistake: The authors tried letting the neural network guess the shape freely. The network, being a "lazy" optimizer, tried to cheat. It created a force field that never quite reached zero, leaving a tiny, invisible "tail" of force stretching out to infinity. Even though the math looked okay, the physics was wrong, and the predictions failed.
• The Fix: The authors built the "zero-force" rule directly into the network's architecture. They wrapped the neural network's output in a Gaussian envelope (think of it as a soft, invisible cage that forces the clay to flatten out to zero at a specific distance).
  • Analogy: Imagine trying to sculpt a mountain that must be perfectly flat at the horizon. If you just tell the sculptor, "Try to make it flat," they might leave a tiny bump. If you build a giant, flat floor under the clay and say, "The clay must sit on this floor," the sculptor has no choice but to make it flat. This "hard constraint" was the key to success.

3. The Training Process

The team fed the network real experimental data (how neutrons bounced at different energies). The network then:

1. Made a guess at the force field shape.
2. Ran a simulation (using a mathematical recipe called the "variable-phase equation") to see what ripples that shape would create.
3. Compared its ripples to the real data.
4. Adjusted its internal "clay" to reduce the error.

Because the "zero-force" rule was built into the structure, the network didn't waste time trying to fix impossible shapes. It converged quickly and smoothly to a solution.

4. What They Found

The network successfully reconstructed the invisible force field. Here is what the "sculpture" looked like:

• The Shape: It turned out to be a smooth, purely attractive "well" (like a bowl). There was no repulsive core (no "hard bump" in the middle), which makes sense because the alpha particle is a tight, stable bundle of protons and neutrons.
• The Resonance: When they added the physics of spinning (centrifugal force) to this bowl, it created a barrier-well structure. Imagine a valley with a hill around the edge. A neutron can get trapped in the valley for a moment before rolling over the hill and escaping. This "trapping" explains a famous phenomenon called the P3/2 resonance, where neutrons linger briefly before bouncing off.
• The Numbers: The depth of this valley and the height of the hill matched experimental expectations almost perfectly. The calculated "resonance energy" (how long the neutron stays trapped) was 0.95 MeV, which is very close to the known experimental value of 0.92 MeV.

5. Why It's Reliable

To make sure this wasn't just a lucky guess, the authors ran three stress tests:

• Starting Over: They restarted the training 10 times with different random starting points. Every time, the network found the exact same shape. This means the solution is unique and stable, not a fluke.
• Time Check: They stopped training early and late. The shape settled down perfectly after a certain point and didn't change much after that.
• The "One Missing" Test: They removed one single data point from the training set and retrained. They did this 22 times (removing each point once). The resulting shapes were almost identical every time. This proves that no single "bad" data point was controlling the whole result; the network learned the true physics from the whole picture.

Summary

This paper shows that by teaching a computer the fundamental rules of physics (like "the force must stop at a certain distance") before it starts learning, rather than just asking it to be polite about it, we can solve incredibly difficult nuclear puzzles. The result is a clear, smooth, and accurate map of the invisible forces inside the nucleus, derived entirely from how particles scatter.

Technical Summary

1. Problem Statement

The paper addresses the inverse scattering problem in nuclear physics: reconstructing the interaction potential $V(r)$ between nucleons and nuclei from measured scattering phase shifts $\delta(E)$.

• Challenges: The problem is inherently ill-posed, meaning multiple distinct potentials can reproduce the same finite set of phase shift data. Traditional methods (e.g., phenomenological Woods-Saxon models, R-matrix theory, or integral equation inversions like Gel'fand-Levitan) suffer from model bias, sensitivity to unmeasured energy ranges, or computational complexity.
• Specific Case: The authors focus on the neutron-alpha ($n-\alpha$) elastic scattering system, specifically the $P_{3/2}$ partial wave. This system is critical due to the well-known $^5$He resonance near 0.92 MeV, which serves as a sensitive test for P-wave interactions.
• Gap: While Machine Learning (ML) has been applied to nuclear physics, standard Physics-Informed Neural Networks (PINNs) often treat physical constraints (like finite range) as "soft" penalties in the loss function. The authors hypothesize that for nuclear scattering, these constraints must be hard-coded into the architecture to ensure physically meaningful solutions.

2. Methodology

The authors developed a custom PINN framework that integrates the variable-phase approach with a specific neural network architecture designed to enforce physical boundary conditions.

A. Physics-Informed Framework

• Governing Equation: Instead of solving the radial Schrödinger equation directly, the authors use Calogero's variable-phase approach. This recasts the problem into a first-order nonlinear initial-value problem:
  $$\frac{d\delta(r; E)}{dr} = -\frac{1}{k} V(r) \left[ \cos \delta(r; E) \hat{j}_\ell(kr) - \sin \delta(r; E) \hat{n}_\ell(kr) \right]^2$$
  where $\delta(r; E)$ is the radial phase function, $k$ is the wave number, and $\hat{j}, \hat{n}$ are Riccati-Bessel functions.
• Differentiability: The integration is performed using a 4th-order Runge-Kutta scheme. Crucially, every step is differentiable, allowing the use of automatic differentiation to compute gradients of the final phase shift with respect to the neural network parameters. This creates an end-to-end differentiable pipeline.

B. Neural Network Architecture & Hard Constraints

The core innovation lies in how the potential $V(r)$ is represented:
$$V_\theta(r) = N_\theta(r) \times \exp\left[ -\left(\frac{r}{R}\right)^2 \right]$$

• $N_\theta(r)$: A feed-forward network (2 hidden layers, 64 nodes, tanh activation) that learns the shape of the potential.
• Gaussian Envelope: The network output is multiplied by a Gaussian envelope with scale $R=3.0$ fm.
  • Significance: This architecturally enforces the finite-range condition ($V(r) \to 0$ as $r \to \infty$). The authors demonstrate that without this envelope, the optimizer produces potentials with non-vanishing tails, leading to systematic phase shift errors and failure to converge, regardless of training duration.
• Input Scaling: The radial coordinate is rescaled ($r/6$) to keep inputs within the linear regime of the tanh activation function.

C. Loss Function

The total loss $L$ combines four terms:

1. Data Fidelity ($L_{data}$): Weighted mean squared error between model and experimental phase shifts. Weights are higher near the resonance energy to prioritize the most dynamic region.
2. Smoothness ($L_{smooth}$): Penalizes the second finite difference of the potential to prevent unphysical high-frequency oscillations.
3. Finite-Range Reinforcement ($L_{range}$): A soft penalty for $r > 4$ fm to accelerate convergence (complementing the hard architectural constraint).
4. Low-Energy Threshold ($L_{low}$): Ensures the low-energy behavior follows effective-range theory ($\delta/k^3 \to \text{const}$) without fixing the scattering volume a priori.

D. Training & Validation

• Optimizer: Adam optimizer with a fixed learning rate of $10^{-4}$.
• Data: 22 experimental phase shift points from Satchler et al. (0.3–18 MeV).
• Robustness Tests:
  • Initialization Sensitivity: 10 different random seeds.
  • Epoch Sensitivity: Monitoring convergence up to 10,000 epochs.
  • Leave-One-Out (LOO): Retraining 22 times, each time removing one data point to test stability.

3. Key Results

A. Reconstructed Potential

• Central Potential: The learned potential $V_\theta(r)$ is purely attractive and smooth, with a minimum depth of $-60.47$ MeV at $r \approx 1.71$ fm. It vanishes exponentially beyond $r \approx 4$ fm. No repulsive core was found, consistent with the Pauli blocking effects in the $P_{3/2}$ channel.
• Effective Potential: Adding the centrifugal barrier ($\ell=1$) creates a barrier-well structure:
  • Well Depth: $-13.60$ MeV at $r = 1.71$ fm.
  • Barrier Height: $+2.03$ MeV at $r = 4.98$ fm.
  • Physical Interpretation: This topology explains the $P_{3/2}$ resonance as a quasi-bound state formed by a neutron temporarily trapped behind the centrifugal barrier.

B. Resonance Parameters

From the reconstructed potential, the resonance parameters for the $^5$He system were extracted:

• Resonance Energy ($E_r$): $0.95$ MeV (Experimental: $0.92 \pm 0.04$ MeV).
• Resonance Width ($\Gamma_r$): $0.78$ MeV (Experimental FWHM $\approx 1.2$ MeV; R-matrix $\approx 0.64$ MeV).
• The results fall naturally between R-matrix and experimental values, validating the approach.

C. Effective-Range Parameters

The model successfully reproduced low-energy P-wave parameters without explicit constraints:

• Scattering Volume ($a$): $-62.18$ fm$^3$ (Reference: $-62.95$ fm$^3$; error < 1.2%).
• Effective Range ($r$): $-0.87$ fm (Reference: $-0.88$ fm; error < 1.7%).

D. Robustness

• Convergence: The loss converged smoothly to $\approx 3 \times 10^{-4}$ within 6,000 epochs.
• LOO Analysis: The reconstructed potential remained stable even when any single data point was removed. The $\pm 1\sigma$ spread in the potential was negligible (sub-hundredths of an MeV), proving the solution is not driven by outliers.
• Initialization: Results were identical across 10 random seeds, suggesting a unimodal loss landscape.

4. Key Contributions

1. Architectural Hard Constraints: The paper establishes that for nuclear inverse problems, finite-range conditions must be embedded in the network architecture (via the Gaussian envelope) rather than treated as soft penalties. This is shown to be a necessary condition for convergence to a physical solution.
2. End-to-End Differentiable Pipeline: The integration of the variable-phase equation with automatic differentiation allows for direct gradient-based optimization of the potential, avoiding the need for metaheuristic algorithms (like genetic algorithms) which lack analytical gradients.
3. Data-Driven Potential Reconstruction: The method successfully reconstructs a smooth, physically interpretable potential from sparse experimental data, recovering resonance parameters and effective-range parameters with high accuracy.
4. Validation Framework: The comprehensive use of LOO analysis and initialization sensitivity tests provides a rigorous benchmark for the reliability of PINNs in nuclear physics.

5. Significance

This work demonstrates that Physics-Informed Neural Networks are a robust, reliable, and interpretable tool for solving inverse scattering problems in nuclear physics.

• Beyond Theory: It moves beyond simple interpolation, providing a method to extract fundamental interaction potentials directly from data.
• Methodological Lesson: It highlights a critical design principle for PINNs: hard physical constraints (boundary conditions) should be architectural features, not just loss function penalties.
• Future Applications: The framework is scalable and can be extended to include spin-orbit coupling, heavier nuclei, and systems with more complex reaction channels, offering a new pathway for nuclear data evaluation and ab initio benchmarking.