Neural Wavefunction Calculations of μSR Spectra with Quantum Muons and Protons

This study demonstrates that using variational quantum Monte Carlo with neural-network trial wavefunctions to explicitly treat muons as quantum particles significantly improves the accuracy of predicted muon hyperfine constants compared to standard density functional theory methods that treat muons as fixed classical particles.

The Gist

Imagine you are trying to take a photograph of a tiny, invisible dancer (a muon) spinning inside a crowded room full of other dancers (electrons). You want to know exactly how fast the muon is spinning relative to its neighbors. This "spin relationship" is called the hyperfine constant, and it's the secret code scientists use to understand the magnetic personality of materials using a technique called muon spin spectroscopy (µSR).

For a long time, scientists tried to predict this code using a method called Density Functional Theory (DFT). But they had a major problem: they treated the muon like a statue. They assumed the muon was frozen in place, a solid, classical object that didn't move or wiggle.

The Problem with the Statue
In reality, a muon is not a statue. It's a quantum particle, meaning it's fuzzy, jittery, and constantly dancing around due to "zero-point motion" (a fundamental quantum jitter that never stops). It's about 200 times heavier than an electron but still light enough to be very wobbly.

By treating the muon as a frozen statue, the old DFT method was like trying to predict the path of a bouncing ball by pretending the ball is glued to the floor. It gave okay results sometimes, but it missed the true, chaotic nature of the dance.

The New Solution: The Neural Network Dance Floor
This paper introduces a new, high-tech way to solve the problem. The authors, Jamie Carr and his team, used a Neural Network (a type of artificial intelligence) to build a "wavefunction."

Think of a wavefunction as a 3D map of probability. Instead of saying "the muon is here," the map says, "there is a 10% chance the muon is here, a 20% chance it's there, and it's mostly dancing in this cloud."

They used a specific AI architecture called Psiformer. Imagine this AI as a super-smart choreographer who doesn't just watch the muon; it watches the entire dance floor at once. It understands that:

1. The muon is dancing.
2. The electrons are dancing.
3. The muon and electrons are reacting to each other's moves in real-time.

The Experiment: Methyl and Ethyl Radicals
The team tested this new method on two chemical "dancers": the muoniated methyl radical and the muoniated ethyl radical. These are molecules where a muon has attached itself to a carbon chain.

They ran three types of simulations:

1. The Statue (Classical Muon): The muon is frozen. (Like the old DFT method).
2. The Jitterbug (Quantum Muon): The muon is allowed to dance and wiggle, but the rest of the molecule is still.
3. The Full Party (Full Quantum): The muon dances, and the nearby hydrogen atoms dance too.

The Results: Why the AI Won
When they compared their results to real-world experiments, the "Statue" method was off by a significant margin. It was like predicting a dancer's speed while ignoring their jumps.

However, the Quantum Muon and Full Quantum methods (the AI approach) were much closer to reality.

• The Methyl Radical: The AI predicted the spin interaction was about 8-11% closer to the real experiment than the old methods.
• The Ethyl Radical: The AI was even better, getting within 3% of the experimental value.

The "Environmental" Twist
The paper also notes that real experiments happen in a "crowded room" (like a liquid or a mineral crystal), which pushes the dancers around. The AI results were so accurate that when the team adjusted for this "crowd pressure," their predictions matched the real-world data almost perfectly.

Why This Matters
This is a big deal because:

• Accuracy: It proves that to understand these tiny particles, you must treat them as quantum dancers, not frozen statues.
• Speed vs. Power: While this AI method is computationally expensive (it takes days on powerful supercomputers), it is far more accurate than the "cheap" DFT method for these specific problems.
• Future Tools: It suggests that in the future, scientists will use these "Neural Wavefunctions" as a standard tool to decode the secrets of new materials, from batteries to superconductors.

In a Nutshell
The authors replaced the old, rigid "statue" model of the muon with a flexible, AI-driven "quantum cloud" model. By letting the muon dance and wiggle in their calculations, they finally got the math to match the real world, showing that when it comes to the quantum realm, you can't just freeze time—you have to let the particles move.

Technical Summary

Here is a detailed technical summary of the paper "Neural Wavefunction Calculations of µSR Spectra with Quantum Muons and Protons."

1. Problem Statement

Muon Spin Rotation/Relaxation/Resonance ($\mu$SR) spectroscopy is a powerful technique for probing local magnetic environments in materials. A key measurable quantity is the muon hyperfine constant ($A_\mu$), which depends on the electron spin density at the muon's location.

• Limitation of Standard Methods: Conventional computational approaches, primarily Density Functional Theory (DFT), typically treat the muon as a fixed classical particle (Born-Oppenheimer approximation). This approach fails to account for the muon's significant quantum mechanical nature (mass $\approx 207 m_e$), specifically its zero-point motion and delocalization.
• The Correlation Gap: While DFT can approximate electron densities, it struggles to accurately describe electron-muon pair correlations because it is a mean-field theory that does not directly yield two-particle correlation functions (pair density functions).
• Goal: To develop a method that treats both electrons and muons as quantum particles on equal footing to accurately calculate the muon-electron pair density and, consequently, the hyperfine constant, surpassing the accuracy of standard DFT.

2. Methodology

The authors employ Variational Quantum Monte Carlo (VMC) using Neural Network Wavefunctions (specifically the Psiformer architecture) to solve the many-body Schrödinger equation.

• Hamiltonian: The system is described by the Coulomb Hamiltonian plus the hyperfine interaction. Since the Coulomb interaction is orders of magnitude stronger, the ground state is first determined using the Coulomb Hamiltonian, and the hyperfine constant is calculated via first-order perturbation theory.
• Neural Wavefunction Ansatz:
  • The trial wavefunction $\Psi(R)$ is constructed using a Psiformer architecture, which utilizes a transformer-based neural network to generate many-particle orbitals.
  • The wavefunction includes a Jastrow factor (via a small neural network) to capture short-range correlations.
  • It is explicitly anti-symmetric to satisfy the Pauli exclusion principle.
  • Crucially, the wavefunction depends on the coordinates of both electrons and the muon, treating them as quantum particles within the same many-body framework.
• Calculation of $A_\mu$:
  • The hyperfine constant is proportional to the spin-resolved system-averaged pair density at zero separation: $\Delta\rho(0) = \bar{\rho}_{\uparrow\uparrow}(0) - \bar{\rho}_{\uparrow\downarrow}(0)$.
  • The authors use Monte Carlo integration to sample particle configurations.
  • Because the probability of finding the muon and electron at exactly the same point is zero in a discrete simulation, they calculate the pair density at small separations ($r < 0.5 a_0$) and extrapolate to $r=0$.
  • The extrapolation uses a quadratic fit in the logarithmic domain, constrained by the generalized Kato cusp condition to ensure physical accuracy near the singularity.

3. Systems Studied

The method was applied to two muoniated radicals:

1. Muoniated Methyl Radical ($CH_2Mu$): 1 muon, 9 electrons.
2. Muoniated Ethyl Radical ($C_2H_4Mu$): 1 muon, 17 electrons.

Three levels of approximation were compared for each system:

• Classical Muon: Muon treated as a fixed point (standard DFT-like approach).
• Quantum Muon: Muon treated as a quantum particle; nuclei (C, H) fixed.
• Full Quantum: Muon and Hydrogen nuclei treated as quantum particles; only Carbon fixed (applied only to Methyl radical).

4. Key Results

A. Validation on Muonium

The method was first validated on the Muonium atom (Mu). The Quantum Muon approach yielded a hyperfine constant of 4469(7) MHz, closely matching the experimental value of 4463 MHz. In contrast, the Classical Muon approximation yielded 4529 MHz, demonstrating the failure of the fixed-particle approximation even for simple systems.

B. Muoniated Methyl Radical ($CH_2Mu$)

• Classical Muon (Psiformer): $-228(3)$ MHz.
• Quantum Muon (Psiformer): $-223(3)$ MHz.
• Full Quantum (Psiformer): $-218(3)$ MHz.
• DFT (B3LYP): $-188$ MHz.
• Experiment (73 K): $\pm 201.30(10)$ MHz.

Analysis:

• The Full Quantum result ($-218$ MHz) is the closest to the experimental magnitude (deviation of $\sim 8\%$).
• The Classical Muon result deviates significantly from DFT ($40$ MHz difference), highlighting that even when treating the muon classically, the neural wavefunction captures electron correlations better than DFT.
• When correcting for environmental effects (liquid ketene), the Full Quantum result aligns even more closely with the estimated vacuum experimental value.

C. Muoniated Ethyl Radical ($C_2H_4Mu$)

• Classical Muon (Psiformer): $466(4)$ MHz.
• Quantum Muon (Psiformer): $537(4)$ MHz.
• DFT (O3LYP): $478.9 - 493.8$ MHz.
• Experiment (5-6 K in silicates): $\pm 530 - 542$ MHz.

Analysis:

• The Quantum Muon result ($537$ MHz) is in excellent agreement with experiment (within error bars).
• The Classical Muon result ($466$MHz) is significantly lower than both the Quantum result and experiment, underestimating the hyperfine constant by$\sim 15%$.
• The shift from Classical to Quantum is attributed to the muon's zero-point motion shifting its probability density into regions of higher positive spin density.

5. Significance and Contributions

1. Demonstration of Quantum Nuclear Effects: The paper provides definitive evidence that treating the muon as a quantum particle is essential for accurate $\mu$SR predictions. The "zero-point motion" of the muon significantly alters the electron spin density at the muon site, a factor missed by standard DFT and classical approximations.
2. Superiority over DFT: Even when comparing "Classical Muon" treatments, the neural wavefunction approach yields results distinct from DFT, suggesting that DFT's mean-field approximation introduces errors in electron-muon correlations that the neural network captures via explicit many-body correlations.
3. Accurate Pair Density: The method successfully computes the muon-electron pair density function directly, bypassing the need for approximate correlation factors used in two-component DFT.
4. Scalability and Future Outlook: While the computational cost ($O(N^{3-4})$) is higher than DFT, it is feasible for systems with $\sim 100$ electrons. The authors propose a hybrid strategy: using DFT for geometry optimization and neural wavefunctions for high-accuracy hyperfine constant evaluation.
5. Impact on Material Science: This approach offers a path to more reliable interpretation of $\mu$SR data, enabling better identification of chemical species and local magnetic structures in complex materials, free radicals, and solids.

Conclusion

The authors successfully demonstrate that neural-network-based variational Monte Carlo provides a highly accurate framework for calculating muon hyperfine constants. By explicitly treating the muon as a quantum particle within the many-body wavefunction, the method captures critical quantum effects (delocalization and zero-point motion) that standard DFT and classical approximations miss, resulting in predictions that align significantly better with experimental $\mu$SR data.