Real-time Dynamics in 3D for up to 1000 Qubits with Neural Quantum States: Quenches and the Quantum Kibble--Zurek Mechanism

This paper introduces a scalable residual-based convolutional Neural Quantum State architecture that enables the first large-scale simulation of real-time dynamics in 3D quantum systems up to 1000 qubits, successfully demonstrating the 3D quantum Kibble-Zurek mechanism and its unique logarithmic corrections at the critical point.

The Gist

Imagine you are trying to predict how a massive crowd of people will move if you suddenly shout a command. If the crowd is just a single line (1D), it's easy to track. If they are in a square room (2D), it gets harder. But if they are in a giant, multi-story skyscraper (3D), with thousands of people rushing, bumping, and influencing each other in every direction, predicting their exact movement becomes a nightmare for even the smartest computers.

This is the challenge physicists face with quantum systems. When you have thousands of tiny particles (qubits) interacting in 3D space, the math explodes in complexity. Traditional computers simply run out of memory before they can finish the calculation.

This paper is about a breakthrough: Teaching a computer to "dream" its way through this complexity.

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

1. The Problem: The "Spaghetti" of Reality

In the quantum world, particles are entangled. Think of them as a giant bowl of spaghetti where every noodle is tied to every other noodle. If you pull one, the whole bowl moves.

• The Old Way: Traditional computers try to map every single noodle. In 3D, with 1,000 noodles, the map is so huge it would fill the universe.
• The New Way (Neural Quantum States): Instead of mapping every noodle, the researchers used an Artificial Neural Network (AI). Think of the AI as a super-smart chef who has tasted the spaghetti a million times. The chef doesn't need to see every noodle to know how the whole bowl will react if you stir it. The AI learns the patterns and feel of the quantum system, allowing it to simulate 1,000 particles at once.

2. The Tool: A 3D "Brain" for Cubes

The researchers built a specific type of AI architecture (a "ResNet-CNN") designed specifically for 3D cubes.

• The Analogy: Imagine trying to understand a 3D object. A standard 2D camera (like a phone) can only see a flat picture. This new AI is like a 3D X-ray vision that understands depth, width, and height simultaneously. It respects the symmetry of the cube, meaning it knows that moving a particle from the left corner to the right corner is just a shift, not a totally new universe.

3. The Experiments: Two Ways to Stir the Pot

They tested this AI on a model called the Transverse-Field Ising Model. Imagine a grid of tiny magnets (spins) that can point up or down. They did two things:

A. The Sudden Shock (The "Drop"):
They started with all magnets pointing one way, then suddenly flipped the rules so they had to point the other way.

• What happened: The magnets didn't just flip instantly. They wobbled, collapsed, and then bounced back (like a spring). This is called "collapse-and-revival."
• The Result: The AI perfectly predicted these wobbles and showed that the magnets became deeply "entangled" (tied together) in complex groups of 30 or more. This is a level of detail no previous computer could reach for 3D systems.

B. The Slow Push (The "Kibble-Zurek Mechanism"):
This is the big discovery. Instead of a sudden shock, they slowly turned the dial from one state to another, passing right through a "critical point" (a tipping point where the system changes phase, like water turning to ice).

• The Analogy: Imagine driving a car toward a cliff. As you get closer, the road gets foggy and slippery. You try to steer, but the car's reaction time lags behind your steering wheel. There is a specific moment where you stop being able to react in time.
• The Physics: In quantum physics, this lag creates "defects" or glitches in the system. This is the Quantum Kibble-Zurek Mechanism (QKZM). It's like the universe getting a "traffic jam" because it couldn't change fast enough.
• The 3D Twist: Usually, these traffic jams follow simple math rules. But in 3D, the rules are weird. They are modified by logarithms (a type of math that grows very slowly, like a snail). It's like the traffic jam isn't just a line; it's a line with a weird, slow-growing spiral attached to it.

4. The Big Win: Proving the "Snail Math"

The researchers didn't just simulate the crash; they proved the math behind it.

• They derived a new, ultra-precise formula that accounts for those "snail-like" logarithmic corrections.
• They ran their AI simulation on grids up to 10x10x10 (1,000 qubits).
• The Result: When they plotted the data, everything collapsed onto a single, perfect curve. The AI's "dream" matched the complex math perfectly.

Why Does This Matter?

1. It's a New Superpower: This proves we can now simulate 3D quantum systems with 1,000 particles. Before, this was impossible.
2. It's a Benchmark: As we build real quantum computers (quantum simulators), we need a way to check if they are working correctly. This paper provides the "answer key" for 3D quantum experiments.
3. It Solves a Mystery: It confirms how the universe behaves when it changes states in 3D, specifically how "entanglement" (the spooky connection between particles) spreads during these changes.

In a nutshell:
The authors built a 3D AI chef that can taste a quantum soup of 1,000 ingredients and predict exactly how it will boil. They used it to prove that when you slowly change the heat, the soup creates a specific, predictable pattern of bubbles (defects) that follows a very tricky, logarithmic recipe. This is the first time we've been able to see this recipe clearly in a 3D world.

Technical Summary

1. Problem Statement

Simulating the real-time dynamics of quantum many-body systems is a central challenge in modern physics, particularly for dimensions higher than one.

• The Bottleneck: The exponential growth of the Hilbert space dimension with system size ($N = L^d$) renders exact numerical methods intractable for large systems.
• Limitations of Existing Methods:
  • 1D: Matrix Product States (MPS) are highly effective.
  • 2D: Tensor Network approaches (e.g., PEPS) struggle with rapid entanglement growth, limiting simulations to short timescales.
  • 3D: Conventional methods face severe computational costs due to tensor contractions and entanglement growth, making large-scale 3D nonequilibrium dynamics essentially inaccessible.
• The Gap: While Neural Quantum States (NQS) have shown promise in 1D and 2D, their scalability and accuracy in 3D remain an open challenge due to the complexity of optimization and the need for architectures that efficiently capture 3D locality and symmetries.

2. Methodology

The authors propose a scalable framework based on Neural Quantum States (NQS) combined with the Time-Dependent Variational Principle (TDVP).

A. Neural Architecture: 3D ResNet-CNN

To represent the wavefunction $\Psi_\theta(s)$ on a cubic lattice ($L \times L \times L$), the authors introduce a specialized convolutional neural network:

• Structure: A translationally equivariant 3D Convolutional Neural Network (CNN) utilizing Residual Connections (ResNet).
• Components:
  • Stacks of $n$ residual blocks.
  • Each block contains two $3 \times 3 \times 3$ convolutional layers with unit stride and $2c$ output channels.
  • Activation: Gaussian Error Linear Unit (GELU).
  • Normalization: Residual contributions are rescaled by $1/\sqrt{i+1}$ (where $i$ is the block index) to stabilize training for deep networks.
  • Boundary Conditions: Periodic boundary conditions are enforced via circular padding to ensure translation equivariance on the 3D torus.
  • Complex Output: A "Pair Complex" layer splits the final channels into real and imaginary parts, which are exponentiated and summed to produce the complex log-wavefunction $\ln \Psi_\theta(S)$.
• Scalability: Due to weight sharing in convolutions, the number of variational parameters is independent of the system size $L$, enabling simulations on large lattices (up to $1000$ spins).

B. Dynamics Solver

• TDVP: The Schrödinger equation is projected onto the variational manifold of the NQS, yielding an effective equation of motion for the network parameters $\theta$.
• Integration: Solved using a second-order Heun scheme with adaptive time steps.
• Sampling: Expectation values are estimated via Markov-Chain Monte Carlo (MCMC) sampling from $|\Psi_\theta|^2$.
• Rotating Frame: For specific quenches, the authors employ a rotating frame strategy to handle rapid global precession, isolating the non-trivial interaction dynamics.

C. Physical Model

The framework is benchmarked on the 3D Transverse-Field Ising Model (TFIM):
$$H = -J \sum_{\langle ij \rangle} \sigma^z_i \sigma^z_j - h \sum_i \sigma^x_i$$
The system is studied across the quantum critical point ($h_c \approx 5.158 J$) under two protocols:

1. Sudden Quenches: Instantaneous changes in the Hamiltonian parameters.
2. Finite-Rate Quenches (Ramps): Continuous variation of parameters to probe the Quantum Kibble–Zurek Mechanism (QKZM).

3. Key Contributions

1. First Scalable 3D NQS Framework: The introduction of a 3D ResNet-CNN architecture tailored for cubic spin lattices, enabling exact numerical simulation of real-time dynamics for systems up to 1000 qubits ($10 \times 10 \times 10$).
2. First Numerical Demonstration of 3D QKZM: The paper provides the first large-scale numerical evidence of the Quantum Kibble–Zurek mechanism in 3D.
3. Refined Scaling Theory for Upper Critical Dimension:
   • The 3D TFIM lies at the upper critical dimension ($d+z = 4$).
   • Standard power-law scaling is modified by logarithmic factors and sub-leading inverse-logarithmic corrections.
   • The authors derived these corrections by integrating Renormalization Group (RG) flow equations up to two-loop order, providing a robust theoretical benchmark that includes sub-leading terms often neglected in lower-dimensional studies.

4. Key Results

A. Collapse-and-Revival Dynamics

• Setup: Quench from a ferromagnetic state deep into the paramagnetic regime ($h = 2h_c$).
• Observation: The NQS accurately captures collapse-and-revival oscillations in longitudinal magnetization.
• Entanglement: The Quantum Fisher Information (QFI) density grows stepwise with each oscillation, reaching values indicating multipartite entanglement spanning at least 30 spins. This confirms the framework's ability to capture strong entanglement growth in 3D.

B. Sudden Quench to Criticality

• Setup: Quench from a paramagnetic state directly to the critical point ($h = h_c$).
• Performance: The NQS captures the rapid buildup of non-trivial multipartite entanglement and sets a new state-of-the-art benchmark for simulation timescales in 3D, surpassing previous methods like Sparse Pauli Dynamics.

C. Quantum Kibble–Zurek Mechanism (QKZM)

The authors performed finite-rate quenches across the critical point to test universal scaling laws.

• Theoretical Refinement: By incorporating two-loop RG corrections, they derived a refined freeze-out time scaling:
  $$\hat{t} \propto \tau_q^{1/3} [\ln(C\tau_q)]^{1/9} \left(1 + \frac{G_\mu(\ln C\tau_q)}{\ln(C\tau_q)}\right)$$
  This accounts for the slow convergence of scaling at the upper critical dimension.
• Data Collapse: Using this refined scaling, the authors achieved compelling data collapse for three distinct observables across system sizes $L=5$ to $L=10$:
  1. Equal-time Correlation Function: Decays exponentially with the rescaled distance $R/\hat{\xi}$.
  2. Excess Energy: Scales as $Q \sim \hat{\xi}^{-4}$, consistent with the scaling dimension $\Delta_Q = d+z = 4$.
  3. Quantum Fisher Information (QFI): Scales as $f_Q \sim \hat{\xi}$, revealing universal multipartite entanglement dynamics.
• Significance: The QFI remains in the universal scaling regime over a broader range of system sizes compared to excess energy, highlighting its sensitivity to entanglement structure.

5. Significance and Outlook

• Methodological Breakthrough: This work establishes NQS as a reliable, scalable tool for 3D quantum dynamics, overcoming the limitations of tensor networks in higher dimensions.
• Benchmarking Quantum Simulators: The results provide concrete numerical benchmarks for emerging 3D quantum simulators (e.g., Rydberg atom arrays, superconducting qubits) which are currently realizing 3D models.
• Universal Physics: The successful derivation and verification of logarithmic corrections in the 3D QKZM deepen the understanding of nonequilibrium universality at the upper critical dimension.
• Future Directions: The framework is applicable to other 3D models (e.g., 3D Rydberg atoms) and can explore dynamics beyond the QKZM paradigm, such as coarsening dynamics and universal entanglement in programmable quantum simulators.

In summary, the paper successfully bridges the gap between theoretical scaling predictions and large-scale numerical simulation in 3D, validating the Quantum Kibble–Zurek mechanism with unprecedented precision and demonstrating the power of deep learning architectures in quantum many-body physics.