Neural Networks Meet Quantum Physics: A Machine Learning Approach to Solving the Transverse Field Ising Model

This project demonstrates how to use Neural Quantum States (NQS) combined with the variational Monte Carlo method to solve the ground state properties of the transverse field Ising model, representing a cutting-edge attempt at the intersection of quantum physics and machine learning. The project covers one-dimensional model validation, two-dimensional model extension, sampling and training processes, and visualization analysis, providing an innovative solution to the computational challenges of quantum many-body systems.

The Ising model is a classic model for understanding magnetic phase transitions. The transverse field Ising model introduces quantum effects (spin tunneling caused by a transverse magnetic field), exhibiting rich quantum phase transitions. The one-dimensional version can be solved analytically, while the two-dimensional version has no closed-form solution in the thermodynamic limit and relies on numerical methods. Traditional exact diagonalization is limited by the exponential growth of the Hilbert space, which becomes a computational bottleneck.

The core method of the project combines variational Monte Carlo with neural networks: parameterizing the trial wave function using a neural network, sampling spin configurations via Markov Chain Monte Carlo (MCMC), calculating local energy, and optimizing network parameters using the variational principle to minimize the expected energy. The advantage of this method is that neural networks can capture complex quantum entanglement, and the computational complexity grows polynomially with the system size.

The project first validates the method's effectiveness on the one-dimensional transverse field Ising model (using analytical solutions as a benchmark), implementing spin sampling, network forward propagation, and gradient descent optimization. The two-dimensional model is the key and difficult part, with more complex quantum entanglement. Restricted Boltzmann Machines (RBM) or graph neural networks may be used to handle the lattice structure, requiring fine-tuning of hyperparameters and longer computation time.

Sampling uses the MCMC proposal-acceptance mechanism to generate configurations that conform to the wave function distribution. The training process iterates through sampling, calculating variational energy and gradients, and updating parameters until convergence. Visualization components include curves of energy over iterations, order parameter evolution, and spin correlation function displays, which intuitively present quantum phase transitions and the neural network learning process.

The project demonstrates the value of machine learning in scientific research: neural networks are not only computational tools; their representations may reveal the intrinsic structure of quantum systems (e.g., hidden layer activations corresponding to quantum correlations). As a course project, it helps students understand core concepts across multiple disciplines. Extension directions include trying complex architectures, studying more difficult quantum models, and exploring the application of reinforcement learning in quantum control.