[Paper Review] High-precision Monte Carlo study of the three-dimensional XY model on GPU

Ti-Yen Lan, Yun-Da Hsieh+1 more|arXiv (Cornell University)|Nov 5, 2012

Quantum, superfluid, helium dynamics3 citations

TL;DR

This study presents a high-precision Monte Carlo simulation of the three-dimensional XY model on GPU using Metropolis, over-relaxation, and parallel-tempering updates, achieving system sizes up to L=160. The finite-size scaling analysis yields critical exponents ν=0.67098(16) and α=−0.01293(48), which are consistent with the most recent experimental estimate ν_exp=0.6709(1) for the superfluid transition in ⁴He, resolving prior discrepancies with theory.

ABSTRACT

We perform large-scale Monte Carlo simulations of the classical XY model on a three-dimensional $L imes L imes L$ cubic lattice using the graphics processing unit (GPU). By the combination of Metropolis single-spin flip, over-relaxation and parallel-tempering methods, we simulate systems up to L=160. Performing the finite-size scaling analysis, we obtain estimates of the critical exponents for the three-dimensional XY universality class: $α=-0.01293(48)$ and $ν=0.67098(16)$. Our estimate for the correlation-length exponent $ν$, in contrast to previous theoretical estimates, agrees with the most recent experimental estimate $ν_{ m exp}=0.6709(1)$ at the superfluid transition of $^4$He in a microgravity environment.

Motivation & Objective

• To perform large-scale Monte Carlo simulations of the 3D XY model on GPU to achieve higher precision in critical exponent estimation.
• To resolve discrepancies between previous theoretical estimates and experimental measurements of the correlation-length exponent ν in the 3D XY universality class.
• To leverage GPU acceleration to simulate system sizes up to L=160, enabling improved finite-size scaling analysis.
• To test the consistency of the 3D XY universality class with experimental data from the superfluid transition in ⁴He under microgravity.

Proposed method

• Implementation of GPU-accelerated Monte Carlo simulations using the NVIDIA CUDA framework for massive parallelism.
• Use of three update schemes: Metropolis single-spin flip, over-relaxation, and parallel-tempering, with one MCS defined as one sweep of each.
• Lattice decomposition into 16×16×16 blocks, each assigned to a thread block of 8×8×8 threads, with each thread updating 8 spins in a sub-block.
• Application of bootstrap resampling (1000 samples per temperature) to decorrelate data from parallel-tempering simulations.
• Finite-size scaling via data collapse of magnetization m and Binder cumulant q₂ using polynomial fits to universal scaling functions.
• Simultaneous fitting of T_c, β, ν, and α using hyperscaling relations, with error bars estimated over 100 bootstrap repetitions.

Experimental results

Research questions

• RQ1 Does the 3D XY universality class accurately describe the superfluid transition in ⁴He, as indicated by recent experimental data?
• RQ2 Can GPU-accelerated Monte Carlo simulations with system sizes up to L=160 reduce finite-size effects and improve precision in critical exponent estimation?
• RQ3 Are the theoretical estimates of ν and α in the 3D XY universality class consistent with the most recent experimental value ν_exp=0.6709(1)?
• RQ4 How do sub-leading corrections to scaling affect the finite-size scaling analysis, and can they be minimized through large system sizes and high statistics?
• RQ5 Can data collapse techniques with high-dimensional fitting and bootstrap resampling yield reliable error estimates for critical exponents?

Key findings

• The critical temperature is estimated as T_c = 2.2018312(6) from q₂ data collapse, with high consistency across multiple analyses.
• The correlation-length exponent ν is determined as ν = 0.67098(16), in excellent agreement with the experimental value ν_exp = 0.6709(1).
• The specific-heat exponent is estimated as α = -0.01293(48), consistent with the hyperscaling relation and experimental constraints.
• The magnetization exponent β is estimated as β = 0.34910(12), and the susceptibility exponent γ as γ = 1.31594(41), both derived via hyperscaling.
• The data collapse for both m and q₂ shows excellent consistency, with reduced chi-square values around 1.25, indicating reliable scaling behavior.
• The results resolve prior theoretical inconsistencies and confirm that the 3D XY universality class accurately describes the λ-transition in ⁴He.