Abstract
International audience. This review examines the impact of high-performance numerical simulations on the understanding of spin glasses, specifically addressing the competition between Replica Symmetry Breaking (RSB) and the Droplet Model. Through the use of dedicated FPGA-based supercomputers, evidence for RSB in 3D systems is presented, alongside a discussion of the unresolved behavior of these systems in external magnetic fields.
Introduction
Spin glasses are disordered magnetic systems characterized by slow equilibration and complex low-temperature phases. For decades, the "droplet/RSB polemic" has dominated the field. The Replica Symmetry Breaking (RSB) theory, proposed by Giorgio Parisi, suggests that spin glasses possess an infinite number of "pure states" and behave like a critical phase. Conversely, the Droplet Model argues for only two pure states, interpreting complex experimental behaviors as long-lived transients. Distinguishing between these theories requires simulations at timescales and system sizes previously considered unreachable, a challenge addressed by modern high-performance computing.
Methodology
The study focuses on the Edwards-Anderson (EA) model, a standard framework for spin glasses. To overcome computational hurdles, researchers utilized dedicated hardware, specifically the Janus and Janus II supercomputers. These FPGA-based machines are hardwired for spin glass Monte Carlo simulations, reaching timescales equivalent to hours of real-world aging. Key techniques include:
- Parallel Tempering: An algorithm allowing simulations to escape local energy traps by swapping between temperatures.
- Finite-Size Scaling (FSS): A mathematical method to extrapolate infinite-system properties from finite simulated samples.
- Statics-Dynamics Dictionary: A conceptual framework relating equilibrium properties of small simulated systems to the non-equilibrium behavior of large experimental samples, such as CuMn crystals.
Evidence for Replica Symmetry Breaking
In the absence of an external magnetic field, numerical evidence strongly supports the RSB framework in three dimensions. A primary indicator is the overlap distribution, $P(q)$. RSB predicts that $P(q)$ remains non-trivial as the system size $L$ increases, signifying the presence of multiple states.
As illustrated in Figure 4, the peaks of $P(q)$ narrow with increasing $L$, yet the distribution between them persists. This non-trivial distribution is a hallmark of the infinite states predicted by RSB. Furthermore, the identification of a non-zero "replicon" exponent ($\theta \approx 0.35–0.4$) directly contradicts the Droplet Model's prediction of $\theta = 0$.
Scaling and Experimental Alignment
Simulations show that the probability of zero overlap, $P(q=0)$, is independent of system size $L$ and proportional to temperature $T$, exactly as RSB predicts. This quantitative alignment is crucial for validating the theory against both numerical and experimental benchmarks.
As shown in Figure 3, the data for $P(q=0)/T$ collapses onto a single curve for different $L$, confirming the scaling behavior. The "statics-dynamics dictionary" successfully describes experimental spin glasses of size $L \sim 100$ lattice spacings, showing a remarkable match between the logarithm of effective time in CuMn experiments and the scaling found in Janus II simulations.
The Magnetic Field Challenge
While the zero-field case is well-described by RSB, the introduction of an external magnetic field remains one of the most challenging open questions. In 4D systems, a phase transition (the de Almeida-Thouless or dAT line) is clearly visible. However, in 3D, this transition is either absent or obscured by statistical fluctuations. This suggests that the lower critical dimension for the dAT transition may be higher than 3, or that current simulations still lack the necessary scale to resolve it.
Conclusion
The integration of dedicated hardware like Janus II with advanced algorithms has allowed simulations to probe the same physical regimes as real-world experiments. While RSB is the leading theory for 3D spin glasses in zero field, the behavior in a magnetic field remains a significant frontier. Future progress will likely require a new generation of dedicated computers and novel algorithmic approaches to fully map the in-field phase diagram.
Summary updated