A spin glass is a type of disordered magnetic material in which the atomic magnetic moments, or “spins,” freeze into a random, jumbled pattern rather than lining up neatly as they do in a regular magnet. This frozen disorder produces bizarre physical behavior that has captivated physicists for over half a century and has made spin glasses one of the most influential concepts in modern science, reaching far beyond magnetism into optimization, neuroscience, and protein biology. The real punch of spin glasses, though, is not the materials themselves but the mathematics needed to describe them, which turned out to be so deep that proving the key theoretical result earned a Fields Medal.
What Makes a Spin Glass Different from an Ordinary Magnet
In a normal ferromagnet like iron, cooling the material below a certain temperature causes all the atomic spins to snap into alignment. The result is a net magnetization you can feel with a refrigerator magnet. In an antiferromagnet, neighboring spins alternate in a regular checkerboard-like pattern, canceling each other out but still forming an orderly arrangement. A spin glass does neither of these things. Instead, when cooled below its freezing temperature, the spins lock into positions that appear completely random. There is no repeating pattern, no net magnetization, and no obvious order at all. Yet the system is genuinely frozen: heat it back up and cool it again, and it settles into a different random arrangement. The state is “glassy” in much the same sense that window glass is a liquid that stopped flowing, except here it is magnetic orientation rather than atomic position that gets stuck.
Spin glasses share a family resemblance with structural glasses, the everyday kind you drink from, but they differ in an important way. Structural glasses lack the time-reversal symmetry that spin glasses possess, which makes spin glasses a cleaner theoretical laboratory for studying glassy dynamics in general.1PubMed Central. Thermodynamic glass transition in a spin glass without time-reversal symmetry That cleanliness is part of why theorists have lavished so much attention on them.
Frustration and Disorder
The key ingredient that turns an ordinary magnet into a spin glass is frustration, and not the human kind. Imagine three spins sitting at the corners of a triangle. Two of them want to point in opposite directions from each other (antiferromagnetic coupling), and the third also wants to oppose both of its neighbors. It cannot: if spin A is “up” and spin B is “down,” spin C cannot simultaneously be opposite to both. No matter what it does, at least one pairing is unhappy. That unsatisfied bond is frustration. Spread it across billions of atoms and you get a system with an enormous number of nearly equivalent low-energy arrangements, none of which is clearly the “right” one.
Frustration alone is not enough, though. You also need randomness, meaning the strengths and signs of the magnetic couplings between atoms vary unpredictably from place to place. In a classic spin glass material like copper doped with a small percentage of manganese atoms (CuMn), the manganese impurities sit at random positions inside the copper lattice, and the interactions between them oscillate between ferromagnetic and antiferromagnetic depending on the distance. The combination of random placement and oscillating interactions creates a thicket of conflicting demands that the spins cannot all satisfy at once.
When geometric frustration is already present in the crystal structure itself, even a small amount of random disorder can stabilize the spin-glass phase. Numerical studies of antiferromagnetic clusters on frustrated lattices have shown that intrinsic geometric frustration helps the spin-glass order survive at lower levels of disorder than would otherwise be needed.2PubMed Central. Interplay between spin-glass clusters and geometrical frustration
What Happens When You Cool a Spin Glass
At high temperatures a spin glass behaves like any paramagnet: the spins jiggle around randomly, responding weakly to an applied magnetic field. As the temperature drops toward the freezing temperature (often labeled Tf), something peculiar happens. The magnetic susceptibility, which measures how easily the material magnetizes in a field, develops a sharp cusp. Below that temperature the spins freeze, but unlike a ferromagnet’s crisp phase transition the freezing is gradual and frequency-dependent. Probe the material with a slowly oscillating field and you see the cusp at one temperature; probe it faster and the cusp shifts upward. That frequency dependence is a hallmark of glassy dynamics and distinguishes the spin-glass transition from conventional magnetic ordering.
Once frozen, spin glasses display spectacular memory effects. Cool a sample to a particular temperature, wait while the spins slowly rearrange themselves (a process called aging), then cool further. If you later reheat through the original temperature, the system “remembers” the aging it underwent there, as if the intervening cooling had not happened. This memory is accompanied by rejuvenation: at each new lower temperature, the aging process restarts as though the system were freshly quenched. These effects have long been seen as a potential experimental window into the exotic mathematical structure theorists predict for spin glasses. Reproducing them in computer simulations, however, proved extraordinarily difficult for more than two decades. Only recently has the custom-built Janus II supercomputer managed to simulate memory and rejuvenation reliably, by carefully matching the length scales in simulation to those in real CuMn crystals.3Nature Physics. Memory and rejuvenation effects in spin glasses are governed by more than one length scale
The Energy Landscape Picture
A useful way to visualize why spin glasses behave the way they do is the energy landscape. Think of a mountainous terrain with countless valleys separated by ridges of varying heights. Each valley is a configuration of spins that is locally stable: small rearrangements would cost energy. The spin glass freezes when the thermal energy available to the system is no longer enough to carry it over the ridges between valleys, so it gets stuck in whichever valley it happens to occupy. Because the landscape has an astronomical number of valleys with nearly identical depths, tiny perturbations like a small temperature change can redirect the system into a completely different valley, which is why the frozen state is irreproducible from one cooling run to the next.
Large-scale simulations of the three-dimensional Edwards-Anderson spin glass, a standard workhorse model, confirm this picture directly. Autocorrelation times in parallel-tempering Monte Carlo simulations are closely correlated with the roughness of the free-energy landscape: rougher landscapes produce longer trapping times, just as you would expect if the system were literally getting stuck in deeper valleys with higher surrounding ridges.4PubMed Central. Correlations between the dynamics of parallel tempering and the free-energy landscape in spin glasses
Parisi’s Solution and Its Proof
In 1975, David Sherrington and Scott Kirkpatrick proposed a simplified spin-glass model in which every spin interacts with every other spin through random couplings, ignoring the spatial structure of real materials. This “infinite-range” or mean-field model, known as the SK model, was supposed to be easy to solve. It was not. A straightforward calculation gave nonsensical results, including a negative entropy at low temperatures, which would mean fewer than one possible state, a physical absurdity.
Giorgio Parisi rescued the SK model in 1979 with an extraordinary and initially mysterious mathematical scheme called replica symmetry breaking (RSB). The idea, loosely, is that the many valleys in the energy landscape are not all independent. They are organized in a hierarchical, tree-like structure: clusters of similar states nested inside larger clusters, which sit inside still larger clusters, all the way up. Parisi wrote down an exact formula for the free energy of the SK model based on this picture. It gave a physically sensible entropy that approaches zero as the temperature goes to zero, exactly as it should.5PubMed. Sherrington-Kirkpatrick model for spin glasses: Solution via the distributional zeta function method
For more than two decades, Parisi’s formula was a conjecture, brilliant and widely believed but unproven. In 2006, Michel Talagrand provided a rigorous mathematical proof that the formula is correct, confirming it for the SK model and a broad class of related models.6Annals of Mathematics. The Parisi formula 7arXiv. Michel Talagrand and the Rigorous Theory of Mean Field Spin Glasses Parisi himself received the 2021 Nobel Prize in Physics in part for this work, and Talagrand was awarded the 2024 Abel Prize, the mathematics equivalent of the Nobel, largely for his proof of the Parisi formula. Few problems in statistical physics have produced results of such mathematical depth.
Ultrametricity and the Tree of States
One of the strangest predictions of Parisi’s RSB picture is ultrametricity. In everyday geometry, distances obey the triangle inequality: the distance from A to C is at most the sum of the distances from A to B and from B to C. In an ultrametric space, a much stronger rule holds: among any three points, the two largest distances are always equal. This is the geometry of a tree. Think of it this way: pick any three leaves on a tree, measure the path length between each pair along the branches, and the two longest paths will always be the same length because they share the same trunk segment.
Parisi’s theory predicts that the overlap between different spin-glass states, a measure of how similar two configurations are, obeys this ultrametric rule. If you pick three random equilibrium states and measure how alike each pair is, the two most dissimilar pairs will always be dissimilar by the same amount. Verifying this on finite systems has been extremely challenging. Numerical studies of the mean-field model in its low-temperature RSB phase have found hints of the emerging hierarchical structure, but the lattice sizes accessible to current computers are too small to make the signal unambiguous.8arXiv. A Quantitative Clustering Approach to Ultrametricity in Spin Glasses Detecting ultrametricity demands very large system sizes, which is one reason the question of whether real, three-dimensional spin glasses share this property remains open.
The RSB Versus Droplet Debate
Parisi’s RSB picture describes the mean-field SK model beautifully, but real spin glasses live in three dimensions, not the infinite-dimensional world the SK model assumes. Whether the full RSB structure, with its hierarchy of states and ultrametric overlaps, survives in realistic low-dimensional systems is one of the longest-running open questions in condensed matter physics.
The main rival is the droplet picture, developed in the 1980s by Daniel Fisher and David Huse. In this view, the low-temperature phase of a real spin glass has only a single pair of ground states (related by flipping every spin), and the apparent complexity of the energy landscape disappears in the limit of very large systems. Excitations above the ground state are compact “droplets” of flipped spins whose energy grows as a power of their size. If the droplet picture is right, the elaborate tree of states predicted by RSB is an artifact of the infinite-range SK model and does not describe nature.
The two pictures make different predictions for a quantity called the replicon exponent, which characterizes how the overlap between states fluctuates. RSB predicts it should be nonzero, while the droplet picture predicts it should be zero in the infinite-system limit. Simulations over the past decade have consistently found a nonzero value, which looks like a win for RSB. But proponents of the droplet picture argue that the crossover to zero happens only at system sizes far larger than any simulation has reached, possibly hundreds of lattice spacings in three dimensions.9PubMed. Droplet-scaling versus replica symmetry breaking debate in spin glasses revisited Recent theoretical work has shown that the two pictures may be less contradictory than they appear: the existence of low-energy interfaces between different ground states, as predicted by RSB, turns out to be equivalent to the presence of special “critical droplets” whose boundaries span a finite fraction of the lattice.10Frontiers in Physics. Critical droplets and replica symmetry breaking
Related to this debate is the question of how many spatial dimensions a spin glass needs in order to have a true phase transition at a nonzero temperature. In two dimensions, the transition temperature is zero, meaning the spin-glass phase exists only at absolute zero and is not a real thermodynamic phase. In three dimensions it survives at finite temperature. The boundary, called the lower critical dimension, appears to lie around 2.5.11PubMed Central. Support for the value 5/2 for the spin glass lower critical dimension at zero magnetic field 12Journal of Applied Physics. Lower critical dimensionality of the spin-glass phase transition That fractional number is not physically meaningful on its own (you cannot build a 2.5-dimensional crystal), but it tells theorists that three-dimensional spin glasses sit only somewhat above the edge, which may explain why many signatures of the transition are hard to observe cleanly.
Spin Glasses as Optimization Problems
Finding the true ground state of a spin glass, the single configuration with the absolute lowest energy among the staggering number of possibilities, is one of the hardest problems in computational physics. For the general case it belongs to the class of NP-hard problems, meaning no known algorithm can solve it efficiently as the number of spins grows. This is not just an academic concern. A huge family of real-world optimization problems, from routing delivery trucks to designing microchips to scheduling airline crews, can be rewritten in the language of spin glasses. Minimizing the cost function of the practical problem is mathematically identical to finding the ground state of a suitably constructed spin glass.13PubMed Central. Searching for spin glass ground states through deep reinforcement learning
This connection has driven a lively exchange between physics and computer science. Physicists borrow algorithmic tools like simulated annealing (inspired by metallurgical cooling) and parallel tempering from computational spin-glass research. Computer scientists, in turn, use the structure of spin-glass energy landscapes to understand why certain optimization problems are hard and where algorithmic phase transitions occur: parameter regimes where a problem flips from easy to practically unsolvable. Deep reinforcement learning agents have recently been trained to search for spin-glass ground states, demonstrating that machine-learning methods can compete with traditional physics-based algorithms on these problems.14PubMed Central. Searching for spin glass ground states through deep reinforcement learning
Quantum Annealing and Spin Glasses
Classical simulated annealing escapes energy-landscape valleys by using thermal fluctuations: the algorithm heats the system, letting it hop over barriers, then slowly cools it, hoping it settles into a deeper valley. Quantum annealing replaces thermal hopping with quantum tunneling. Instead of going over a barrier, the system tunnels through it, which can be faster for certain barrier shapes.
Theoretical and numerical studies comparing the two approaches on the two-dimensional random Ising model, a standard spin-glass benchmark, found that quantum annealing reaches lower residual energies than classical annealing for the same amount of computational time. Both methods produce residual energy that shrinks as an inverse power of the logarithm of the annealing time, but the power is larger for the quantum version, which translates to faster convergence.15PubMed. Theory of quantum annealing of an Ising spin glass The picture is not entirely rosy, however. Different quantum simulation techniques yield different performance levels. Green’s function Monte Carlo implementations of quantum annealing, for example, have so far performed worse than path-integral Monte Carlo versions, largely because constructing a good trial wave function for a quantum glass is itself an unsolved problem.16PubMed. Quantum annealing of an Ising spin-glass by Green’s function Monte Carlo
Companies like D-Wave have built physical quantum annealers whose hardware is essentially a programmable spin glass etched into a superconducting chip. Whether these machines deliver a genuine quantum speedup over classical algorithms for practical problems remains an active and occasionally contentious area of research, but the underlying motivation traces directly back to spin-glass physics.
Neural Networks and the Hopfield Connection
In 1982, John Hopfield proposed a model of associative memory, a network of neurons that can store and retrieve patterns, that turned out to be mathematically identical to a certain type of spin glass. Each neuron is a spin that can be “on” or “off,” and the connection strengths between neurons play the role of the random couplings. When the network stores too many patterns relative to its size, it undergoes a phase transition into a spin-glass phase where retrieval fails: the network gets trapped in spurious states that do not correspond to any stored pattern, much like a spin glass getting stuck in a random valley.
Rigorous studies of Hopfield-type models with higher-order interactions have confirmed this picture. Above a critical temperature the overlap between different configurations concentrates at zero, meaning the system retains no memory. Below it, the overlap distribution splits, signaling that the system breaks into distinct states. When the number of stored patterns is large enough, the accessible states no longer include the original patterns at all, which is a genuine spin-glass transition: the network’s memory has been overwhelmed by disorder.17arXiv. The spin-glass phase-transition in the Hopfield model with p-spin interactions Hopfield shared the 2024 Nobel Prize in Physics for this work, making spin-glass theory one of the few threads in physics to contribute to two separate Nobel Prizes.
Protein Folding and Biological Landscapes
The energy landscape metaphor that makes spin glasses so useful in physics turns out to be equally powerful in biology. A protein is a chain of amino acids that must fold into a precise three-dimensional shape to function. The folding process navigates an energy landscape with an enormous number of possible configurations, and since the 1980s theorists have used spin-glass models to study it. The analogy is natural: the amino-acid sequence provides a form of quenched disorder (fixed and random-looking), and the competing interactions between residues create frustration very similar to what happens in a magnetic spin glass.18PubMed Central. Spin glasses and the statistical mechanics of protein folding
More recently, computational studies using Markov state models of real protein systems have found evidence for distinct dynamical phases resembling the active and inactive phases seen in glass physics. Some proteins fold through dynamics that look “glass-like,” with long trapping times in metastable states, while others fold more smoothly.19PubMed Central. Emergence of glass-like behavior in Markov state models of protein folding dynamics The extent to which real protein folding is governed by glassy kinetics remains debated, but the mathematical tools developed for spin glasses, including the energy-landscape framework and the statistical mechanics of disordered systems, have become standard equipment in the protein-folding theorist’s toolkit.
Why Spin Glasses Remain Unfinished Business
Despite decades of progress, spin glasses still resist a clean, settled understanding in the way that, say, superconductivity has one. The mean-field theory is rigorously proven and spectacularly rich, but whether its key features survive in three dimensions is unresolved. Simulations are constrained by the sheer difficulty of equilibrating disordered systems at low temperatures; even the Janus II supercomputer, a machine built specifically for this problem, can explore only modest system sizes on the relevant time scales.20Nature Physics. Memory and rejuvenation effects in spin glasses are governed by more than one length scale Experimental spin glasses like CuMn are well characterized but hard to probe at the microscopic level needed to test the theoretical predictions directly. And the lower critical dimension sitting so close to three means that three-dimensional spin glasses may be perpetually on the edge, displaying mean-field-like behavior at accessible scales while hiding their true asymptotic nature at scales no experiment or simulation can reach.21PubMed. Droplet-scaling versus replica symmetry breaking debate in spin glasses revisited
That unresolved character is arguably what keeps spin glasses fascinating. They sit at the intersection of condensed matter physics, probability theory, computer science, and biology, and each of those fields keeps finding new reasons to care about disordered systems with rugged energy landscapes. The mathematics born from trying to understand a curious magnetic alloy in the 1970s has become one of the most versatile conceptual exports in all of modern science.