Gradient-based search of quantum phases: discovering unconventional fractional Chern insulators

The Big Picture

Imagine searching for a specific crystal in a vast, featureless desert. No map, no compass, just a vague sense it exists somewhere. For decades, that’s roughly how physicists hunted for exotic quantum phases of matter: navigating enormous parameter spaces with mostly intuition and luck.

These phases aren’t academic curiosities. Superconductivity gave us MRI machines and particle accelerators. The quantum mechanics of crystalline solids built the semiconductor industry. Today, physicists believe that phases with topological order could form the backbone of fault-tolerant quantum computers. In topologically ordered materials, quantum information is encoded in the material’s global structure rather than any single site, making it resistant to local errors.

A team from MIT, Harvard, and IAIFI has turned this search into a directed one. They treated quantum phase discovery as a machine learning optimization problem and built a method that locates exotic quantum phases on demand, including ones no one expected to find. The central move: encode the spectral fingerprints of a desired quantum state into a differentiable loss function, then recast phase discovery as gradient descent over Hamiltonian parameter space.

How It Works

Every quantum phase leaves a characteristic fingerprint in a system’s energy spectrum: particular patterns of ground-state degeneracy distributed across different symmetry sectors. Write down what that fingerprint should look like mathematically, define a loss function measuring how far any Hamiltonian is from producing it, then use gradient descent to minimize it. That’s the core idea.

Here’s how the target-phase loss function works. Given a Hamiltonian parameterized by a vector of numbers (hopping amplitudes, interaction strengths, lattice geometry), the algorithm partitions the low-energy spectrum into a “target manifold” of states you want as ground states and a “complement manifold” of everything else. The loss measures the energy gap between the highest target state and the lowest complement state. Negative gap means you’ve found the phase. Positive gap means gradient descent can still push you closer.

The workflow has three steps:

1. Specify the target: Encode the expected spectral fingerprints, including which symmetry sectors host ground states, how many, and how the spectrum evolves under magnetic flux insertion.
2. Minimize the loss: Run gradient descent through Hamiltonian parameter space, adjusting physical parameters to push the system toward the target phase.
3. Verify with exact diagonalization: Once the optimizer converges, use exact diagonalization (ED), a numerically exact method for small quantum systems, to confirm the phase is genuine and map the surrounding phase diagram.

The loss function is differentiable with respect to Hamiltonian parameters, which is what makes gradient optimization possible. Quantum energy spectra involve eigenvalues that can change discontinuously, so the team had to handle this carefully to keep the optimization landscape smooth.

They tested their framework on spinless fermions on the kagome lattice, a tiling of corner-sharing triangles found in materials like herbertsmithite. Their quarry: fractional Chern insulators (FCIs), topological phases where electrons organize into collective states carrying fractional charges. FCIs are lattice analogues of the fractional quantum Hall effect, potentially accessible without enormous magnetic fields.

First target: an Abelian FCI at filling fraction ν = 1/3. Most known FCIs live near the “ideal” regime, where band geometry closely mimics the lowest Landau level. The optimizer found a non-ideal FCI whose band geometry lies far outside any ideal description, contradicting a widespread assumption in the field.

The second target was more ambitious: a non-Abelian FCI at filling ν = 1/2. Non-Abelian phases are the holy grail of topological quantum computing. Their excitations, called non-Abelian anyons, store quantum information in ways that are intrinsically protected from errors. Previously known examples required either three-body interactions or unrealistically long interaction ranges. The search turned up a Moore–Read state, the prototypical non-Abelian FCI first proposed in the 1990s, stabilized purely by finite-range, two-body interactions. No one had found an explicit realization before.

Both discoveries were confirmed through exact diagonalization, spectral flow analysis under flux insertion, and phase diagram mapping.

Why It Matters

The individual discoveries are striking on their own, but the methodology is what carries long-term weight. The gradient-based framework applies to any phase with identifiable spectral fingerprints and continuous Hamiltonian parameters: charge density waves, quantum spin liquids, various topological orders. Researchers encode domain knowledge about what a phase should look like, and the optimizer finds where it lives.

The non-Abelian FCI discovery raises immediate theoretical questions. Why does this phase appear with finite-range interactions? What framework describes fractionalization in bands this far from the ideal limit? Concrete examples like this tend to accelerate theoretical understanding, much the way experimental discoveries do.

Extensions are straightforward to imagine: finite-temperature phases, disordered systems, open quantum systems. The method could also interface with materials databases, running searches over realistic parameters and pointing experimentalists toward specific compounds worth synthesizing.

This team treated quantum phase discovery as an optimization problem and found two fractional Chern insulators that conventional wisdom said shouldn’t exist. The framework they built could systematically map the exotic quantum phases hiding throughout condensed matter physics.