Tracing the Flumen of Topological Order Through Disordered Lattices

Topological order in crystalline systems is well charted, but real materials are rarely perfect. Disordered lattices—amorphous networks, heavily doped semiconductors, or quasicrystals—can still host topological phases, but identifying them demands a different toolkit. This guide is for researchers who already know what a Chern number is and want to know how to compute it when there is no Brillouin zone. We focus on practical workflows, common pitfalls, and the physical intuition that separates genuine topological order from trivial localization.

Who Needs This and What Goes Wrong Without It

Anyone working with strongly disordered or non-crystalline systems—amorphous thin films, topological Anderson insulators, or moiré heterostructures with twist disorder—needs a reliable method to detect topological order. Without one, you risk misinterpreting Anderson-localized edge states as topological, or missing a topological phase because the conventional momentum-space invariants are undefined.

The standard approach of computing Berry curvature over a periodic Brillouin zone fails when periodicity is broken. Researchers who naively apply k-space formulas to disordered supercells often obtain noisy or meaningless results. Worse, the disorder itself can induce a topological phase—the topological Anderson insulator—where a clean system is trivial but disorder opens a gap and drives a topological transition. Without a real-space invariant, such a phase is invisible.

Another common failure mode is misidentifying trivial edge states. In a disordered system, surface roughness can produce localized states that look like topological edge modes in a density-of-states plot. Only a careful computation of a local topological marker—like the local Chern marker or Bott index—can distinguish them. Teams that skip this step often publish results that cannot be reproduced by others.

The stakes are high: false positives waste time and misdirect follow-up experiments. For example, a purported topological phase in an amorphous film might later be traced to a trivial surface resonance, undermining the paper's conclusions. Conversely, false negatives cause researchers to overlook promising topological behavior in materials that are inherently disordered, such as high-entropy alloys or doped topological insulators.

This guide equips you with the conceptual framework and step-by-step procedures to avoid these traps, saving months of dead-end work.

Who Should Read This

This article is for condensed matter theorists and computational materials scientists who have a working knowledge of tight-binding models and topological band theory. If you have never computed a Chern number in a crystalline system, start with a standard textbook. We assume you are comfortable with Python or Julia and have access to HPC resources for larger simulations.

Prerequisites and Context to Settle First

Before diving into workflows, we need to clarify the core ideas that make topological order tractable in disordered lattices. The key insight is that topological invariants can be defined in real space without translational symmetry. The most common are the local Chern marker (LCM) and the Bott index. Both are derived from the projection operator onto occupied states, which is well-defined even in a disordered system.

The local Chern marker at a lattice site r is computed as: C(r) = (2π) Im Tr[P [X, P] [Y, P]] where P is the projector onto occupied eigenstates, and X, Y are position operators. The spatial average of C(r) over the bulk gives the Chern number. For a topological phase, the bulk average is quantized to an integer, while the edges show deviations. The Bott index is similar but uses unitary matrices U = exp(i2π X/L) and V = exp(i2π Y/L) and computes the commutator of projected versions. Both methods are robust to disorder as long as the gap remains open.

Another prerequisite is understanding the role of the spectral gap. In disordered systems, the gap can be filled with localized states (Anderson insulator) or remain clean. Topological order requires a mobility gap—a range of energies where all states are localized—but the topological invariant is still quantized if the Fermi level lies in the gap. This is the essence of the topological Anderson insulator: disorder creates a mobility gap and drives a topological transition.

It is also important to distinguish between different types of disorder. Weak disorder (small on-site potential fluctuations) may preserve the topological phase, while strong disorder can close the gap or induce a trivial Anderson insulator. Correlated disorder (e.g., random but smooth potential variations) can have different effects than uncorrelated white-noise disorder. The spatial correlation length matters: long-range correlated disorder can mimic a superlattice potential and induce new topological phases.

Finally, finite-size effects are severe in disordered systems. The local Chern marker converges slowly with system size, and spurious quantization can appear in small samples. We recommend using at least 50×50 sites for 2D systems and averaging over many disorder realizations (100 or more) to obtain reliable statistics.

Core Workflow: Identifying Topological Order in a Disordered Lattice

We now present a step-by-step workflow that has been tested on a variety of disordered systems, from amorphous graphene to topological Anderson insulators.

Step 1: Construct the Real-Space Tight-Binding Model

Start with a lattice geometry that includes disorder. For a square lattice with on-site disorder, the Hamiltonian is H = ∑_i ε_i c_i† c_i + ∑_⟨ij⟩ t c_i† c_j, where ε_i are random numbers uniformly distributed in [-W/2, W/2] (W is disorder strength). For amorphous systems, generate a random point pattern (e.g., using a Poisson disk sampling) and connect nearest neighbors within a cutoff radius. Ensure the connectivity is consistent with the desired coordination number.

Step 2: Diagonalize and Identify the Occupied Subspace

Diagonalize the Hamiltonian for a finite cluster with open boundary conditions. Sort the eigenvalues and identify the number of occupied states N_occ based on the filling fraction. For a half-filled system, N_occ = N_sites/2. The projector onto occupied states is P = ∑_n=1^N_occ |ψ_n⟩⟨ψ_n|. Use a sparse eigensolver (e.g., ARPACK) for large systems.

Step 3: Compute the Local Chern Marker

For each site i, compute the matrix elements [X, P] and [Y, P] in the site basis. The position operators X and Y are diagonal in the site basis (x_i, y_i). Then compute the trace over the site index within a bulk region (excluding edges to avoid boundary effects). The average over the bulk gives the Chern number. For a system of size L×L, we typically exclude a boundary of width L/10. The LCM converges to an integer when the system size is large enough and the gap is open.

Step 4: Perform Disorder Averaging

Repeat steps 1–3 for many disorder realizations (typically 100–500). For each realization, compute the bulk average of the LCM. Then average over realizations and compute the standard deviation. A topological phase will show a quantized average (e.g., 1 or -1) with small variance. If the average is non-integer or the variance is large, either the system is trivial or the gap is closing.

Step 5: Check the Spectral Gap and Edge States

Plot the density of states (DOS) and identify the mobility gap. For a topological phase, the gap should be finite (or at least a mobility gap). Compute the local density of states (LDOS) at the edges; topological edge states appear as a peak in the LDOS within the gap. However, beware: trivial edge resonances can also appear. The true test is the quantized LCM.

Tools, Setup, and Environment Realities

Choosing the right computational tools can make or break your workflow. Here we compare the most common options.

KWANT

KWANT is a Python package designed for quantum transport, but it can also compute the local Chern marker via its built-in functionality for real-space invariants. It handles arbitrary lattice geometries and disorder naturally. The downside is that KWANT's tight-binding builder is optimized for transport, and computing projectors for very large systems ( >10⁵ sites) becomes memory-intensive. For moderate sizes (up to 10⁴ sites), it is excellent.

Custom ED Codes (Python/Julia)

Writing your own exact diagonalization code gives full control. In Python, use NumPy/SciPy with sparse eigensolvers (scipy.sparse.linalg.eigsh). For larger systems, consider using the SLEPc library via petsc4py. In Julia, the Arpack.jl and KrylovKit.jl packages are efficient. The main challenge is implementing the LCM correctly; a common mistake is forgetting to use the bulk average or miscomputing the commutator.

GPU Acceleration

For systems up to about 2000 sites, full diagonalization on a GPU (using CuPy or Julia's CUDA.jl) can speed up the computation by an order of magnitude. However, for larger systems, sparse iterative methods are still faster on CPUs due to memory bandwidth limitations. We recommend using GPU for disorder averaging over many small realizations rather than a single large system.

Finite-Size Scaling

Always perform a finite-size scaling analysis. Compute the LCM for system sizes L = 10, 20, 30, 40, 50 and check if the average approaches an integer. Extrapolate to the thermodynamic limit using a linear fit in 1/L. If the extrapolated value is not integer, the system is likely trivial or at a critical point.

Variations for Different Constraints

Not all disordered systems are the same. Here we discuss how to adapt the workflow for different scenarios.

Weak vs. Strong Disorder

In the weak disorder regime (W < t), the topological phase is stable and the LCM converges quickly. You can use smaller system sizes (e.g., 30×30) and fewer realizations (50). For strong disorder (W > t), the gap may close or become a mobility gap. The LCM may show large fluctuations; increase the number of realizations to 500 and use larger systems (50×50 or more). Check the DOS to ensure a mobility gap exists.

Correlated Disorder

If the disorder potential is spatially correlated (e.g., Gaussian-correlated with correlation length ξ), the effective disorder strength is reduced at long wavelengths. This can stabilize topological phases even for large W. To model correlated disorder, generate random Fourier modes with a power-law spectrum. The LCM workflow remains the same, but you need to average over fewer realizations because the disorder is smoother. However, finite-size effects become important when ξ is comparable to the system size; ensure L >> ξ.

Amorphous Lattices

For amorphous systems, there is no underlying lattice; you generate random points and connect them. The local Chern marker is still well-defined, but the position operators are the actual coordinates. A challenge is that the coordination number varies, leading to fluctuations in the local density of states. To reduce noise, use a larger bulk region and average over many realizations of the amorphous geometry (each with the same density but different point patterns). The topological phase in amorphous systems often requires a certain degree of local order (e.g., hexagonal rings in amorphous graphene).

Three-Dimensional Systems

In 3D, the local Chern marker is replaced by the real-space Chern number, which is a 3D trace. The computational cost scales as N² (since the projector is N×N), so 3D systems are limited to about 20³ sites. Use the Bott index instead, which is more efficient. The workflow is analogous: compute the Bott index for each realization and average. For 3D topological insulators, the index is a Z₂ invariant, which can be computed from the parity of the Bott index.

Pitfalls, Debugging, and What to Check When It Fails

Even with a solid workflow, things go wrong. Here are the most common issues and how to diagnose them.

Non-Quantized Average

If the disorder-averaged LCM is not close to an integer, first check that the Fermi level lies in a gap. Plot the DOS and look for a dip at the Fermi energy. If there is no gap, the system is metallic and the LCM is not defined. If there is a gap but the average is non-integer, the system may be too small. Increase the system size and check if the value approaches an integer. Also check that you are averaging over the bulk only; edge contamination can spoil the quantization.

Large Variance Across Realizations

A large standard deviation ( >0.3) indicates that some realizations are topological and some are trivial, or that the gap is closing in some realizations. Plot the distribution of LCM values; if it is bimodal, the system is at a transition. Increase the number of realizations and check if the variance decreases with system size (it should scale as 1/√N_real). If not, the disorder is strong enough to destroy the topological phase in a fraction of samples.

Spurious Edge States

Edge states that appear in the LDOS but are not topological are common. To test, compute the LCM on a cylinder geometry (periodic in one direction, open in the other). Topological edge states will carry a quantized conductance, while trivial ones will not. Alternatively, add a small perturbation to the edge (e.g., a potential step); topological edge states are robust, while trivial ones shift in energy.

Finite-Size Quantization

In small systems, the LCM can accidentally be quantized even in a trivial phase due to finite-size effects. Always perform a finite-size scaling. If the LCM is quantized at L=10 but not at L=20, the system is trivial. Also check the Bott index as a cross-validation; if both agree, the result is more trustworthy.

FAQ and Checklist for Verifying Results

We distill the key checks into a brief FAQ and a practical checklist.

Frequently Asked Questions

Q: Can I use the Chern number from a supercell with periodic boundary conditions?
A: Yes, if the supercell is large enough to sample the disorder. However, the supercell must be commensurate with the disorder periodicity, which is artificial. Real-space methods are more reliable.

Q: How many disorder realizations are needed?
A: For a clear topological phase, 100 is often enough. Near a transition, 500 or more may be needed to resolve the critical behavior.

Q: What if my system has spin-orbit coupling?
A: The local Chern marker generalizes to the spin Chern number or Z₂ invariant. For time-reversal invariant systems, use the real-space spin Bott index or the parity of the LCM.

Verification Checklist

• Fermi level lies in a mobility gap (check DOS).
• Disorder-averaged LCM is integer (within 0.1).
• Standard deviation decreases with system size.
• Finite-size scaling extrapolates to integer.
• Edge states are present in LDOS and robust to perturbations.
• Result is consistent with Bott index (cross-check).
• No trivial edge resonances (check conductance).

What to Do Next: Specific Next Moves

Once you have identified a topological phase in a disordered lattice, the next steps depend on your goal. If you are studying a specific material, compare your model parameters to experimental data. For example, if you find a topological Anderson insulator in a disordered HgTe quantum well, compute the conductance and compare with transport measurements. If you are exploring new physics, consider the following directions:

• Compute the topological phase diagram as a function of disorder strength and filling. Use the workflow to map out the phase boundaries with high resolution.
• Study the effect of interactions. Add a Hubbard term and use mean-field theory or exact diagonalization for small systems to see if the topological phase survives.
• Calculate the entanglement spectrum. The entanglement spectrum of a topological phase shows a distinctive degeneracy pattern, which can be computed using the real-space projector.
• Design a device. Propose a heterostructure where the disordered topological phase is sandwiched between trivial insulators, and predict quantized transport signatures.
• Publish your code. Sharing your LCM implementation facilitates reproducibility and accelerates the field. Include a repository with example input files and a tutorial.

Finally, always validate your results with a second method. If you used the local Chern marker, also compute the Bott index. If both agree, you can be confident. If they disagree, revisit your implementation—it is easy to misplace a factor of 2π or mis-index the projector. The field of topological order in disordered systems is rapidly evolving, and careful numerics are the foundation for credible discoveries.