When Machine Learning Loses Its Way: Symmetry’s Role in Predicting Material States

Author: Denis Avetisyan

A new study reveals that convolutional neural networks struggle to accurately classify topological materials when key symmetries are disrupted, hindering their ability to generalize beyond familiar configurations.

The model describes a system where electrons traverse a lattice-defined by intra-cell hopping amplitudes <span class="katex-eq" data-katex-display="false">\{t_{1,1}, t_{1,2}, t_{1,3}, ...\}</span> and inter-cell amplitudes <span class="katex-eq" data-katex-display="false">\{t_{2,1}, t_{2,2}, t_{2,3}, ...\}</span>-supplemented by a long-range hopping term <span class="katex-eq" data-katex-display="false">t_3</span> and modulated by on-site potentials <span class="katex-eq" data-katex-display="false">\{V_1, V_2, V_3, ...\}</span>, enabling electron transport across the structure. — The model describes a system where electrons traverse a lattice-defined by intra-cell hopping amplitudes $\{t_{1,1}, t_{1,2}, t_{1,3}, ...\}$ and inter-cell amplitudes $\{t_{2,1}, t_{2,2}, t_{2,3}, ...\}$ -supplemented by a long-range hopping term $t_3$ and modulated by on-site potentials $\{V_1, V_2, V_3, ...\}$ , enabling electron transport across the structure.

Researchers demonstrate that the predictive power of neural networks for identifying topological and Anderson insulators in the Su-Schrieffer-Heeger model critically depends on the preservation of chiral symmetry.

Predicting the behavior of quantum materials often relies on computationally expensive simulations, limiting exploration of complex disorder effects. This is addressed in ‘Machine Learning of Topological Insulator and Anderson Insulator in One-Dimensional Extended Su-Schrieffer-Heeger Chain’, which utilizes convolutional neural networks to map reduced correlation matrices to topological phase diagrams, revealing a surprising sensitivity to chiral symmetry. Specifically, the trained model accurately predicts phase transitions in systems with symmetry-preserving disorder but fails when chiral symmetry is broken, indicating a transition towards an Anderson localization regime. This suggests machine learning can move beyond classification to become a powerful probe of the subtle symmetry-protected properties defining quantum matter – but how can these techniques be extended to higher-dimensional and more strongly interacting systems?

The Fragile Order of Topology: Introducing the SSH Model

The Su-Schrieffer-Heeger (SSH) model, initially developed to explain the unusual electronic properties of polyacetylene, has blossomed into a cornerstone for comprehending topological phases of matter. This simplified yet powerful model describes a one-dimensional chain where hopping strengths between neighboring atoms alternate – a seemingly subtle detail that gives rise to profoundly different behaviors. Unlike conventional materials characterized by symmetry breaking, the SSH model exhibits topological phases distinguished not by a change in local order, but by a global property of the wavefunction. This topological protection ensures the existence of robust edge states – conducting channels appearing at the boundaries of the material – which are immune to imperfections and disorder. Consequently, the SSH model serves as an accessible and versatile platform for investigating more complex topological systems and exploring their potential applications in areas like robust electronics and quantum computing, offering a crucial bridge between theoretical physics and materials science.

Chiral symmetry, at the heart of the Su-Schrieffer-Heeger (SSH) model, acts as a powerful guardian of topologically protected edge states. This symmetry, which essentially means the system remains invariant under a specific type of transformation, prevents certain perturbations from scattering or destroying these states. These edge states aren’t simply localized; they are fundamentally different from states within the bulk material, existing at the boundaries due to the topological properties of the system. The protection afforded by chiral symmetry is vital because it ensures these edge states remain robust even in the presence of imperfections or disorder – a critical feature for potential applications in areas like lossless data transmission and fault-tolerant quantum computing. Without this symmetry, the delicate balance maintaining these states would be easily disrupted, rendering them unusable for technological advancements. $\hat{H} = \sum_{n} c^\dagger_n c_{n+1} + h.c.$

The promise of topologically protected edge states extends beyond fundamental physics, offering pathways to revolutionize information technology. These states, guaranteed by the bulk-boundary correspondence – a direct link between a material’s internal topology and its surface properties – exhibit an extraordinary robustness against imperfections and disorder. Unlike conventional conductors where resistance inevitably leads to energy loss, these edge states can, in principle, support dissipationless transport, meaning electrons flow without losing energy. This characteristic is crucial for developing highly efficient electronic devices and could dramatically reduce energy consumption. Furthermore, the inherent stability of these states makes them ideal candidates for encoding and manipulating quantum information, providing a pathway toward building robust quantum computers less susceptible to the errors that plague current quantum computing architectures. The potential for both lossless energy transfer and resilient quantum computation positions these topologically protected edge states as a cornerstone of future technological advancements.

The energy spectrum of the Su-Schrieffer-Heeger model with long-range hopping reveals that diagonal disorder (<span class="katex-eq" data-katex-display="false">V_{on} = 0.1</span>) eliminates edge states and introduces bulk localizations, while both disorder-free and off-diagonal disorder conditions (<span class="katex-eq" data-katex-display="false">W_{off} = 0.1</span>) support edge states and lack bulk localizations, as indicated by IPR coloring. — The energy spectrum of the Su-Schrieffer-Heeger model with long-range hopping reveals that diagonal disorder ( $V_{on} = 0.1$ ) eliminates edge states and introduces bulk localizations, while both disorder-free and off-diagonal disorder conditions ( $W_{off} = 0.1$ ) support edge states and lack bulk localizations, as indicated by IPR coloring.

Disorder’s Impact: Breaking the Symmetry

The Su-Schrieffer-Heeger (SSH) model relies on chiral symmetry for the existence of topologically protected edge states; however, introducing on-site energy disorder – often termed diagonal disorder – directly disrupts this symmetry. Diagonal disorder manifests as a random variation in the potential experienced by each lattice site, effectively altering the hopping integrals within the SSH Hamiltonian. This random perturbation breaks the symmetry between the two sublattices fundamental to the SSH model, removing the condition required for the existence of a zero-energy edge state. The loss of chiral symmetry fundamentally alters the band structure, eliminating the topological protection and leading to the localization of what were previously extended edge states.

The disruption of chiral symmetry due to disorder directly impacts the topologically protected edge states characteristic of the Su-Schrieffer-Heeger (SSH) model. These edge states, normally robust against perturbations, experience induced localization when symmetry is broken. This localization is quantitatively measured by the Inverse Participation Ratio (IPR), where a significant increase in IPR indicates a greater degree of wavefunction confinement and, consequently, diminished functionality of the edge states. A higher IPR value signifies that the probability amplitude is concentrated on fewer lattice sites, effectively reducing the states’ ability to conduct or perform their intended topological function.

As diagonal disorder increases within the Su-Schrieffer-Heeger (SSH) model, the system transitions from a topologically protected state to an Anderson localization regime. This shift is characterized by the loss of extended edge states and the emergence of localized states throughout the bulk material. Specifically, the introduction of disorder disrupts the alternating pattern of hopping integrals, destroying the necessary conditions for topological protection and inducing a gradual increase in the Inverse Participation Ratio (IPR). This process results in a diminished ability to conduct along the edges and a confinement of electron wavefunctions within the disordered potential, effectively transforming the system’s insulating behavior from a topological origin to one arising from strong localization effects.

Disorder significantly alters the energy spectrum and localization, as off-diagonal disorder <span class="katex-eq" data-katex-display="false">W_{\mathrm{off}}</span> preserves edge states while diagonal disorder <span class="katex-eq" data-katex-display="false">V_{\mathrm{on}}</span> eliminates them and localizes bulk states, as indicated by inverse participation ratio (IPR) values. — Disorder significantly alters the energy spectrum and localization, as off-diagonal disorder $W_{\mathrm{off}}$ preserves edge states while diagonal disorder $V_{\mathrm{on}}$ eliminates them and localizes bulk states, as indicated by inverse participation ratio (IPR) values.

Preserving Topology: The Role of Off-Diagonal Disorder

Within the Su-Schrieffer-Heeger (SSH) model, chiral symmetry is directly linked to the preservation of topological edge states. Diagonal disorder, which affects on-site energies, typically breaks this symmetry, leading to localization of these states. However, off-diagonal disorder, specifically random variations in the hopping terms – the parameters governing electron transfer between lattice sites – does not inherently disrupt chiral symmetry. This is because the symmetry is defined by the relationship between the hopping amplitudes; alterations to these amplitudes, provided they maintain a specific balance, do not invalidate the symmetry condition $H = \gamma \sum_{i} c^{\dagger}_{i} c_{i+1} + h.c.$ . Consequently, off-diagonal disorder can be introduced without destroying the conditions necessary for robust topological protection and the existence of protected edge states.

The preservation of chiral symmetry is fundamentally important to the stability and functionality of topological insulators within the Su-Schrieffer-Heeger (SSH) model. Chiral symmetry guarantees the existence of topologically protected edge states, which are states localized at the boundaries of the material and immune to backscattering from non-magnetic impurities. These robust edge states are critical for maintaining high accuracy in the convolutional neural network (CNN) because they provide a reliable pathway for signal propagation, even in the presence of disorder. Disruption of chiral symmetry would lead to the localization or hybridization of these edge states, degrading performance and reducing the CNN’s ability to accurately classify inputs.

Controlled introduction of off-diagonal disorder allows for the creation of topologically protected insulators exhibiting resilience to imperfections. Analysis of the resulting feature spaces, specifically using Random Component Mapping (RCM), demonstrates that the first two principal components account for 71.3% of the total variance. This high variance explained by a limited number of components indicates a clear and dominant topological signature, confirming the effectiveness of off-diagonal disorder in maintaining robust topological properties despite the presence of disorder. The methodology effectively engineers systems where topological invariants are preserved, leading to stable and predictable behavior even with introduced imperfections.

Principal Component Analysis reveals that while disorder-free and off-diagonal disordered samples share similar feature distributions, diagonal disorder demonstrably diverges in feature space.

Beyond Nearest Neighbors: Long-Range Hopping and Enhanced Robustness

The Su-Schrieffer-Heeger (SSH) model, a cornerstone in understanding topological insulators, traditionally focuses on interactions between nearest-neighboring sites within a material. However, extending this model to incorporate long-range hopping – where electrons can ‘jump’ between sites further apart – dramatically increases the design possibilities for topological systems. This isn’t merely an addition; it introduces new degrees of freedom, allowing researchers to finely tune the topological invariants that dictate a material’s unique properties. By manipulating the strength and range of these hopping terms, it becomes possible to engineer materials with customized topological phases and potentially unlock novel functionalities beyond those achievable with strictly localized interactions. This expanded framework provides a powerful tool for creating designer topological materials tailored for specific applications, offering greater control over electron behavior and potentially leading to more robust and efficient devices.

The Su-Schrieffer-Heeger (SSH) model, a cornerstone of topological physics, gains considerable nuance through the introduction of adjustable parameters when long-range hopping is incorporated. These parameters don’t merely expand the model’s complexity; they provide a dedicated toolkit for precisely manipulating the system’s topological invariants-quantities that define its fundamental properties and protect its unique edge states. Crucially, this refined control extends to bolstering the robustness of these edge states against perturbations. By carefully tuning these long-range hopping parameters, researchers can engineer systems where the protected states remain stable even in the presence of disorder or imperfections, potentially leading to more resilient and practical topological materials and devices. This ability to sculpt the topological landscape offers an unprecedented degree of freedom in designing materials with tailored electronic and optical properties.

The deliberate combination of long-range hopping – where electrons can ‘jump’ further than to adjacent sites – and the introduction of off-diagonal disorder presents a powerful pathway towards crafting topological materials with precisely engineered characteristics. Recent investigations demonstrate that this interplay not only allows for greater control over material properties, but also significantly enhances robustness against imperfections. Specifically, convolutional neural networks (CNNs) utilized to analyze these systems maintain a remarkably high degree of accuracy in identifying the topological phase – even when subjected to disorder strengths as high as $W_{off} \approx 3.0$ . This suggests that these ‘designer’ topological materials, created through careful manipulation of hopping parameters and disorder, could be far more resilient and practical for future technological applications than previously anticipated, opening doors to stable and reliable topological devices.

Increasing disorder strength <span class="katex-eq" data-katex-display="false">W_{off}</span> progressively blurs predicted winding number boundaries and reduces prediction confidence while increasing uncertainty, though topological accuracy is maintained until <span class="katex-eq" data-katex-display="false">W_{off} \approx 3.0</span>. — Increasing disorder strength $W_{off}$ progressively blurs predicted winding number boundaries and reduces prediction confidence while increasing uncertainty, though topological accuracy is maintained until $W_{off} \approx 3.0$ .

The research highlights a critical dependency on symmetry preservation for accurate machine learning predictions – a principle echoed in philosophical thought. As Jean-Jacques Rousseau observed, “The more we are accustomed to order, the more odious everything appears to us which is irregular.” This study demonstrates that when chiral symmetry-the ‘order’ in the system-is disrupted, transitioning to an Anderson insulator, the convolutional neural network’s ability to generalize falters. The model, like any system seeking stability, struggles with the ‘irregularity’ introduced by broken symmetry. The failure isn’t inherent in the network’s architecture, but in its inability to extrapolate beyond the conditions under which it was trained-a limitation exposed by the transition to a topologically distinct, disordered phase. If replication fails under altered symmetry, the initial prediction remains suspect.

What Remains to be Seen?

The demonstrated sensitivity of convolutional neural networks to chiral symmetry breaking in the Su-Schrieffer-Heeger model isn’t a limitation of the network itself, but a stark reminder of what constitutes meaningful prediction. The model doesn’t ‘see’ topology; it correlates input features with labels assigned under specific, symmetry-defined conditions. The failure to generalize to the Anderson localization regime isn’t a bug-it’s an honest response to a fundamentally different question being asked. Data isn’t the goal-it’s a mirror of human error.

Future work will undoubtedly explore more robust network architectures, perhaps those explicitly incorporating known symmetries as priors. However, a more fruitful direction lies in acknowledging the inherent limitations of any model attempting to categorize phases based solely on local observables. The transition to an Anderson insulator introduces long-range correlations, and a complete description requires grappling with the information lost when focusing solely on short-range order.

Ultimately, the exercise highlights that even what can’t be measured still matters-it’s just harder to model. The search isn’t for a perfect predictor, but for a better understanding of the information required to even pose the right question about topological phases, especially in disordered systems.

Original article: https://arxiv.org/pdf/2603.11688.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

The Fragile Order of Topology: Introducing the SSH Model

Disorder’s Impact: Breaking the Symmetry

Preserving Topology: The Role of Off-Diagonal Disorder

Beyond Nearest Neighbors: Long-Range Hopping and Enhanced Robustness

What Remains to be Seen?

See also: