Data-driven reconstruction of band dispersion and quantum geometry via Koopman dynamic mode decomposition
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.
Kai: I'm Kai.
Mira: And I'm Mira, here with Lev.
Kai: We're looking at the paper: "Data-driven reconstruction of band dispersion and quantum geometry via Koopman dynamic mode decomposition".
Title and authors: Mira: So, to recap what we’ve covered, this paper introduces a data-driven framework using Koopman dynamic mode decomposition to reconstruct band structures and quantum geometry directly from spatiotemporal data, bypassing the need for an explicit Hamiltonian. We've touched on how this connects to Floquet-Bloch theory and how it can extract spectral properties.
Kai: Exactly, and the central question they are asking is whether an unknown system can be effectively described in terms of quasiparticle-like excitations like Bloch modes, localized modes, or collective oscillatory modes associated with pattern formation. The answer they give is yes.
Lev: That’s a strong claim for a data-driven method. What does that actually mean when we're dealing with real experimental noise and finite measurements? Does the reconstruction hold up under those kinds of conditions?
Mira: They demonstrate this connection numerically in prototypical one and two dimensional lattice models, specifically the SSH model, graphene, and Haldane models. Across these typical lattice models, they show that Koopman dynamic mode decomposition is computationally efficient for extracting these modes.
Kai: And beyond just the structure of the bands themselves, they show that you can infer topological invariants like winding numbers or Chern numbers through discrete gauge invariant formulas that are formally parallel to those computed in Hamiltonian Floquet-Bloch analysis.
Lev: Inferring topology from data is ambitious. If you’re building an error correction code, knowing the topological properties without solving the full Hamiltonian analytically would save us time, but what if the DMD modes are just noisy representations of something else?
Mira: They show that you can reconstruct the quantum geometric tensor by estimating the quantum metric from the fidelity between adjacent dynamic mode decomposition modes. This metric is insensitive to local phase fluctuations, which is a nice feature for robustness.
Kai: So, it’s about using data to get these topological and geometric properties, which we normally derive from a known Hamiltonian's structure. It bridges that gap between theoretical modeling and experimental observation in some way.
The paper's summary: Lev: If the authors are looking at this as an improvement over existing methods, what specific advantages are they claiming for their Koopman dynamic mode decomposition approach compared to traditional Floquet-Bloch methods?
Mira: They present a comparative formulation of Hamiltonian Floquet-Bloch decomposition and Koopman dynamics. The key difference they highlight is that the Floquet-Bloch method is equation based, while Koopman DMD is equation free but data driven.
Kai: So the improvement lies in this shift from top-down derivation to bottom-up inference, where you don't need a fully specified microscopic model to get these results. They claim that DMD can recover the band gap and group velocity, and group velocity dispersion directly from spatiotemporal data instead of an explicit Hamiltonian.
Lev: That’s appealing because it reduces the requirement for having perfect knowledge of the underlying physics at every single point in space or time. But what about systems where we only have limited measurements?
Mira: They show that you can reconstruct band dispersion and modal dynamics directly from spatiotemporal data including wavefunctions and observables, which is a key improvement over methods that rely on prior knowledge of the governing equations.
Kai: And they also claim this method is powerful enough to capture the crossover from spatially extended states to exponentially localized states in models like the Anderson model by analyzing things like the inverse participation ratio. That’s how you diagnose localization without solving for it first.
Lev: Diagnosing localization through an observable like the inverse participation ratio is helpful, but what if we're looking at something more complex than just simple localization? Like driven non-Hermitian systems where things are really messy?
Mira: They show that using the Center of Mass distribution of DMD modes can clearly distinguish between topological edge pinning, Non-Hermitian skin effect induced boundary accumulation, and random Anderson localization in these driven non-Hermitian systems.
The paper's improvements: Kai: So, to wrap up the paper "Data-driven reconstruction of band dispersion and quantum geometry via Koopman dynamic mode decomposition," the main point is that we can use data to reconstruct quasiparticle excitations and spatiotemporal coherent structures from data without needing a fully specified microscopic model.
Mira: And they show that this framework, by bridging Hamiltonian Floquet-Bloch decomposition and Koopman-DMD, allows us to infer band structure features, topological invariants like the quantum metric, and even the quantum geometric tensors directly from experimental observations.
Lev: From an error correction standpoint, it’s interesting because it offers a way to probe topological properties without requiring a full Hamiltonian diagonalization upfront. But I still have my concerns about how well this scales when we move this from the numerical demonstrations on lattices to actual hardware with limited measurement capability.
Kai: I agree, but the demonstration shows that for models like SSH and graphene, the extracted DMD modes agree closely with exact Floquet-Bloch results. It’s a strong link between data and theory in those specific settings.
Mira: So, while it's not a replacement for every theoretical tool, this paper provides a way to look at complex systems from the bottom up through observation rather than starting top down with an equation. We should definitely keep an eye on how this is applied to driven, disordered systems where we don't have perfect knowledge of the system dynamics.
Lev: Yeah, so it’s a complementary framework, where Hamiltonian methods derive structure from equations top-down and Koopman dynamic mode decomposition infers dynamics bottom-up from observation. That complementarity is really what makes this work useful for driven or disordered systems where the underlying model isn't perfectly known yet.
Kai: That's the picture we have now for this paper. It gives us a data-driven route to understanding band dispersion and quantum geometry by analyzing spatiotemporal data through Koopman dynamic mode decomposition.
Conclusion: Kai: So, we're wrapping up on "Data-driven reconstruction of band dispersion and quantum geometry via Koopman dynamic mode decomposition." The main idea is that you can figure out band structure and quantum geometry by looking at time-resolved data without needing a full equation for the system.
Mira: Exactly. They show that this Koopman dynamic mode decomposition method lets us extract things like the band gap and even the quantum metric from just observing spatiotemporal data, rather than starting with an explicit Hamiltonian model.
Lev: From an error-correction standpoint, that's a big deal because it means we can probe topological phases without having to solve for them analytically first. But I still wonder about the overhead for real hardware; how much measurement record do you actually need before this method becomes practical?
Kai: They tested it on models like the SSH and graphene systems, and those results matched what we get from standard Floquet-Bloch analysis. It shows the method is robust enough to work across different types of lattice physics.
Mira: The paper also highlights how you can reconstruct topological invariants like winding numbers or Chern numbers using formulas that are formally similar to the ones used in traditional Hamiltonian methods. That's a direct link between data and topology.
Lev: Inferring those topological properties from data is ambitious, but if the fidelity holds up under noisy measurements, it could be useful for characterizing real-world materials where we don't have perfect control over every parameter.
Kai: They also showed you can diagnose localization regimes by looking at how the DMD modes are distributed in energy versus the inverse participation ratio plane. That helps separate extended states from localized ones clearly.
Mira: And they even found a way to reconstruct the quantum metric and Berry curvature of those dynamic modes using gauge invariant formulas derived from link variables. That’s getting into genuine quantum geometry inference here.
Lev: That geometric tensor reconstruction is what I find most interesting for hardware applications; if we can map that to physical observables, it could give us new metrics to monitor the health of our qubit states.
Kai: So, in summary, this paper shows that Koopman dynamic mode decomposition provides a data-driven route to extract effective quasiparticle excitations and spatiotemporal coherent structures from data without needing a fully specified microscopic model.
Mira: It’s a powerful tool because it bridges the gap between rigorous theoretical modeling and mode-resolved inference from experimental data, especially in driven or disordered systems.
Lev: It's a good framework for checking what we think is happening in complex systems, but we still need to figure out the practical measurement limits before we can really rely on this for designing new hardware architectures.
Kai: That's right. We’ve seen how this Koopman dynamic mode decomposition approach lets us reconstruct band dispersion and quantum geometry directly from spatiotemporal data. Next up, we’re looking at how AI is changing the game in inverse design of free energy landscapes with diffusion models.
ShanghaiTech University · Newcastle University
physics.comp-ph
Submitted: 2026-05-07
Updated: 2026-11-17
Comments: 28 pages, 6 figures, 1 table
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 91/100
The gist: Data-driven reconstruction of band dispersion and quantum geometry via Koopman dynamic mode decomposition presents a framework to reconstruct band structures and topological properties directly from…
Key concepts
- Hamiltonian Floquet-Bloch Decomposition (FBD)
- This is a traditional method where physical properties like energy bands are calculated using an explicit mathematical model called a Hamiltonian. It relies on knowing the exact governing equations of the system, which can be complex, making it equation-based and requiring prior knowledge.
- Koopman Dynamic Mode Decomposition (DMD)
- This is a data-driven technique that approximates complex spatiotemporal data by decomposing it into fundamental spatial modes and their temporal dynamics. It works directly on observed data, allowing researchers to infer physical properties like band structure without needing the underlying equations.
- Quantum Metric
- The quantum metric is a geometric property of the system's quantum states that describes how distances and angles are measured in its state space. This paper shows it can be reconstructed from DMD modes using fidelity calculations between adjacent modes, providing insight into the system's underlying geometry.
- Inverse Participation Ratio (IPR)
- The IPR is a measure used to diagnose whether a state or mode is localized or extended in space. By analyzing how the IPR changes as disorder increases, researchers can distinguish between different physical regimes, such as extended states and exponentially localized states.
Terminology
Summary
Data-driven reconstruction of band dispersion and quantum geometry via Koopman dynamic mode decomposition presents a framework to reconstruct band structures and topological properties directly from spatiotemporal data without relying on an explicit Hamiltonian <ref:2605. p. 0 1
Data-driven reconstruction of band dispersion and quantum geometry via
Koopman dynamical mode decomposition
Yiming Pan1*
, Jinze He1
, Jiapeng Yang1
, Zhiwei Fan2
1
State Key Laboratory of Quantum Functional Materials, School of Physical Science and
Technology and Center for Transformative Science, ShanghaiTech University, Shanghai 200031,
China
2
School of Engineering, Newcastle University, Newcastle upon Tyne, NE1 7RU, UK
Abstract
We present a data-driven framework for reconstructing band structures using Koopman operator
analysis and dynamic mode decomposition (Koopman-DMD). Instead of deriving spectra from an
explicit Hamiltonian, the approach reconstructs band dispersion and modal dynamics directly from
spatiotemporal data, including wavefunctions and observables. This framework establishes a
correspondence between Hamiltonian Floquet-Bloch decomposition and Koopman-DMD,
whereby the extracted DMD modes encode frequencies, decay or growth rates, spatial profiles and
projection weights. These quantities allow the reconstruction of spectral functions, local density
of states, and delocalized-to-localized measures such as the inverse participation ratio.Where the paper says it
Where the paper says it
Framework Comparison
The paper presents a comparative formulation of Hamiltonian Floquet-Bloch decomposition (FBD) and Koopman dynamical mode decomposition (DMD), highlighting their methodological difference where FBD is equation-based and Koopman DMD is equation-free but data-driven <ref:2605.06199#pg4> The primary distinction lies in how physical insight is derived, as FBD depends on the accuracy of a model Hamiltonian, whereas Koopman-DMD encodes it within the dataset itself <ref:2605. p. 2 When the data are sparse, noisy, or poorly resolved,
the extracted DMD modes become unreliable. In Hamiltonian FBD, physical insight depends on
the accuracy of the model Hamiltonian and governing equations, whereas in Koopman-DMD, it is
encoded in the dataset itself.Where the paper says it
Basic DMD and FBD
The basic DMD framework approximates complex spatiotemporal data by decomposing them into spatial modes, modal amplitudes, and temporal dynamics <ref:2605.06199#pg5> This decomposition allows the data to be written as a sum of extended modes and localized modes <ref:2605. p. 2 In this decomposition, the spatial modes 𝜙j
represent coherent structures in space, including
extended waves with characteristic momenta and localized modes with finite confinement lengths,
where the complex frequencies 𝜔𝑗 determine the temporal evolution of each mode, including
oscillation and decay or growth. Accordingly,
𝑋 ≈ [
| | |
𝜙1 ⋯ 𝜙𝑟
| | |
] [
𝑤1
⋱
𝑤𝑟
] [
𝑒
−𝑖𝜔1𝑡1 ⋯ 𝑒
−𝑖𝜔1𝑡𝑚
⋮ ⋱ ⋮
𝑒
−𝑖𝜔𝑟𝑡1 ⋯ 𝑒
−𝑖𝜔𝑟𝑡𝑚
], (1)
so that the data can be written as Basic FBD framework. The Floquet–Bloch theorem provides a standard framework for systems
with spatial and temporal periodicity.Where the paper says it
Where the paper says it
Band Physics Extraction
In the Koopman-DMD framework, band dispersion is extracted by projecting each extended DMD mode into momentum space to yield a band relation E(k) ≃ ħω(k) <ref:2605.06199#pg5> The paper demonstrates that DMD can recover the band gap, group velocity, and group-velocity dispersion from spatiotemporal data rather than an explicit Hamiltonian <ref:2605. p. 3 Once the eigenvalues and eigenstates are obtained, the Hamiltonian framework can be used to
construct several classes of physical observables. Firstly, the band structure determines dynamical
properties such as the energy gap, group velocity, and group-velocity dispersion, which controls
pulse broadening and nonlinear propagation. This framework establishes a
correspondence between Hamiltonian Floquet-Bloch decomposition and Koopman-DMD,
whereby the extracted DMD modes encode frequencies, decay or growth rates, spatial profiles and
projection weights. These quantities allow the reconstruction of spectral functions, local density
of states, and delocalized-to-localized measures such as the inverse participation ratio.Where the paper says it
Where the paper says it
Topological Invariants and Geometry
Topological invariants are inferred from DMD modes through discrete gauge-invariant formulas, such as the Zak phase in 1D or the Chern number in 2D, which are formally parallel to those computed in Hamiltonian FBD <ref:2605.06199#pg5> The quantum geometric tensor (QGT) can be reconstructed by estimating the quantum metric from the fidelity between adjacent DMD modes, which is insensitive to local phase fluctuations <ref:2605. p. 8 For instance, the real part of the QGT, namely the quantum metric 𝑔𝜇𝜈, measures the Fubini-Study
distance between nearby quantum states. Rather than evaluating derivatives explicitly, the discrete
quantum metric can be extracted from the fidelity (inner product overlap) between adjacent DMD
modes:
𝑑𝑠
2 = 𝑔𝜇𝜈𝑑𝑘
𝜇𝑑𝑘
𝜈 ≈ 1 − |⟨𝑢
𝐷𝑀𝐷(𝑘𝑖
)|𝑢
𝐷𝑀𝐷(𝑘𝑗)⟩|
2
.Where the paper says it
Benchmark Results
The study validates the Koopman-DMD approach across several tight-binding models, including disordered Su-Schrieffer-Heeger (SSH) and graphene models <ref:2605.06199#pg5> The results show that DMD can resolve the crossover from spatially extended states to exponentially localized states in the Anderson model by analyzing the Inverse Participation Ratio (IPR) <ref:2605. p. 8 Applying DMD to spatiotemporal
wavefunction yields spatial modes ϕ𝑗
(𝑥) that, in the localized regime, exhibit exponential
envelopes,
|ϕ𝑗
(𝑥)| ∼ 𝑒
−|𝑥−𝑥j
|/ξj
, (13)
from which the localization length ξj can be estimated by envelope fitting or, more robustly, by
through inverse participation ratio (IPR) [18]. The disappearance of extended modes and the
emergence of localized modes in the DMD spectrum therefore provide a direct dynamical
signature of the localization-to-delocalization Anderson transition (see Fig. This can be
implemented using the Fukui–Hatsugai–Suzuki (FHS) construction [37], in which the link variable
𝑈𝛿
(k) =
⟨ 𝑢(𝑘) | 𝑢(𝑘 + 𝛿) ⟩
|⟨ 𝑢(𝑘) | 𝑢(𝑘 + 𝛿) ⟩|
, (10)
removes the arbitrary 𝑈(1) phase of independently extracted DMD modes. In this way, the
computation of Chern numbers or winding numbers from DMD modes becomes formally parallel
to that in Hamiltonian FBD, while retaining sensitivity to boundaries, defects, and finite-size
structure.Where the paper says it
Where the paper says it
Koopman Perspective on Observables
The appropriate variant of DMD is determined by the physical nature of the data, with exact DMD providing a natural approximation to the time-evolution operator for state-level data like wavefunctions <ref:2605.06199#pg5> For strongly projected observables, time-delayed (Hankel) DMD is more robust because it induces effective memory, and its extracted modes are better interpreted as effective density waves or scattering channels <ref:2605.06199#pg5>
Finale Remarks
Koopman-DMD provides a data-driven route to extracting effective quasiparticle excitations and spatiotemporal coherent structures from data without requiring a fully specified microscopic model <ref:2605.06199#pg5> Hamiltonian FBD and Koopman-DMD are complementary frameworks, where Hamiltonian methods derive structure from equations top-down, and Koopman-DMD infers dynamics bottom-up from observation <ref:2605. p. 2 Their primary difference is methodological:
Hamiltonian FBD is equation-based, whereas Koopman DMD is equation-free but data-driven. Here we focus on crystalline solid-state and photonic lattice systems, for which the relevant
descriptions include the Schrödinger equation, Maxwell’s equations, and reduced forms such as
tight-binding models, and slow-varying-enveloped coupled-mode equations. Koopman–DMD is particularly useful when
disorder creates or suppresses boundary-localized modes. In this sense, DMD provides a datadriven route to detecting topological transitions in finite disordered systems.Where the paper says it
Where the paper says it
Acknowledgments
Y. P. acknowledges the support of the NSFC (No. p. 1 The conventional equation-based Floquet–Bloch decomposition can, in principle, be
complemented by a data-driven Koopman–DMD decomposition performed directly on
spatiotemporal observables. Without prior knowledge of the equations of motion or the detailed
system parameters, Koopman–DMD enables the direct extraction of Floquet–Bloch-like extended
modes, as well as localized modes, from state data, and may further support reconstruction of an
effective Hamiltonian and its associated band structure. 1
Data-driven reconstruction of band dispersion and quantum geometry via
Koopman dynamical mode decomposition
Yiming Pan1*
, Jinze He1
, Jiapeng Yang1
, Zhiwei Fan2
1
State Key Laboratory of Quantum Functional Materials, School of Physical Science and
Technology and Center for Transformative Science, ShanghaiTech University, Shanghai 200031,
China
2
School of Engineering, Newcastle University, Newcastle upon Tyne, NE1 7RU, UK
Abstract
We present a data-driven framework for reconstructing band structures using Koopman operator
analysis and dynamic mode decomposition (Koopman-DMD). Instead of deriving spectra from an
explicit Hamiltonian, the approach reconstructs band dispersion and modal dynamics directly from
spatiotemporal data, including wavefunctions and observables.Where the paper says it
Where the paper says it
Data and Code Availability
The code repository will be made publicly available on acceptance through an open-source link <ref:2605.06199#pg5>
Conflict of Interest
The authors declare no competing interests <ref:2605.06199#pg5>
References: [1] P. Coleman, Introduction to Many-Body Physics (Cambridge University Press, 2015). p. 2 3
disordered, periodically-driven, non-Hermitian, nonlinear, and many-body interacting systems
from data, particularly through the identification of coherent oscillatory modes and spatial
structures. In this paper, we establish a theoretical connection between Koopman–DMD analysis and band
structures obtained from the Floquet–Bloch formalism, and support this connection with numerical
demonstrations in prototypical one- and two-dimensional lattice models. By designing periodicity and material
composition, one can create targeted band dispersion and bandgaps that govern electrical transport,
optical absorption and emission, and light propagation. In solid-state physics, for example, bandstructure engineering modifies semiconductor bandgaps to regulate conductivity and optical
response, enabling devices such as transistors, laser diodes, and high-efficiency solar cells. 2
Introduction
Band structure engineering seeks to manipulate the spatial or spatiotemporal periodicity of
materials so as to tailor the energy ranges (bands) in which electrons in solids [1,2] or photons in
dielectrics [3] are allowed or forbidden to propagate. By designing periodicity and material
composition, one can create targeted band dispersion and bandgaps that govern electrical transport,
optical absorption and emission, and light propagation. By designing periodicity and material
composition, one can create targeted band dispersion and bandgaps that govern electrical transport,
optical absorption and emission, and light propagation. In solid-state physics, for example, bandstructure engineering modifies semiconductor bandgaps to regulate conductivity and optical
response, enabling devices such as transistors, laser diodes, and high-efficiency solar cells. Disorder and defects further break translational symmetry, limiting
momentum-space analysis and often necessitating computationally expensive real-space supercell
calculations. In addition, periodically driven (Floquet) systems require a quasienergy band
description of nonequilibrium dynamics [8–10], which further increases theoretical and numerical
complexity. The Floquet–Bloch theorem provides a standard framework for systems
with spatial and temporal periodicity. For a periodically driven quantum system with
𝐻(𝒓 + 𝑹,𝑡) = 𝐻(𝒓,𝑡) and 𝐻(𝒓,𝑡 + 𝑇) = 𝐻(𝒓,𝑡), where 𝑹 is a lattice vector and 𝑇 is the driving
period, the wavefunction can be written in Bloch form, ψ𝑛,𝒌
(𝒓,𝑡) = 𝑒
𝑖𝒌⋅𝒓𝑢𝑛,𝒌
(𝒓,𝑡), where 𝑢𝑛,𝒌 is
lattice-periodic in space. In addition, periodically driven (Floquet) systems require a quasienergy band
description of nonequilibrium dynamics [8–10], which further increases theoretical and numerical
complexity. To address these challenges, data-driven approaches have been increasingly important in modern
physics. Applied to DMD modes, the IPR provides a practical measure of real-space mode
concentration and hence of localization. The same criterion can be generalized to other settings,
including spin chains, and many-body Hilbert or Fock spaces [18]. Firstly, the band structure determines dynamical
properties such as the energy gap, group velocity, and group-velocity dispersion, which controls
pulse broadening and nonlinear propagation. In systems with boundaries, defects, or disorder, the
same framework also yields localized states such as edge and defect modes. Koopman operator theory has emerged as a powerful framework in fluid
dynamics, control theory, pure mathematics, and, more recently, physics [17,27,28]. Its central
idea is to lift a finite-dimensional nonlinear dynamical system into a linear evolution in an infinitedimensional observable space. 1
Data-driven reconstruction of band dispersion and quantum geometry via
Koopman dynamical mode decomposition
Yiming Pan1*
, Jinze He1
, Jiapeng Yang1
, Zhiwei Fan2
1
State Key Laboratory of Quantum Functional Materials, School of Physical Science and
Technology and Center for Transformative Science, ShanghaiTech University, Shanghai 200031,
China
2
School of Engineering, Newcastle University, Newcastle upon Tyne, NE1 7RU, UK
Abstract
We present a data-driven framework for reconstructing band structures using Koopman operator
analysis and dynamic mode decomposition (Koopman-DMD). Instead of deriving spectra from an
explicit Hamiltonian, the approach reconstructs band dispersion and modal dynamics directly from
spatiotemporal data, including wavefunctions and observables. Koopman linearization. Koopman operator theory has emerged as a powerful framework in fluid
dynamics, control theory, pure mathematics, and, more recently, physics [17,27,28]. 1
Data-driven reconstruction of band dispersion and quantum geometry via
Koopman dynamical mode decomposition
Yiming Pan1*
, Jinze He1
, Jiapeng Yang1
, Zhiwei Fan2
1
State Key Laboratory of Quantum Functional Materials, School of Physical Science and
Technology and Center for Transformative Science, ShanghaiTech University, Shanghai 200031,
China
2
School of Engineering, Newcastle University, Newcastle upon Tyne, NE1 7RU, UK
Abstract
We present a data-driven framework for reconstructing band structures using Koopman operator
analysis and dynamic mode decomposition (Koopman-DMD). Instead of deriving spectra from an
explicit Hamiltonian, the approach reconstructs band dispersion and modal dynamics directly from
spatiotemporal data, including wavefunctions and observables. 4
x(t) ≈ ∑𝜙𝑗𝑤𝑗𝑒
−𝑖𝜔𝑗
𝑡
𝑟
𝑗=1
= ∑𝑢𝑖(𝑘)𝑤𝑖𝑒
−𝑖𝜔𝑖
𝑡+𝑖𝒌𝒊
⋅𝒓
′
⏟𝑖
Extended modes
+∑𝜙𝑗𝑤𝑗𝑒
−𝑖𝜔𝑗
𝑡
′′
⏟𝑗
Localized modes
. (2)
We broadly classify the DMD modes into extended and localized branches, which will be later
quantified using the inverse participation ratio (IPR) [18–20]. Also, these
extended DMD modes enable inference of quantum-geometric and topological properties,
including the quantum metric, Berry curvature and geometric phases. Applications to prototypical
one- and two-dimensional tight-binding models, including disordered Su-Schrieffer-Heeger model
and its Floquet and non-Hermitian variants, graphene and Haldane models, demonstrate that
Koopman-DMD provides a unified route for the data-driven analysis of wave propagation,
localization, and topological phases in condensed matter, photonics, and related fields. Also, these
extended DMD modes enable inference of quantum-geometric and topological properties,
including the quantum metric, Berry curvature and geometric phases. Applications to prototypical
one- and two-dimensional tight-binding models, including disordered Su-Schrieffer-Heeger model
and its Floquet and non-Hermitian variants, graphene and Haldane models, demonstrate that
Koopman-DMD provides a unified route for the data-driven analysis of wave propagation,
localization, and topological phases in condensed matter, photonics, and related fields. Also, these
extended DMD modes enable inference of quantum-geometric and topological properties,
including the quantum metric, Berry curvature and geometric phases. Applications to prototypical
one- and two-dimensional tight-binding models, including disordered Su-Schrieffer-Heeger model
and its Floquet and non-Hermitian variants, graphene and Haldane models, demonstrate that
Koopman-DMD provides a unified route for the data-driven analysis of wave propagation,
localization, and topological phases in condensed matter, photonics, and related fields. 1
Data-driven reconstruction of band dispersion and quantum geometry via
Koopman dynamical mode decomposition
Yiming Pan1*
, Jinze He1
, Jiapeng Yang1
, Zhiwei Fan2
1
State Key Laboratory of Quantum Functional Materials, School of Physical Science and
Technology and Center for Transformative Science, ShanghaiTech University, Shanghai 200031,
China
2
School of Engineering, Newcastle University, Newcastle upon Tyne, NE1 7RU, UK
Abstract
We present a data-driven framework for reconstructing band structures using Koopman operator
analysis and dynamic mode decomposition (Koopman-DMD). Instead of deriving spectra from an
explicit Hamiltonian, the approach reconstructs band dispersion and modal dynamics directly from
spatiotemporal data, including wavefunctions and observables. 1
Data-driven reconstruction of band dispersion and quantum geometry via
Koopman dynamical mode decomposition
Yiming Pan1*
, Jinze He1
, Jiapeng Yang1
, Zhiwei Fan2
1
State Key Laboratory of Quantum Functional Materials, School of Physical Science and
Technology and Center for Transformative Science, ShanghaiTech University, Shanghai 200031,
China
2
School of Engineering, Newcastle University, Newcastle upon Tyne, NE1 7RU, UK
Abstract
We present a data-driven framework for reconstructing band structures using Koopman operator
analysis and dynamic mode decomposition (Koopman-DMD). Instead of deriving spectra from an
explicit Hamiltonian, the approach reconstructs band dispersion and modal dynamics directly from
spatiotemporal data, including wavefunctions and observables. Also, these
extended DMD modes enable inference of quantum-geometric and topological properties,
including the quantum metric, Berry curvature and geometric phases. Applications to prototypical
one- and two-dimensional tight-binding models, including disordered Su-Schrieffer-Heeger model
and its Floquet and non-Hermitian variants, graphene and Haldane models, demonstrate that
Koopman-DMD provides a unified route for the data-driven analysis of wave propagation,
localization, and topological phases in condensed matter, photonics, and related fields. 1
Data-driven reconstruction of band dispersion and quantum geometry via
Koopman dynamical mode decomposition
Yiming Pan1*
, Jinze He1
, Jiapeng Yang1
, Zhiwei Fan2
1
State Key Laboratory of Quantum Functional Materials, School of Physical Science and
Technology and Center for Transformative Science, ShanghaiTech University, Shanghai 200031,
China
2
School of Engineering, Newcastle University, Newcastle upon Tyne, NE1 7RU, UK
Abstract
We present a data-driven framework for reconstructing band structures using Koopman operator
analysis and dynamic mode decomposition (Koopman-DMD). Instead of deriving spectra from an
explicit Hamiltonian, the approach reconstructs band dispersion and modal dynamics directly from
spatiotemporal data, including wavefunctions and observables. 1
Data-driven reconstruction of band dispersion and quantum geometry via
Koopman dynamical mode decomposition
Yiming Pan1*
, Jinze He1
, Jiapeng Yang1
, Zhiwei Fan2
1
State Key Laboratory of Quantum Functional Materials, School of Physical Science and
Technology and Center for Transformative Science, ShanghaiTech University, Shanghai 200031,
China
2
School of Engineering, Newcastle University, Newcastle upon Tyne, NE1 7RU, UK
Abstract
We present a data-driven framework for reconstructing band structures using Koopman operator
analysis and dynamic mode decomposition (Koopman-DMD). Instead of deriving spectra from an
explicit Hamiltonian, the approach reconstructs band dispersion and modal dynamics directly from
spatiotemporal data, including wavefunctions and observables. Given a sequence of snapshots
𝑥(𝑡𝑘
), we form the data matrix 𝑋 = [𝑥(𝑡1) 𝑥(𝑡2) ⋯ 𝑥(𝑡𝑚)], and seek a low rank-r approximation
of the form 𝑋 ≈ Φ 𝑑𝑖𝑎𝑔(𝑤)𝑇, where Φ = [ϕ1, …, ϕr
] contains the spatial DMD modes, 𝑤 =
(𝑤1, … , 𝑤𝑟
)
𝑇 gives the projection weights and the temporal matrix 𝑇 is built from exponentials
𝑒
𝑖𝜔𝑖
𝑡
. In this decomposition, the spatial modes 𝜙j
represent coherent structures in space, including
extended waves with characteristic momenta and localized modes with finite confinement lengths,
where the complex frequencies 𝜔𝑗 determine the temporal evolution of each mode, including
oscillation and decay or growth. DMD approximates complex spatiotemporal data by decomposing them
into spatial modes, modal amplitudes, and temporal dynamics. Given a sequence of snapshots
𝑥(𝑡𝑘
), we form the data matrix 𝑋 = [𝑥(𝑡1) 𝑥(𝑡2) ⋯ 𝑥(𝑡𝑚)], and seek a low rank-r approximation
of the form 𝑋 ≈ Φ 𝑑𝑖𝑎𝑔(𝑤)𝑇, where Φ = [ϕ1, …, ϕr
] contains the spatial DMD modes, 𝑤 =
(𝑤1, … , 𝑤𝑟
)
𝑇 gives the projection weights and the temporal matrix 𝑇 is built from exponentials
𝑒
𝑖𝜔𝑖
𝑡
. 1
Data-driven reconstruction of band dispersion and quantum geometry via
Koopman dynamical mode decomposition
Yiming Pan1*
, Jinze He1
, Jiapeng Yang1
, Zhiwei Fan2
1
State Key Laboratory of Quantum Functional Materials, School of Physical Science and
Technology and Center for Transformative Science, ShanghaiTech University, Shanghai 200031,
China
2
School of Engineering, Newcastle University, Newcastle upon Tyne, NE1 7RU, UK
Abstract
We present a data-driven framework for reconstructing band structures using Koopman operator
analysis and dynamic mode decomposition (Koopman-DMD). Instead of deriving spectra from an
explicit Hamiltonian, the approach reconstructs band dispersion and modal dynamics directly from
spatiotemporal data, including wavefunctions and observables. 1
Data-driven reconstruction of band dispersion and quantum geometry via
Koopman dynamical mode decomposition
Yiming Pan1*
, Jinze He1
, Jiapeng Yang1
, Zhiwei Fan2
1
State Key Laboratory of Quantum Functional Materials, School of Physical Science and
Technology and Center for Transformative Science, ShanghaiTech University, Shanghai 200031,
China
2
School of Engineering, Newcastle University, Newcastle upon Tyne, NE1 7RU, UK
Abstract
We present a data-driven framework for reconstructing band structures using Koopman operator
analysis and dynamic mode decomposition (Koopman-DMD). Instead of deriving spectra from an
explicit Hamiltonian, the approach reconstructs band dispersion and modal dynamics directly from
spatiotemporal data, including wavefunctions and observables. Another practical approach is to infer topology from dynamical
observables whose long-time behavior is known to reflect topological invariants. In onedimensional chiral systems, for instance, the mean chiral displacement converges to the Zak phase
and thus provides a dynamical signature of topology without direct access to Bloch-state phases. In solid-state physics, for example, bandstructure engineering modifies semiconductor bandgaps to regulate conductivity and optical
response, enabling devices such as transistors, laser diodes, and high-efficiency solar cells. In
photonic crystals (PhCs), it controls light propagation and confinement through periodic dielectric
structures [4,5], producing photonic bandgaps and engineered dispersion for applications such as
high-reflectivity mirrors, PhC fibers and optical cavities. Although graphene has zero total Chern number
with time-reversal symmetry breaking, each Dirac valley still carries a nontrivial Berry phase,
underlying its quantum Hall effect. We will discuss the band topology of Haldane model [35] and
graphene [36] in the following sections. The Floquet–Bloch theorem is widely used
to compute band structure and identify band modes, defect states within gaps, and group-velocity
dispersion (GVD) in periodic media. Here we focus on electronic and photonic systems, for which
the band structure – that is, the energy-momentum relation ε𝑛
(𝑘) – is central to understanding the
properties of solids and dielectrics. For instance, the Zak phase can be evaluated using the
gauge-invariant discretized formula
γ𝑍𝑎𝑘 = − Im ln∏⟨ 𝑢𝑘𝑗
𝐷𝑀𝐷 | 𝑢𝑘𝑗+1
𝐷𝑀𝐷 ⟩
𝑁
𝑗=1
, (9)
Similarly, the Berry flux through each plaquette is obtained from Wilson-loop overlap products,
and summation over the discretized Brillouin zone yields the Chern number. This can be
implemented using the Fukui–Hatsugai–Suzuki (FHS) construction [37], in which the link variable
𝑈𝛿
(k) =
⟨ 𝑢(𝑘) | 𝑢(𝑘 + 𝛿) ⟩
|⟨ 𝑢(𝑘) | 𝑢(𝑘 + 𝛿) ⟩|
, (10)
removes the arbitrary 𝑈(1) phase of independently extracted DMD modes. Without prior knowledge of the equations of motion or the detailed
system parameters, Koopman–DMD enables the direct extraction of Floquet–Bloch-like extended
modes, as well as localized modes, from state data, and may further support reconstruction of an
effective Hamiltonian and its associated band structure. The extracted DMD modes can also be
used to infer phase transition, such as the Anderson transition, as well as band topology and
quantum geometry, including Chern numbers, topological edge states, Berry curvature and the
quantum metric. In two dimensions, the corresponding invariant is often the Chern
number,
𝐶 =
1
2𝜋
∬ Ω(𝑘𝑥, 𝑘𝑦)
BZ
𝑑𝑘𝑥𝑑𝑘𝑦, (8)
where Ω = ∇k × 𝒜(𝑘) is the Berry curvature. Although graphene has zero total Chern number
with time-reversal symmetry breaking, each Dirac valley still carries a nontrivial Berry phase,
underlying its quantum Hall effect. Also, these
extended DMD modes enable inference of quantum-geometric and topological properties,
including the quantum metric, Berry curvature and geometric phases. Applications to prototypical
one- and two-dimensional tight-binding models, including disordered Su-Schrieffer-Heeger model
and its Floquet and non-Hermitian variants, graphene and Haldane models, demonstrate that
Koopman-DMD provides a unified route for the data-driven analysis of wave propagation,
localization, and topological phases in condensed matter, photonics, and related fields. The Berry
connection is 𝐴𝑛𝑚(𝑘) = 𝑖⟨𝑢𝑛,𝑘
|∇𝑘
|𝑢𝑚,𝑘
⟩ [21], the Berry phase is obtained by integrating the
diagonal Berry connection along a closed loop in momentum space, and quantum geometric tensor
can be further defined accordingly [22–26]. Thirdly, for interacting or many-body systems, one
often introduces Green’s functions within a mean-field or self-energy framework, 𝐺(ω, 𝑘) =
(ω − 𝐻(𝑘) − Σ(ω, 𝑘))
−1
from which the spectral function, local density of states, and
quasiparticle lifetime can be derived. Also, these
extended DMD modes enable inference of quantum-geometric and topological properties,
including the quantum metric, Berry curvature and geometric phases. Applications to prototypical
one- and two-dimensional tight-binding models, including disordered Su-Schrieffer-Heeger model
and its Floquet and non-Hermitian variants, graphene and Haldane models, demonstrate that
Koopman-DMD provides a unified route for the data-driven analysis of wave propagation,
localization, and topological phases in condensed matter, photonics, and related fields. It enables reconstruction not
only of topological invariants but also of the quantum geometric tensors of the data-driven band
structure. Anderson localization and phase transition. In solid-state physics, for example, bandstructure engineering modifies semiconductor bandgaps to regulate conductivity and optical
response, enabling devices such as transistors, laser diodes, and high-efficiency solar cells. In
photonic crystals (PhCs), it controls light propagation and confinement through periodic dielectric
structures [4,5], producing photonic bandgaps and engineered dispersion for applications such as
high-reflectivity mirrors, PhC fibers and optical cavities. 3
disordered, periodically-driven, non-Hermitian, nonlinear, and many-body interacting systems
from data, particularly through the identification of coherent oscillatory modes and spatial
structures. In this paper, we establish a theoretical connection between Koopman–DMD analysis and band
structures obtained from the Floquet–Bloch formalism, and support this connection with numerical
demonstrations in prototypical one- and two-dimensional lattice models. Also, these
extended DMD modes enable inference of quantum-geometric and topological properties,
including the quantum metric, Berry curvature and geometric phases. Applications to prototypical
one- and two-dimensional tight-binding models, including disordered Su-Schrieffer-Heeger model
and its Floquet and non-Hermitian variants, graphene and Haldane models, demonstrate that
Koopman-DMD provides a unified route for the data-driven analysis of wave propagation,
localization, and topological phases in condensed matter, photonics, and related fields. Also, these
extended DMD modes enable inference of quantum-geometric and topological properties,
including the quantum metric, Berry curvature and geometric phases. Applications to prototypical
one- and two-dimensional tight-binding models, including disordered Su-Schrieffer-Heeger model
and its Floquet and non-Hermitian variants, graphene and Haldane models, demonstrate that
Koopman-DMD provides a unified route for the data-driven analysis of wave propagation,
localization, and topological phases in condensed matter, photonics, and related fields. In
non-Hermitian or dissipative systems, nonzero γ𝑗 measures modal amplification or decay. The spatial profile 𝜙𝑗
(𝑥) gives the corresponding mode shape in real or configuration space. Also, these
extended DMD modes enable inference of quantum-geometric and topological properties,
including the quantum metric, Berry curvature and geometric phases. Applications to prototypical
one- and two-dimensional tight-binding models, including disordered Su-Schrieffer-Heeger model
and its Floquet and non-Hermitian variants, graphene and Haldane models, demonstrate that
Koopman-DMD provides a unified route for the data-driven analysis of wave propagation,
localization, and topological phases in condensed matter, photonics, and related fields. As a result,
observable-driven DMD often requires multiple initial conditions or repeated measurements to
reconstruct the full band structure and its topology reliably. Numerically, DMD in the Schrödinger
picture is conceptually cleaner and provides the most direct access to band physics when full state
data are available. Also, these
extended DMD modes enable inference of quantum-geometric and topological properties,
including the quantum metric, Berry curvature and geometric phases. Applications to prototypical
one- and two-dimensional tight-binding models, including disordered Su-Schrieffer-Heeger model
and its Floquet and non-Hermitian variants, graphene and Haldane models, demonstrate that
Koopman-DMD provides a unified route for the data-driven analysis of wave propagation,
localization, and topological phases in condensed matter, photonics, and related fields. 3
disordered, periodically-driven, non-Hermitian, nonlinear, and many-body interacting systems
from data, particularly through the identification of coherent oscillatory modes and spatial
structures. In this paper, we establish a theoretical connection between Koopman–DMD analysis and band
structures obtained from the Floquet–Bloch formalism, and support this connection with numerical
demonstrations in prototypical one- and two-dimensional lattice models. For a sufficiently smooth family of DMD bulk modes |𝑢
𝐷𝑀𝐷(𝑘𝑗)⟩, discrete Berry phases can be
computed directly from overlap products. For instance, the Zak phase can be evaluated using the
gauge-invariant discretized formula
γ𝑍𝑎𝑘 = − Im ln∏⟨ 𝑢𝑘𝑗
𝐷𝑀𝐷 | 𝑢𝑘𝑗+1
𝐷𝑀𝐷 ⟩
𝑁
𝑗=1
, (9)
Similarly, the Berry flux through each plaquette is obtained from Wilson-loop overlap products,
and summation over the discretized Brillouin zone yields the Chern number. In two dimensions, the corresponding invariant is often the Chern
number,
𝐶 =
1
2𝜋
∬ Ω(𝑘𝑥, 𝑘𝑦)
BZ
𝑑𝑘𝑥𝑑𝑘𝑦, (8)
where Ω = ∇k × 𝒜(𝑘) is the Berry curvature. Although graphene has zero total Chern number
with time-reversal symmetry breaking, each Dirac valley still carries a nontrivial Berry phase,
underlying its quantum Hall effect. Given a sequence of snapshots
𝑥(𝑡𝑘
), we form the data matrix 𝑋 = [𝑥(𝑡1) 𝑥(𝑡2) ⋯ 𝑥(𝑡𝑚)], and seek a low rank-r approximation
of the form 𝑋 ≈ Φ 𝑑𝑖𝑎𝑔(𝑤)𝑇, where Φ = [ϕ1, …, ϕr
] contains the spatial DMD modes, 𝑤 =
(𝑤1, … , 𝑤𝑟
)
𝑇 gives the projection weights and the temporal matrix 𝑇 is built from exponentials
𝑒
𝑖𝜔𝑖
𝑡
. In this decomposition, the spatial modes 𝜙j
represent coherent structures in space, including
extended waves with characteristic momenta and localized modes with finite confinement lengths,
where the complex frequencies 𝜔𝑗 determine the temporal evolution of each mode, including
oscillation and decay or growth. To address these challenges, data-driven approaches have been increasingly important in modern
physics. Rather than relying on prior knowledge of governing equations [11], this paradigm starts
directly from spatiotemporal data generated by experiments or simulations to uncover latent
physical structure.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 20 ---
Figures: Figure 1 Complementary paradigms for analyzing spatiotemporal periodic systems. A schematic comparison between the traditional equation-based framework and the data-driven approach <ref:2605. p. 1 While techniques such as physics-informed neural networks or symbolic
regression excel at discovering explicit governing equations, Koopman operator theory [12–15]
and dynamic mode decomposition (DMD) [15,16] provide a particularly natural, equation-free
bridge between data and physics by linearly embedding the nonlinear dynamics [17]. The conventional equation-based Floquet–Bloch decomposition can, in principle, be
complemented by a data-driven Koopman–DMD decomposition performed directly on
spatiotemporal observables. Within Hamiltonian Floquet–Bloch decomposition (FBD), band dispersion is
obtained by first specifying the Hamiltonian 𝐻, or 𝐻(𝑡) in the periodically driven case. For a static
periodic lattice, Bloch’s theorem gives 𝜓𝑛𝒌
(𝒓) = 𝑒
𝑖𝒌⋅𝒓𝑢𝑛𝒌
(𝒓), where 𝒌 is the crystal momentum
and 𝑛 labels the band. While techniques such as physics-informed neural networks or symbolic
regression excel at discovering explicit governing equations, Koopman operator theory [12–15]
and dynamic mode decomposition (DMD) [15,16] provide a particularly natural, equation-free
bridge between data and physics by linearly embedding the nonlinear dynamics [17]. The conventional equation-based Floquet–Bloch decomposition can, in principle, be
complemented by a data-driven Koopman–DMD decomposition performed directly on
spatiotemporal observables. For localized modes, by contrast, a sharp momentum is generally ill-defined, and localization is
better characterized by a localization length or inverse participation ratio (IPR). In this way, DMD
extracts the frequency spectrum, decay (or growth) rate, spatial mode profile, and projection
weight directly from data, without prior knowledge of the underlying governing equations.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Figure 2 DMD analysis of the Anderson model and the crossover to localization. (a) Spatiotemporal evolution data seeded from the random. p. 4 Dynamic mode decomposition (DMD) provides a practical finite-data realization of Koopman
spectral analysis. Given a sequence of spatiotemporal snapshots, such as wavefunctions 𝜓(𝑥,𝑡𝑛
),
DMD seeks a best-fit linear evolution operator 𝐴 satisfying 𝜓(𝑡𝑛+1
) ≈ 𝐴 𝜓(𝑡𝑛
). This framework establishes a
correspondence between Hamiltonian Floquet-Bloch decomposition and Koopman-DMD,
whereby the extracted DMD modes encode frequencies, decay or growth rates, spatial profiles and
projection weights. These quantities allow the reconstruction of spectral functions, local density
of states, and delocalized-to-localized measures such as the inverse participation ratio. In practice,
however, this strategy faces several limitations. In many systems, including strongly-correlated
electron systems and nonlinear photonic crystals, the exact Hamiltonian may be unknown or only
partially constrained. More generally, zero-energy localized edge states or systematic sublattice polarization in the
extracted spatial modes can provide strong evidence for nontrivial bulk topology through bulk–
boundary correspondence. For a sufficiently smooth family of DMD bulk modes |𝑢
𝐷𝑀𝐷(𝑘𝑗)⟩, discrete Berry phases can be
computed directly from overlap products. However, exact diagonalization rapidly becomes computationally
expensive as system size grows. For Koopman–DMD, one may instead examine the frequency
distribution and spatial support of localized DMD modes extracted from experimentally or
numerically accessible observables. For localized modes, by contrast, a sharp momentum is generally ill-defined, and localization is
better characterized by a localization length or inverse participation ratio (IPR). In this way, DMD
extracts the frequency spectrum, decay (or growth) rate, spatial mode profile, and projection
weight directly from data, without prior knowledge of the underlying governing equations. Its main strength is
the direct extraction of edge or defect spectral modes from individual samples, whereas global
topological invariants are generally less directly accessible. Notably, for disordered many-body systems, exact diagonalization is often used to distinguish
localized and thermal phases through level-spacing statistics, such as Poisson versus Wigner–
Dyson distributions [44].Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved AI can reconstruct band dispersion and modal dynamics directly from spatiotemporal data, including wavefunctions and observables, rather than deriving spectra from an explicit Hamiltonian. This allows for a
data-driven reconstruction of band dispersion
andmodal dynamics
that are independent of prior knowledge of governing equations. -
Improved AI can infer quantum-geometric properties such as the quantum metric, Berry curvature, and geometric phases directly from the extracted DMD modes. This capability enables the inference of
quantum-geometric and topological properties, including the quantum metric, Berry curvature and geometric phases.
-
Improved AI can perform a unified route for analyzing wave propagation, localization, and topological phases in condensed matter systems across disordered, periodically driven, non-Hermitian models. The system can
directly capture from data, the breakdown of Floquet driven topological phases into Anderson localization
by tracking modal quantities like Center of Mass and Inverse Participation Ratio. -
Improved AI can diagnose the type of physical regime (topological vs. Anderson localization) by analyzing the distribution of DMD modes in the energy inverse participation ratio plane, which
sharply distinguishes extended and localized regimes.
-
Improved AI can distinguish between different physical regimes in driven non-Hermitian systems by analyzing the Center of Mass (COM) distribution of DMD modes, which
clearly distinguishes topological edge pinning, NHSE-induced boundary accumulation, and random Anderson localization.
-
Improved AI can reconstruct effective band structures from finite-size measurements to infer bulk dispersion relations. This allows for the inference of
effective dispersion relation from measured data
in 2D systems like graphene without requiring exact Hamiltonian diagonalization. -
Improved AI can distinguish between gapless Dirac semimetals and topological Chern insulators by analyzing reconstructed band structures and Berry curvature, as DMD can
directly distinguish a gapless Dirac semimetal from a topological Chern insulator.
-
Improved AI can reconstruct the quantum metric and Berry curvature of extracted DMD modes using gauge-invariant discretized formulas derived from link variables, providing access to the
quantum geometric tensors
of the data-driven band structure.
Abstract
We present a data-driven framework for reconstructing band structures using Koopman operator analysis and dynamic mode decomposition (Koopman-DMD). Instead of deriving spectra from an explicit Hamiltonian, the approach reconstructs band dispersion and modal dynamics directly from spatiotemporal data, including wavefunctions and observables. This framework establishes a correspondence between Hamiltonian Floquet-Bloch decomposition and Koopman-DMD, whereby the extracted DMD modes encode frequencies, decay or growth rates, spatial profiles and projection weights. These quantities allow the reconstruction of spectral functions, local density of states, and delocalized-to-localized measures such as the inverse participation ratio. Also, these extended DMD modes enable inference of quantum-geometric and topological properties, including the quantum metric, Berry curvature and geometric phases. Applications to prototypical one- and two-dimensional tight-binding models, including disordered Su-Schrieffer-Heeger model and its Floquet and non-Hermitian variants, graphene and Haldane models, demonstrate that Koopman-DMD provides a unified route for the data-driven analysis of wave propagation, localization, and topological phases in condensed matter, photonics, and related fields.
Sources
- Quantum Geometry Phenomena in Condensed Matter Systems
- Topology and Localizations in a 2D Su-Schrieffer-Heeger Model with Domain Walls, Quasi-periodic Disorder and Periodic Hopping Modulations
Related papers
- Gradient-Based Inverse Design of Free-Energy Landscapes with Diffusion Models
- Mixture of Polyconvex Neural Potentials for Parametric Hyperelasticity: Towards Foundation Material Models
- Development and large-scale benchmarks of a protein--ligand absolute binding free energy toolkit
- Spin disorder competing with positional symmetry breaking governs the metal-insulator behavior in oxide paramagnets
- Investigation of the Asymptotic Properties of Active Impedance in Large Finite Array Antennas
- Automatic generation of exchange-correlation response kernels