A Quantum-Classical Surrogate Model for the Collision Operator of the Lattice Boltzmann Method

arXiv:2606.31351 · quant-ph · Submitted 2026-06-30 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Quantum Radio. Generated commentary on the latest quantum physics and condensed matter papers.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "A Quantum-Classical Surrogate Model for the Collision Operator of the Lattice Boltzmann Method".

Mira: Here's the gist of the paper. A hybrid approach utilizing a quantum machine learning surrogate model to approximate the non-linear collision dynamics of the Lattice Boltzmann Method effectively offloads non-unitary operations that challenge pure quantum solvers The proposed surrogate is validated against the classical BGK collision operator using established benchmark problems, including the Taylor–Green vortex for evaluating energy dissipation and the double shear layer for assessing shear-driven instabilities and nonlinear flow evolution These findings suggest that hybrid quantum-classical strategies offer a practical path toward realizing the potential of quantum computing in fluid engineering.

Kai: Model Foundation The Lattice Boltzmann Method (LBM) discretizes space, time, and velocity on a lattice where particle distributions propagate with discrete lattice velocities and interact The collision operator in its exact form involves a high-dimensional integral over all possible pre- and post-collision velocities, with non-linear dependencies on the particle distribution functions and the collision crosssection Simplified models like the Bhatnagar–Gross–Krook (BGK) approximation formulate the collision operator as omegai = −1/τ(fi − feqi), representing relaxation towards its equilibrium state with relaxation time τ The LBE is solved in two steps: a collision step where each population undergoes a local update f∗i(x, t) = fi(x, t) − ∆t τ(fi(x, t) − feqi(x, t)), and a streaming step where post-collision populations are propagated along ci fi(x + ci∆t, t + ∆t) = f∗i(x, t).

Mira: Circuit Architecture Design The proposed quantum circuit is grounded in a Fourier-series interpretation, providing a principled view of quantum circuit learning as a truncated spectral approximation The design relies on established hardware-efficient ansätze tailored to NISQ devices, employing shallow circuits, angle-based encoding, and hardware-native quantum operations consisting of single-qubit rotations and two-qubit CNOT gates The circuit output is obtained as the expectation of a Hermitian observable Oˆ, given by y(xi, θ) = ⟨ψ0 U(x)†U(θ)†OUˆ (θ)U(x)ψ0⟩ Entangling layers control how quantum basis functions are combined and determine the accessible Fourier coefficients cω

2606.31351>. [Kai: Performance Metrics Analysis The expressibility of a VQC is quantified by the Kullback–Leibler (KL) divergence DKL PˆVQC(F; θ) PHaar(F), where smaller values of DKL indicate a higher expressibility The analysis showed that LSW, the number of encoding repetitions, is the primary driver of expressibility, while additional LW layers contribute only marginally beyond the first Entanglement is quantified by the Meyer-Wallach (MW) measure ⟨Q⟩, and for PA-SEL architectures, this measure converges monotonically to the Haar benchmark as depth increases.

Title and authors: Kai: The title of this paper is "A Quantum-Classical Surrogate Model for the Collision Operator of the Lattice Boltzmann Method," and the authors are Lukas Birka, David Wawrzyniakb, Josef Wintera, Steffen Schmidta, Thomas Indingera, Christian Janßenc, Nikolaus Adamsa.

Mira: It’s interesting how they frame it as a hybrid approach because it acknowledges that pure quantum solvers have difficulty handling the non-linear collision dynamics of the LBM.

Lev: What I find noteworthy is that this paper focuses on using quantum machine learning to approximate those difficult parts, which is a practical way to tackle problems that are too complex for current hardware.

Kai: They build the expressivity of their surrogate model based on the ability of parameterised quantum circuits to implement partial Fourier series, and they mention data re-uploading helps extend the spectrum of frequencies they can represent.

Mira: So it’s not just a fixed relaxation parameter like in simpler models; they are able to vary how that parameter works within the quantum circuit structure itself.

The paper's summary: Kai: The core summary of this paper is that they introduce this hybrid model to approximate the non-linear collision dynamics of the Lattice Boltzmann Method, specifically targeting those non-unitary operations that challenge pure quantum solvers.

Mira: They then validate it by testing against classical operators like BGK, using benchmarks such as the Taylor–Green vortex for energy dissipation and the double shear layer to assess instabilities and nonlinear flow evolution.

Lev: What I’m looking at is how they handle the LBM itself, which involves a collision step where each particle population gets an update based on its distribution function and a relaxation time tau.

Kai: They show that they are solving this in two steps: first the collision step with the approximated operator, and then a streaming step where those post-collision populations are propagated along their lattice velocities.

Mira: The results they get show that for the D3Q19 lattice, when they trained their SEL1 L6 scheme, they got median relative errors around ten to the power of negative four for all discrete velocity distributions <ref:2606.31351#pg1>.

The paper's improvements: Kai: One of the key things this paper suggests is that they can improve it by allowing for modular collision model capabilities, like using an entropic multi-relaxation-time scheme to impose an entropy condition on the relaxation step and use different relaxation rates for different moment orders.

Mira: That’s a big move because it means we can train dedicated surrogate pipelines for completely different sets of relaxation parameters, which would let us handle more complex collision models than just the standard BGK approximation.

Lev: And from an error correction standpoint, they suggest enforcing conservation and symmetry constraints by applying them as post-processing to the measured outputs using algebraic corrections like f star equals Omega c of f equals U f plus V omega of f.

Kai: That’s interesting because it means we can achieve physical consistency by design in a full-stack quantum LBM surrogate, instead of just trying to fix errors after the simulation is done.

Conclusion: Mira: To wrap up, this paper shows that these hybrid quantum-classical strategies offer a practical path toward realizing the potential of quantum computing in fluid engineering because they offload those challenging non-unitary parts.

Kai: They successfully demonstrated that the surrogate accurately captures the symmetric, periodic decay of vortices over time in simulations like the Taylor–Green vortex, and they found that small macroscopic velocities with magnitude less than zero point zero one fall within the near-linear regime of feq where accuracy is highest.

Lev: If you’re running this on real hardware right now, you need to be aware that those post-processing corrections for mass and momentum conservation and rotational symmetry are necessary safeguards because they address limitations in the surrogate accuracy and for long integration horizons.

Mira: We're looking forward to seeing how these modular improvements can be used to train surrogates for much more intricate collision models than just BGK, which is what this paper sets up for future work.

Kai: So, "A Quantum-Classical Surrogate Model for the Collision Operator of the Lattice Boltzmann Method" gives us a solid foundation on how to use quantum machine learning to handle complex fluid dynamics problems in a way that’s actually testable right now.

Lukas C. Birka, David M. Wawrzyniakb, Josef M. Wintera, Steffen J. Schmidta, Thomas Indingera, Christian F. Janßenck, ¡ Nikolaus A Adams

Technical University of Munich · Siemens Digital Industries Software

quant-ph

Submitted: 2026-06-30

Updated: 2026-11-13

DOI: 10.1016/j.cpc.2026.110513

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 88/100

The gist: A hybrid approach utilizing a quantum machine learning surrogate model to approximate the non-linear collision dynamics of the Lattice Boltzmann Method effectively offloads non-unitary operations…

Key concepts

Lattice Boltzmann Method (LBM)
A numerical technique used to simulate fluid flow by discretizing space and time onto a lattice. Particles move around this grid, and their interactions determine the overall fluid behavior.
BGK Collision Operator
A simplified mathematical model used in LBM that represents how particles collide and relax towards equilibrium. The quantum model is trained to mimic this specific relaxation process for accurate simulation.
Quantum Surrogate Model
A quantum circuit designed to approximate the complex, non-linear collision operator of the LBM. It uses quantum mechanics principles to learn the dynamics efficiently, replacing computationally expensive classical calculations.
Expressibility (KL Divergence)
A metric used to measure how well a quantum model can represent a function. A lower KL divergence value indicates that the quantum circuit is more capable of accurately approximating the true fluid dynamics.

Terminology

Summary

A hybrid approach utilizing a quantum machine learning surrogate model to approximate the non-linear collision dynamics of the Lattice Boltzmann Method effectively offloads non-unitary operations that challenge pure quantum solvers The proposed surrogate is validated against the classical BGK collision operator using established benchmark problems, including the Taylor–Green vortex for evaluating energy dissipation and the double shear layer for assessing shear-driven instabilities and nonlinear flow evolution These findings suggest that hybrid quantum-classical strategies offer a practical path toward realizing the potential of quantum computing in fluid engineering. p. 0

Where the paper says it

The proposed surrogate is validated against the classical BGK collision operator using established benchmark problems, including the Taylor–Green vortex for evaluating energy dissipation and the double shear layer for assessing shear-driven instabilities and nonlinear flow evolution. Our results demonstrate that the hybrid model achieves high accuracy and generalisability while closely replicating classical solutions.

Model Foundation

The Lattice Boltzmann Method (LBM) discretizes space, time, and velocity on a lattice where particle distributions propagate with discrete lattice velocities and interact The collision operator in its exact form involves a high-dimensional integral over all possible pre- and post-collision velocities, with non-linear dependencies on the particle distribution functions and the collision crosssection Simplified models like the Bhatnagar–Gross–Krook (BGK) approximation formulate the collision operator as omegai = −1/τ(fi − feqi), representing relaxation towards its equilibrium state with relaxation time τ The LBE is solved in two steps: a collision step where each population undergoes a local update f∗i(x, t) = fi(x, t) − ∆t τ(fi(x, t) − feqi(x, t)), and a streaming step where post-collision populations are propagated along ci fi(x + ci∆t, t + ∆t) = f∗i(x, t).

Circuit Architecture Design

The proposed quantum circuit is grounded in a Fourier-series interpretation, providing a principled view of quantum circuit learning as a truncated spectral approximation The design relies on established hardware-efficient ansätze tailored to NISQ devices, employing shallow circuits, angle-based encoding, and hardware-native quantum operations consisting of single-qubit rotations and two-qubit CNOT gates The circuit output is obtained as the expectation of a Hermitian observable Oˆ, given by y(xi, θ) = ⟨ψ0 U(x)†U(θ)†OUˆ (θ)U(x)ψ0⟩ Entangling layers control how quantum basis functions are combined and determine the accessible Fourier coefficients cω [2606.31351>.

Performance Metrics Analysis

The expressibility of a VQC is quantified by the Kullback–Leibler (KL) divergence DKL PˆVQC(F; θ) PHaar(F), where smaller values of DKL indicate a higher expressibility The analysis showed that LSW, the number of encoding repetitions, is the primary driver of expressibility, while additional LW layers contribute only marginally beyond the first Entanglement is quantified by the Meyer-Wallach (MW) measure ⟨Q⟩, and for PA-SEL architectures, this measure converges monotonically to the Haar benchmark as depth increases. The Effective Dimension (ED) analysis confirmed that while absolute ED increases with depth, it decreases sublinearly in Nθ, indicating that successive parameters contribute progressively fewer independent directions. p. 4

Where the paper says it

Established VQC analysis tools expressibility, entanglement, and effective dimension are applied to the circuit design process, and their predictive validity is critically assessed for fluid-dynamics applications. A problem-specific performance analysis further investigates the surrogate’s interpolation and extrapolation behaviour, robustness to variations in the input parameter range, and sensitivity to circuit depth and width, identifying the key architectural parameters that govern accuracy.

Numerical Validation and Corrections

The quantum surrogate replaces the collision operator in full LBM simulations, with conservation and symmetry corrections applied as post-processing at each time step For the D3Q19 lattice, the trained SEL1 L6 scheme achieved median relative errors around 10−4 for all discrete velocity distributions The surrogate accurately captures the symmetric, periodic decay of vortices over time in the Taylor–Green vortex simulation Post-processing corrections for mass and momentum conservation and rotational symmetry were necessary safeguards because they address surrogate-accuracy limitations and long integration horizons.

Generalization Capabilities

The variable relaxation training strategy with ω ∈ [0.5, 1.99] enables the circuit to generalize across the full viscosity range spanned by the three Reynolds numbers without retraining The surrogate achieves its highest accuracy when small macroscopic velocities (ui ≲ 0. p. 0

Where the paper says it

Unlike previous approaches with a fixed relaxation parameter, the surrogate recovers the complete Bhatnagar-Gross-Krook (BGK) collision dynamics across the full physically admissible range of relaxation without retraining. We reassess the relevance of standard variational quantum circuit (VQC) metrics, including expressibility, entanglement, and effective dimension, by relating them directly to task-specific surrogate performance and identifying the key architectural parameters that determine approximation accuracy.

01) fall within the near-linear regime of feq, where it represents high accuracy The results confirm that small macroscopic velocities (ui ≲ 0. p. 5
Where the paper says it

The macroscopic moments, density ρ and momentum ρu, are recovered as ρ = X i fi , ρu = X i fi ci . (3) The collision operator Ωi in its exact form involves a high-dimensional integral over all possible pre- and post-collision velocities, with non-linear dependencies on the particle distribution functions and the collision crosssection.

01) fall within the near-linear regime of feq, where the surrogate achieves its highest accuracy. p. 3
Where the paper says it

Both approaches demonstrate that learned unitary collision circuits can reproduce the non-linear BGK relaxation with controllable accuracy in the low-Reynolds-number regime; however, this capability is shown only for a fixed relaxation time (τ = 1), and extension to other relaxation parameters requires retraining. In the following, we introduce a novel hybrid quantum machine learning surrogate model tailored for NISQ hardware to approximate the complete post-collision distribution of the lattice Boltzmann method.

The median error increases monotonically from ∼ 10−5 at a = 0. p. 1
Where the paper says it

NISQ devices typically consist of approximately 10 to 100 qubits, with interactions restricted by the underlying hardware topology rather than being fully connected. Furthermore, these devices have limited error correction, resulting in significant computational errors.

001 to ∼ 3 × 10−3 at a = 0.1, with the most pronounced jump occurring between a = 0.01 and a = 0.05.

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

  1. Modular collision model capabilities: The framework allows for Entropic multi-relaxation-time schemes, for instance, to impose an entropy condition on the relaxation step and involve multiple relaxation rates for different moment orders. This enables training dedicated surrogate pipelines for distinct sets of relaxation parameters, allowing the system to handle complex collision models beyond BGK.

  2. Physics-based constraint enforcement: The system can enforce conservation and symmetry constraints as post-processing applied to the measured outputs, using algebraic corrections such as f∗ = omegac(f) = Uf + V omega(f). This allows for a full-stack quantum LBM surrogate where physical consistency is achieved by design rather than solely by correction.

  3. Generalization across relaxation parameters: The hybrid model can generalize across the full physically admissible range of relaxation without retraining, recovering the complete Bhatnagar-Gross-Krook (BGK) collision dynamics across the full physically admissible range of relaxation without retraining. This capability is achieved by encoding the parameter as a circuit input.

  4. Architectural parameter selection: The system can select optimal architectural parameters based on performance metrics, identifying that LSW, the number of encoding repetitions, is the primary performance driver, and suggesting configurations like a three-qubit parallel ansatz with a single strongly entangling layer and LSW = 6 encoding repetitions for better resource efficiency.

  5. Robustness to noise and hardware limitations: The use of shallow circuits built exclusively from single-qubit rotation and CNOT gates keeps circuit depth low, minimizing decoherence, making the approach tractable on current hardware.

Abstract

We introduce a hybrid approach utilising a quantum machine learning surrogate model to approximate the non-linear collision dynamics of the LBM. It offloads the non-unitary operations that challenge pure quantum solvers. The surrogate's expressivity relies on variational quantum circuits that implement partial Fourier series, while data re-uploading extends the representable frequency spectrum. Unlike previous approaches that use a fixed relaxation parameter, the relaxation step is embedded analytically in the circuit, allowing the learned equilibrium to be used for any relaxation parameter without retraining. We reassess standard VQC metrics, including expressibility, entanglement, and effective dimension, by relating them directly to task-specific surrogate performance and identifying the architectural parameters that determine approximation accuracy. The proposed surrogate is validated against the classical BGK collision operator using benchmark problems, including the Taylor-Green vortex for energy dissipation and the double shear layer for shear-driven instabilities and nonlinear flow evolution. Our results show that the surrogate closely reproduces the BGK reference on both benchmarks, using a single set of trained circuits across all relaxation parameters. These findings suggest that hybrid quantum-classical strategies offer a practical path toward realising the potential of quantum computing in fluid engineering.

Sources

Related papers