Quantum learning recreates complex dynamics in two new circuit designs

Antonio David Bastida Zamora and colleagues at the University of Aix-Marseille in France are using quantum machine learning to enhance the quantum lattice Boltzmann method, a method for simulating complex fluid dynamics. In their work, they approximated the key nonlinear component of the method’s collision operator with a trained variational quantum circuit. By building an operator that can reproduce the dynamics of the Bhatnagar-Gross-Krook approximation, the team presents two different circuit architectures, the R1 and R2 models, providing flexible options for both continuous multistep simulation and high-precision single-step reconstruction of nonlinear operators. This significantly increases the efficiency and accuracy of fluid dynamics modeling.

Variational quantum circuits enable extended simulations of fluid dynamics

Multistep quantum lattice Boltzmann method (QLBM) simulations that scale up to O(log2 N) are now possible, overcoming previous limitations in capturing nonlinear dynamics without constant quantum state measurements. Previously, quantum algorithms required frequent interruptions of the simulation process to perform quantum state measurements known as “state tomography.” This requirement severely limited the ability to model fluid behavior over long periods of time, as the quantum state collapses and decoherence is introduced with each measurement. New variational quantum circuits (VQCs) circumvent this need by learning how to predict fluid dynamics using the Bhatnagar-Gross-Krook approximation, effectively embedding collision dynamics in the quantum circuit itself. This allows simulations to continue without repeating measurements, allowing longer and more complex simulations.

Two different circuit architectures, R1 and R2, were developed to address different simulation demands, with the former prioritizing continuous evolution and the latter focusing on single-step accuracy. The R1 model prioritizes efficient propagation of quantum states over long periods of time and is designed for simulations that require a large number of time steps. This is achieved through a shallower circuit depth, which reduces the error accumulation inherent in quantum computation. Conversely, the R2 model prioritizes the accuracy of individual time steps and employs deeper and more complex circuits to more faithfully represent the nonlinear collision operator. Current demonstrations are still limited to relatively small systems, and scaling to industrially relevant problem sizes involving millions of grid points remains a major engineering challenge. This advance builds on the foundation of the quantum lattice Boltzmann method and enables multi-step simulations that were previously limited to a single iteration. The trained variational quantum circuit accurately mimics the fluid behavior, and the two circuit designs R1 and R2 provide different strengths. R1 is better at modeling continuous evolution over time, while R2 prioritizes the accuracy of individual simulation steps. The successful reproduction of nonlinear collisional dynamics, which is key to modeling complex fluids, demonstrates the possibility of scaling the number of lattice sites as O(log2 N), opening the possibility of more efficient computational fluid dynamics. Logarithmic scaling is particularly important because it suggests that the computational cost of a simulation increases much more slowly with system size compared to traditional methods that typically scale polynomially.

Quantum machine learning tackles fluid simulations with initial simplifications

Fluid mechanics underpins countless processes, from weather forecasting and climate modeling to designing more efficient aircraft and optimizing industrial processes. Accurately modeling these flows remains a computational challenge. Traditional computational fluid dynamics (CFD) methods often rely on discretizing the fluid domain into a huge number of grid points, which is computationally prohibitively expensive, especially for turbulent flows. Employing machine learning offers a potential quantum shortcut to circumvent the difficulties of simulating complex fluid behavior with traditional computers. Although this approach relies on the Bhatnagar-Gross-Crook approximation, which simplifies the full Boltzmann equation, an important question arises as to how easily these quantum circuits generalize to more realistic and computationally demanding fluid models that lack this convenient simplification. The Bhatnagar-Gross-Krook model introduces a relaxation time τ that governs the rate at which the fluid returns to equilibrium after a disturbance. While this approximation simplifies calculations, it can limit the accuracy of simulations of certain fluid behaviors.

This represents a significant step forward, even though we currently rely on the Bhatnagar-Gross-Crook approximation, a simplification of real-world fluid dynamics. This potentially provides a pathway to speed up the simulation, and future work will focus on refining the model to handle more complex scenarios without this initial simplification. The ultimate goal is to develop quantum machine learning models that can accurately simulate fluid dynamics without relying on such approximations, potentially unlocking new insights into complex fluid phenomena. A functional quantum machine learning approach to fluid dynamics simulations has been established, bypassing the limitations of traditional algorithms. The multi-step simulation was achieved by training a programmable quantum computer to reproduce the behavior of the fluid without continuously measuring the quantum system. This provides more flexibility with separate circuit designs for R1 and R2. The training process involves tuning the parameters of a variational quantum circuit to minimize a cost function that quantifies the difference between predicted and expected fluid behavior based on the Bhatnagar-Gross-Krook approximation. The successful reproduction of nonlinear collision mechanics, an essential element in modeling complex fluids, enables further exploration of more efficient computational fluid dynamics. The ability to accurately model these dynamics is important for simulating a wide range of phenomena such as turbulence, shock waves, and multiphase flows. This research paves the way for investigating the potential of quantum computing to address long-standing challenges in fluid mechanics and related fields.

Researchers have successfully demonstrated a quantum machine learning approach that uses variational quantum circuits to reproduce nonlinear collision dynamics and simulate fluid dynamics. This is important because current fluid simulations often rely on approximations like the Bhatnagar-Gross-Krook model, which can have limited accuracy. Two different circuit architectures, R1 and R2, were developed to provide flexibility in the simulation approach. R1 enables multi-step simulation while R2 focuses on single-step accuracy. The authors plan to improve these models to eliminate the need for initial Bhatnagar-Gross-Krook simplifications and improve the accuracy of complex fluid simulations.