Shao-Bo Zhou and colleagues at TuringQ Co., Ltd. have achieved a 56.9% reduction in mean Jensen-Shannon divergence for the Lorenz-63 model using a novel quantum prior, compared to a classical multilayer perceptron with a similar number of weights across 96 unseen parameter values. The team developed a conditional quantum prior that learns statistical patterns without retraining for new parameter values within a defined range, a surprising feat for quantum models.

This shared prior requires 84.7% fewer trainable weights than separate quantum neural networks would need for 101 parameter values, and also reduces forecast mean squared error by 66.0% for Kolmogorov flow at Reynolds number Re = 1000.

Photonic Quantum Neural Networks for NISQ Devices

A photonic quantum neural network achieved a mean Jensen-Shannon divergence of 4.58× 10^(-3) for the Lorenz-63 model, demonstrating its capacity to model chaotic systems with improved statistical accuracy at parameter values not used during training. The network encodes inputs through manipulations of optical modes with beam splitters and phase shifters, generating detection probabilities that are then used in conjunction with trainable classical components. Experimentally, a photonic quantum neural network was used to implement the conditional prior learning process, using measured optical responses and trainable classical components to construct conditional probability distributions.

Graph QCBM Achieves Low Error Rates for Lorenz-63

A single conditional quantum prior reduces errors in modeling chaotic systems, achieving a mean Jensen-Shannon divergence of 0. The approach, detailed in recent work, relies on a “Graph QCBM” architecture to learn how key statistical properties change with system parameters without requiring retraining for each new value. The efficiency of this method is notable. The Graph QCBM requires only 2,976 trainable weights to achieve its performance on Lorenz-63, a reduction of 84. The researchers also demonstrated the effectiveness of their approach with Kolmogorov flow, indicating a broader applicability beyond the specific Lorenz-63 model.

By using the learned prior to regularize classical networks, the researchers were able to encourage more accurate long-term predictions of chaotic dynamics. The results show that training on only five parameter values is sufficient to achieve the prescribed accuracy threshold at unseen conditions for both Lorenz-63 and Kolmogorov flow, a significant reduction in the amount of training data required.

“One shared set of trainable weights provides statistical references at unseen parameter values within the training range without retraining,” the paper states, emphasizing the adaptability of the system. Further research will explore the benefits of this approach in scenarios with noisy data and investigate the limits imposed by finite-sample error in the statistical penalty.

Conditional Quantum Priors Improve Long-Term Forecasts

The Graph QCBM, a quantum circuit architecture, achieved a 53.36 x 10^-3 mean Jensen-Shannon divergence on unseen parameter values for the Lorenz-63 model, demonstrating improved statistical forecasting capabilities. This level of performance surpasses that of a conditional multilayer perceptron, which exhibited a 1.2385 x 10^-3 divergence using a comparable number of trainable weights. The key to this improvement lies in the development of a conditional quantum prior that learns the relationship between system parameters and dynamical statistics.

Statistical Regularization Reduces Error in Kolmogorov Flow

This improvement demonstrates the potential of quantum-inspired techniques to enhance the accuracy of chaotic fluid dynamics simulations, a field vital for modeling weather patterns and turbulence. The framework employed uses a prior trained on statistical distributions derived from trajectories at multiple system parameter values, offering a significant advantage over independently trained priors. For Kolmogorov flow at Reynolds number Re = 1000, prior regularization reduces forecast mean squared error by 66.0% over 318 steps relative to the classical baseline without prior regularization.

Across 101 parameter values, the shared prior uses 84.7% fewer trainable weights than would be required by a separate quantum neural network at each value. The researchers also implement a photonic quantum prior for noisy intermediate-scale quantum (NISQ) devices, achieving a mean Jensen-Shannon divergence of 4.58× $10^{- 3}$ for Lorenz-63 and 2.59× $10^{- 3}$ for Kolmogorov flow at unseen parameter values.

Parameter Efficiency of Shared Quantum Priors

The use of a single conditional quantum prior markedly reduces the number of trainable weights needed to model chaotic systems, offering a significant advantage over traditional approaches. A Graph QCBM, employing 2,976 shared weights, achieved comparable performance to a classical quantum-inspired machine learning (QIML) network that would require 19,392 weights when extended to 101 different parameter values. Beyond weight reduction, the conditional prior demonstrates an ability to generalize within the training range without requiring retraining for new parameter values.

Quantum Priors Address Limitations of Independent QIML Models

A conditional quantum prior reduced mean Jensen-Shannon divergence by 56.9% relative to a classical multilayer perceptron with a comparable number of weights across 96 unseen parameter values. This improvement demonstrates the framework’s capacity to generalize beyond training data, a critical step toward more robust predictions of chaotic systems.

The team developed a method that learns how dynamical statistics depend on system parameters, unlike earlier independent priors that treated each parameter value in isolation. Across 101 parameter values, the shared prior uses 84.7% fewer trainable weights than would be required by a separate quantum neural network at each value.

Lorenz-63 and Kolmogorov Flow as Chaotic Test Cases

These results indicate the potential for improved generalization in modeling complex fluid dynamics, a long-standing challenge in computational science. This regularization process, applied to classical neural networks, encourages more accurate long-term statistical reproduction of the chaotic system’s behavior. For Kolmogorov flow at Reynolds number Re = 1000, prior regularization reduces forecast mean squared error by 66.0% over 318 steps relative to the classical baseline without prior regularization.

Framework for Training Conditional Quantum Priors

This adaptability stems from learning how selected dynamical statistics depend on a system parameter, a departure from independently trained priors that fail to capture shared dependencies. Regularization with this fixed prior demonstrably improves prediction accuracy for complex fluid dynamics, specifically reducing mean squared error by 66.0% over 318 steps relative to the classical baseline without prior regularization. The results suggest a robust method for improving the accuracy and efficiency of chaotic fluid dynamics modeling.

👉 More information
🗞 Conditional Quantum Priors for Chaotic Fluid Dynamics
✍️ Shao-Bo Zhou, Yu-Ze Zhu, Hang Song, Xiang Zhao, Lin Yang and Xian-Min Jin
🧠 ArXiv: https://arxiv.org/abs/2611.04872

More like this