Palqo Achieves 30x Speedup in Large-Scale Quantum Optimization with Physics-Informed Neural Networks

Quantum optimisation represents a promising pathway towards realising the potential of near-term quantum computers, but a fundamental limitation of these algorithms, the no-cloning theorem, creates significant computational demands as problems scale up. Yiming Huang from Peking University, Yajie Hao from the University of Electronic Science and Technology of China, Jing Zhou from Fudan University, and colleagues, now present a new approach that dramatically reduces these costs. The team reformulates the training process of these algorithms as a problem governed by a partial differential equation, and then utilises physics-informed neural networks to efficiently model and predict how parameters should be updated. This innovative method achieves up to a 30-fold increase in speed and reduces resource consumption by as much as 90% for problems involving up to 40 qubits, paving the way for more practical and efficient quantum optimisation techniques.

Koopman Operator Optimizes Variational Quantum Algorithms

Scientists have developed a new optimization algorithm, PALQO, for variational quantum algorithms (VQAs) that significantly improves efficiency and reduces the demand on quantum computing resources. Recognizing the challenges of optimizing VQAs, the team leveraged the Koopman operator to model the dynamics of the optimization process, allowing PALQO to predict future parameter updates and adapt the optimization strategy accordingly. The method combines the power of quantum computation for evaluating the cost function with classical computation for learning and adapting the optimization strategy. Through extensive simulations, PALQO consistently outperformed other algorithms, achieving lower energy values, faster convergence, and higher accuracy in quantum machine learning tasks, while also demonstrating good scalability and robustness to quantum noise. This work addresses a critical bottleneck in VQAs, offering a novel and comprehensive solution with practical implications for realizing the full potential of quantum computation. The underlying principles of PALQO could potentially be applied to other optimization problems beyond VQAs, opening new avenues for research and development.

Scientists Method

Scientists developed a novel methodology to accelerate variational quantum algorithms (VQAs) and reduce their reliance on quantum computing resources. Recognizing the limitations imposed by fundamental physical constraints, the team reformulated the parameter optimization process within VQAs as a nonlinear partial differential equation, shifting computational burden from quantum to classical hardware. The core of this work lies in the development of a protocol, termed PALQO, which leverages physics-informed neural networks (PINNs) to approximate solutions to the derived partial differential equation. By effectively learning from a limited initial dataset collected from quantum devices, the PINN predicts future parameter updates, dramatically reducing the number of costly quantum measurements required during optimization. Extensive numerical simulations, including systems of up to 40 qubits, demonstrate that PALQO achieves up to a 30x speedup compared to standard VQA optimization techniques, while simultaneously reducing quantum resource costs by as much as 90%. The team proved that a polynomial number of training samples is sufficient to ensure PALQO’s generalization ability, further solidifying its potential for scalability.

PINNs Accelerate Variational Quantum Algorithm Training

Scientists have developed a new method to significantly accelerate the training of variational quantum algorithms (VQAs), a leading strategy for realizing practical applications of near-term quantum devices. The research addresses a fundamental limitation of VQAs, the prohibitive resource costs associated with large-scale tasks. The team reformulated the training dynamics of VQAs as a nonlinear partial differential equation, then leveraged physics-informed neural networks (PINNs) to efficiently model this system. This innovative approach predicts parameter updates for VQAs using only a small amount of initial training data collected from quantum devices, dramatically reducing the computational burden on quantum hardware.

Experiments demonstrate the method achieves up to a 30x speedup compared to conventional VQA training methods, while simultaneously reducing resource costs by as much as 90% for tasks involving up to 40 qubits. These improvements were validated across several ground state preparation tasks, including simulations of the transverse-field Ising model, the Heisenberg model, and multiple molecule systems. The team proved that a polynomial number of training samples is sufficient to ensure the method’s ability to generalize effectively.

Physics-Informed Neural Networks Accelerate Quantum Algorithms

This research presents a novel approach to optimizing variational quantum algorithms (VQAs), which are leading strategies for utilizing near-term quantum devices. Recognizing the resource limitations imposed by fundamental physical constraints, the team reformulated the training dynamics of VQAs as a nonlinear partial differential equation. They then developed a protocol, termed PALQO, that employs physics-informed neural networks to efficiently model this dynamical system, significantly reducing the need for repeated evaluations on quantum hardware. Through extensive numerical experiments involving systems of up to 40 qubits, PALQO demonstrated substantial improvements in efficiency, achieving up to a 30x speedup and reducing resource costs by as much as 90% compared to conventional methods, while maintaining competitive accuracy.

The method predicts parameter updates on the classical side, effectively learning the optimization path and minimizing the demand for quantum queries. While acknowledging that it remains uncertain whether PALQO will scale to a regime where quantum hardware definitively outperforms classical approaches, the results at currently accessible scales are highly encouraging. Future research will focus on incorporating adaptive strategies and variance reduction techniques to further enhance PALQO’s potential and unlock even greater efficiency in large-scale VQAs.