TLDR: A new research paper introduces two classical learning models, hcs and hqs, designed to efficiently emulate the mean-value behavior of large-scale quantum processors. These “predictive surrogates” significantly reduce the need for direct quantum hardware access. Experimentally validated on a 20-qubit superconducting processor, hcs successfully pre-trained variational quantum eigensolvers with drastically fewer measurements and improved accuracy, while hqs effectively identified quantum phase transitions, also with substantial measurement overhead reduction. This work offers a practical approach to broaden the impact of advanced quantum computing by leveraging classical AI techniques.
The rapid advancement of quantum processors is opening new frontiers in scientific discovery. However, the immense cost associated with building and maintaining these large-scale quantum machines means they will remain a scarce resource for the foreseeable future. This scarcity poses a significant challenge to their widespread adoption and application.
To overcome this hurdle, researchers have introduced an innovative concept: predictive surrogates. These are classical learning models specifically designed to mimic the average behavior of a given quantum processor. The key advantage of these surrogates is their provable computational efficiency, meaning they can perform complex emulations without requiring extensive access to actual quantum hardware.
The paper, titled “Demonstration of Efficient Predictive Surrogates for Large-scale Quantum Processors,” proposes two distinct types of predictive surrogates, each tailored for different practical scenarios, that can substantially reduce the need for direct quantum processor access. These surrogates were put to the test by emulating a quantum processor with up to 20 programmable superconducting qubits, showcasing their potential in advancing digital quantum simulation.
Two Powerful Surrogates: hcs and hqs
The first surrogate, known as hcs, is an extension of existing models adapted for noisy quantum environments. It functions as a “classical shadow predictor,” capable of estimating mean values for various local measurements on quantum states, even those previously unseen. This is achieved without needing further interaction with the quantum processor. The theoretical analysis suggests that hcs is computationally efficient, especially when noise levels are relatively high or when the quantum circuit’s behavior is smooth. A significant application demonstrated in the research is using hcs to pre-train Variational Quantum Eigensolvers (VQEs). VQEs are algorithms used to find the ground state energies of quantum systems, a process that typically demands extensive quantum resources. By pre-training VQEs with hcs, the entire estimation of ground state energies can be performed classically, drastically cutting down on quantum measurement overhead. Experimental results showed that hcs not only reduced measurements by orders of magnitude (using just 0.023% of measurements) but also yielded more accurate ground state energy estimates than conventional VQEs trained directly on the quantum processor.
The second surrogate, hqs, is designed for scenarios where quantum circuit parameters are correlated, a common setup in quantum simulation studies. Unlike hcs, hqs can handle input parameters sampled from arbitrary distributions. The research proves that hqs is computationally efficient, particularly when the quantum processor’s noise is modeled by Pauli noise channels and the input parameters are within a small range. The potential of hqs was explored in identifying Floquet symmetry-protected topological (FSPT) phases, which are unique non-equilibrium states of matter. Traditionally, identifying these phases requires numerous quantum measurements across different parameters. The proposed hqs method enables a massive reduction in this measurement overhead by training a set of hqs models to emulate the mean-value behavior. Experimental findings confirmed that the optimized hqs models could reliably pinpoint the transition region of FSPT phases, aligning with expected signatures.
Also Read:
- Fisher Information Enhances Quantum Federated Learning Performance
- Tackling High-Dimensional Control Games with PINN-Based Policy Iteration
Real-World Validation on a 20-Qubit Processor
All experiments were conducted on a 20-qubit chain within a 66-qubit programmable superconducting processor. The results consistently demonstrated the efficacy of both hcs and hqs. For VQE pre-training, hcs showed strong agreement with actual quantum processor outputs, with errors decreasing as the training dataset size increased. It also exhibited favorable scaling behavior, maintaining comparable performance across varying qubit counts (up to 20 qubits), suggesting its applicability to larger systems. For FSPT phase identification, the set of hqs surrogates successfully captured the distinct signatures of FSPT and thermal phases, and accurately identified the critical transition point. This was achieved with significantly fewer quantum measurements compared to traditional methods.
These findings collectively establish predictive surrogates as a practical and powerful pathway to broadening the impact and accessibility of advanced quantum processors. By offloading computationally intensive tasks to classical models, they pave the way for more efficient and widespread use of quantum computing resources. For more in-depth details, you can refer to the full research paper: Demonstration of Efficient Predictive Surrogates for Large-scale Quantum Processors.
The researchers envision several exciting future directions, including applying predictive surrogates to other variational quantum algorithms, exploring their efficiency with quantum circuits possessing intrinsic symmetries, extending their capabilities to learn nonlinear functions like entanglement entropy, and developing surrogates that can emulate full output distributions rather than just mean values. The integration of deep learning techniques with these surrogates also remains a promising area for further enhancement.
