Quantum Circuits Now ADAPT to Hardware, Boosting Accuracy and Speed

A new system, TuniQ, optimises the compilation of quantum circuits for integration with high-performance computing systems. Mohammad Abrarul Hasanat at Rice University, and colleagues, in collaboration with the University of Utah, have developed the system which dynamically selects compilation passes based on circuit characteristics, hardware specifications, and current noise levels. TuniQ improves both the fidelity and compilation speed compared to existing state-of-the-art compilers like IBM Qiskit, and exhibits strong scalability and generalisation across different quantum backends without the need for retraining. The system offers a flexible set of tools vital for quantum computation.

Reinforcement learning optimises quantum circuit fidelity beyond established thresholds

Total Variation Distance, a measure of how accurately a quantum computation reproduces the ideal result, fell to under 0.03 across all benchmarks using TuniQ, a substantial improvement over the fidelity-optimised IBM Qiskit transpiler. This 0.03 threshold is important because it represents a level of accuracy previously unattainable for complex circuits, unlocking the potential for more reliable quantum algorithms. Quantum algorithms, particularly those intended for practical applications like materials science or drug discovery, are highly sensitive to errors. Reducing the Total Variation Distance to below 0.03 signifies a considerable reduction in the probability of obtaining incorrect results, thereby enhancing the reliability of these computations. TuniQ achieves this by dynamically selecting from over 100 compilation passes, a process impossible with conventional fixed-sequence compilers. These compilation passes encompass a range of transformations, including qubit mapping, gate decomposition, and optimisation techniques, each impacting the circuit’s fidelity and runtime.

The system’s reinforcement learning approach selects compilation passes at each pipeline stage, adapting to circuit structure, quantum hardware, and the current noise profile of the processor, improving fidelity and reducing compilation time. Improvements were observed when TuniQ was applied to circuits of varying complexities, utilising a dual-encoder design that captures logical qubit interactions and device noise characteristics. The dual-encoder architecture is crucial; one encoder focuses on the logical structure of the quantum circuit, the qubits and the gates that operate on them, while the other encoder models the physical characteristics of the quantum hardware and its inherent noise. This allows TuniQ to learn a policy that considers both the algorithmic requirements and the limitations of the underlying hardware. Its efficacy was also demonstrated on multiple IBM Quantum Cloud processors, including Torino, Pittsburgh, and Fez, where optimal pass configurations differed depending on the hardware. This highlights the system’s adaptability and its ability to tailor compilation strategies to specific hardware constraints. In particular, TuniQ responds to fluctuating noise levels, adjusting pass selection based on simulated noise decreasing from baseline to 30%, prioritising layout quality under high noise and refinement under lower noise. Under high noise conditions, prioritising layout quality, minimising the physical distance between interacting qubits, reduces the impact of decoherence and gate errors. Conversely, when noise is lower, the system can focus on more complex refinement passes to further optimise the circuit’s performance. While TuniQ consistently achieves improved fidelity over the IBM Qiskit transpiler, these gains currently focus on single circuit optimisation and do not yet address maintaining performance during prolonged operation or with complex, multi-layered quantum algorithms.

Assessing portability limitations of an adaptive quantum compilation system

As quantum processors become integrated into larger high-performance computing systems, optimising quantum circuits for execution presents a significant hurdle. The integration of quantum coprocessors into HPC workflows necessitates efficient compilation to minimise overhead and maximise the benefits of quantum acceleration. Current compilation techniques often struggle to balance fidelity, runtime, and resource utilisation, particularly as circuit complexity increases. Although TuniQ demonstrably improves compilation fidelity and speed over existing tools, its current validation relies heavily on IBM Quantum Cloud processors. This raises a key question: can TuniQ maintain its performance advantage when deployed on radically different hardware architectures, or will its learned strategies prove brittle outside the IBM ecosystem.

The reliance on IBM Quantum Cloud processors for initial validation introduces a potential limitation regarding portability. Quantum hardware varies significantly in terms of qubit connectivity, gate fidelity, and noise characteristics. A system trained exclusively on one platform may not generalise well to others. Future work should investigate TuniQ’s performance on a wider range of quantum backends, including those based on superconducting qubits from different manufacturers, trapped ions, and photonic systems. This will require either transfer learning techniques to adapt the learned policy to new hardware or continuous retraining with data from diverse platforms. Furthermore, assessing the system’s robustness to variations in qubit calibration and control parameters is crucial for real-world deployment.

Improving compilation speed and accuracy is fundamental to unlocking the potential of quantum computing as a coprocessor within larger high-performance computing systems, and TuniQ demonstrably advances both metrics. A shift in quantum compilation is represented by TuniQ, moving beyond static methods to a dynamic system adapting to specific circuit and hardware conditions. By employing reinforcement learning, the system intelligently selects the optimal sequence of transformations, known as compilation passes, at each stage of the process. The reinforcement learning agent learns a policy that maps circuit and hardware states to optimal pass selections, maximising a reward function that balances fidelity and runtime. This adaptability demonstrably improves both the accuracy of quantum calculations and the speed at which circuits are prepared for execution on IBM Quantum Cloud processors. The potential impact extends beyond simply improving existing quantum algorithms; it also facilitates the development of new algorithms that were previously impractical due to compilation bottlenecks. The ability to efficiently compile complex circuits opens up opportunities for exploring more sophisticated quantum error correction schemes and tackling larger, more challenging computational problems.

Further research will focus on extending TuniQ’s capabilities to handle multi-layered quantum algorithms and maintain performance during prolonged operation. Addressing these challenges will require developing techniques for managing resource allocation, optimising circuit scheduling, and mitigating the accumulation of errors over time. Ultimately, the goal is to create a compilation system that seamlessly integrates quantum co-processors into the broader HPC landscape, enabling a new era of scientific discovery and technological innovation.

TuniQ successfully improved both the fidelity and compilation time of quantum circuits. This matters because efficient compilation is essential for integrating quantum processors with existing high-performance computing systems. The system achieves this by using reinforcement learning to dynamically select the best sequence of compilation passes, adapting to both the circuit and the specific hardware on IBM Quantum Cloud processors. Researchers intend to extend TuniQ to handle more complex, multi-layered algorithms and maintain performance over longer periods of operation.